Category: Tech News

I am a passionate health, beauty, and lifestyle writer based in the south of the UK. Balancing my love for writing with raising my two children, I hopefully bring a relatable, insightful perspective to my articles.
  • Continuous Glucose Monitors Are Being Marketed to Healthy Britons, Here Is What the Evidence Actually Justifies

    Continuous Glucose Monitors Are Being Marketed to Healthy Britons, Here Is What the Evidence Actually Justifies

    Continuous glucose monitors were, not long ago, a clinical tool for people managing type 1 diabetes or unstable type 2. Now they are being advertised on Instagram, sold through wellness subscription services, and worn by healthy thirtysomethings who want to understand their “metabolic health”. In the UK, companies like Levels, Zoe and Supersapiens have made CGM for non-diabetics a genuine consumer category, with prices typically starting around £99 to £200 per month depending on the service. That is real money for what, in many cases, amounts to real-time blood glucose data that a healthy person may have no clinical need for.

    I want to be clear upfront: this is not an attack on the technology. CGMs are genuinely impressive medical devices, and the data they produce is real. The question is whether the interpretation layers being sold alongside them, and the lifestyle changes being recommended on the basis of individual glucose responses, are justified by the evidence we actually have.

    Woman checking CGM for non-diabetics use on her arm in a home kitchen
    Photo by Nataliya Vaitkevich on Pexels

    What CGMs are approved to do in the UK

    The MHRA regulates CGMs as medical devices under the UK Medical Devices Regulations 2002 (as amended). Devices like the Abbott FreeStyle Libre, which is also available on NHS prescription for eligible patients with diabetes, are approved as aids for glucose monitoring in people with diabetes. They are not approved as metabolic optimisation tools for healthy adults. That distinction matters, because it means the claims made in wellness marketing sit outside the scope of what regulators have actually evaluated.

    The Advertising Standards Authority has already taken action against some health brands for making unsubstantiated claims. If you see a CGM subscription service promising it will help you “unlock fat burning” or “end energy crashes”, those are marketing claims, not clinical findings. Treat them accordingly.

    What does the research actually say about CGM for non-diabetics?

    The honest answer is: not much, and what exists is mixed. A 2023 study published in Nature Medicine (the Stanford-led DIETFITS follow-up work examining glycaemic variability) found that individual glucose responses to food varied considerably even among healthy people. That finding has been heavily used by CGM companies to argue that personalised nutrition guidance, informed by your own glucose data, produces better outcomes. But the leap from “glucose responses vary” to “wearing a CGM will improve your health or performance” is not supported by robust clinical trials in non-diabetic populations.

    A 2024 review in The Lancet Diabetes & Endocrinology noted that while CGMs generate large volumes of data in healthy individuals, there is currently insufficient evidence that acting on that data produces meaningful improvements in weight, cardiovascular risk, or athletic performance compared to established dietary approaches. The reviewers called for randomised controlled trials in non-clinical populations before widespread recommendation. We are still waiting for those trials.

    In healthy adults, blood glucose generally stays within a tight range regardless of what you eat. The dramatic-looking spikes that some CGM apps flag as alarming are often physiologically normal. A glucose rise after a bowl of porridge is not the same thing as the sustained hyperglycaemia seen in diabetes. Yet some CGM coaching services treat any excursion above a certain threshold as a problem to be solved through dietary restriction, which risks encouraging unnecessary anxiety about food.

    Smartphone showing blood glucose data relevant to CGM for non-diabetics tracking
    Photo by Tessy Agbonome on Pexels

    The performance claims deserve particular scrutiny

    Supersapiens, which markets its CGM service specifically to athletes, has attracted a following among cyclists, runners, and endurance competitors. The premise is that knowing your glucose in real time helps you fuel more precisely during training and racing. I find this genuinely interesting as a concept. Sports nutrition is a field where small margins matter, and glucose management during prolonged effort is well established as important.

    But “interesting concept” and “proven performance advantage” are different things. The peer-reviewed evidence base for CGM-guided fuelling in non-diabetic athletes is thin. Most sports nutrition guidance from bodies like the British Dietetic Association still centres on periodised carbohydrate intake based on training load and intensity, not real-time glucose response. Until we have properly controlled studies showing CGM-guided fuelling outperforms standard evidence-based nutrition practice in healthy athletes, the claims should be held lightly.

    Who might genuinely benefit from CGM data outside a diabetes diagnosis?

    This is where it gets more constructive. There are populations where CGM use in non-diabetic contexts has real clinical logic. People with prediabetes, those with a strong family history of type 2 diabetes, women with a history of gestational diabetes, and people with polycystic ovary syndrome (PCOS) all have elevated risk profiles where understanding glucose variability has some clinical rationale. For these groups, the conversation belongs with a GP or endocrinologist, not a wellness app.

    If you are curious about your metabolic health and you fall into one of these categories, that is worth raising with your NHS GP rather than self-directing via a subscription service. Given how the NHS waiting list pressures are already pushing people towards self-diagnosis, the last thing anyone needs is another layer of uninterpreted data creating unnecessary worry or driving inappropriate dietary restriction.

    The supplement and nutrition industry connection

    CGM services rarely exist in isolation. Most are bundled with dietary coaching, meal planning, and frequently supplements. Once a company has your glucose data, they have a compelling hook for selling you products designed to “smooth” your glucose response: berberine, cinnamon extract, chromium, and various fibre supplements appear regularly in CGM-adjacent marketing. Some of these have limited supporting evidence in clinical populations; most are sold well beyond what the evidence justifies.

    I came across a comparison tool from Nusan whilst researching how UK consumers are trying to make sense of health product claims more broadly, and the pattern is consistent: the supplement market tends to rush into spaces where clinical data is incomplete and consumer interest is high. CGM is no different. If you are being told by a CGM service that you need specific supplements based on your glucose patterns, ask for the clinical evidence before spending your money.

    This also connects to what I have written previously about the supplement boom needing more scrutiny in the UK generally. The CGM category is just the newest entry point.

    What about Zoe specifically?

    Zoe is the highest-profile CGM-adjacent wellness company in the UK, co-founded by Professor Tim Spector of King’s College London. Their programme involves a CGM, gut microbiome testing, and personalised dietary recommendations. Zoe has published peer-reviewed research, including a 2024 paper in Nature Medicine suggesting their personalised dietary programme outperformed standard healthy eating guidance for certain metabolic markers. That is genuinely more rigorous than most commercial wellness programmes.

    But even here, the research involves Zoe’s own programme participants and has not yet been independently replicated at scale. The scientific advisory involvement does not mean every claim in their marketing is substantiated. The NHS guidance on CGM remains focused on people with diabetes and does not endorse its use in healthy adults for weight or performance management.

    The psychological risk nobody talks about

    One concern I have not seen discussed enough is the psychological effect of continuous metabolic monitoring on people who do not have a clinical reason for it. Wearing a device that generates a new data point every few minutes, and then receiving alerts about “high” readings, could reasonably increase health anxiety in some people. There is also a risk of what researchers call “orthorexic” thinking: becoming so focused on optimising food choices around a metric that eating becomes a source of stress rather than pleasure or nourishment.

    If you are already prone to metabolic health anxiety or have a history of disordered eating, a CGM subscription service is not where I would start. Talk to your GP first.

    The bottom line

    CGMs are legitimate, well-engineered medical devices. The data they produce is real. For people managing diabetes, they are genuinely life-changing. For healthy adults with no clinical indication, the evidence that wearing one and adjusting behaviour accordingly produces meaningful health outcomes is currently insufficient to justify the cost, the anxiety, or the supplement spending that tends to follow. The commercial ecosystem around CGM for non-diabetics has moved considerably faster than the science. That gap is worth understanding before you subscribe.

    Frequently Asked Questions

    Can I get a CGM on the NHS if I don't have diabetes?

    Generally no. NHS CGM prescriptions are reserved for people with type 1 diabetes and some people with type 2 diabetes who meet specific clinical criteria. If you are interested in CGM for metabolic monitoring without a diabetes diagnosis, you would need to purchase it privately through a commercial service.

    Is it safe to use a CGM if I am not diabetic?

    The devices themselves are safe for most people; they involve a small sensor inserted just under the skin. The main risks for non-diabetic users are psychological, including health anxiety from misinterpreting normal glucose fluctuations, and financial, given the ongoing subscription costs. If you have a history of disordered eating, speak to a GP before using one.

    Do CGMs actually help with weight loss in healthy people?

    Current evidence does not support CGM use as an effective weight loss tool in people without diabetes or prediabetes. A 2024 Lancet review found insufficient evidence that acting on CGM data in healthy adults produces better weight outcomes than established dietary approaches. The personalised nutrition framing is compelling, but the clinical proof is not there yet.

  • AI Self-Diagnosis in the UK: What Happens When Patients Use Chatbots to Interpret Their Symptoms

    AI Self-Diagnosis in the UK: What Happens When Patients Use Chatbots to Interpret Their Symptoms

    Something has quietly shifted in how people in the UK engage with their health before they ever book a GP appointment. Rather than reaching for NHS 111 or typing symptoms into a search engine, a growing number of patients are having full-length, back-and-forth conversations with large language models, asking things like “could this be lupus?” or “should I be worried about this lump?” The results of those conversations are shaping what people believe, how urgently they act, and in some cases, what they tell their doctors when they eventually do get a consultation.

    AI self-diagnosis in the UK is not a fringe behaviour any more. A 2024 survey by the Nuffield Trust found that digital health tools, including AI-driven ones, were being used by patients in ways that outpaced any formal guidance on appropriate use. The question worth asking plainly is: how accurate is AI when it comes to interpreting human symptoms, and who is responsible when it gets it wrong?

    Woman using a smartphone to look up symptoms, illustrating AI self-diagnosis UK behaviour
    Photo by Laura James on Pexels

    What the research actually shows about chatbot diagnostic accuracy

    The headline findings on accuracy are mixed, and I’d argue the nuances matter more than the averages. A 2023 study published in JAMA Internal Medicine found that ChatGPT performed reasonably well on standardised clinical vignettes, ranking the correct diagnosis in its top three suggestions around 72% of the time. Sounds reassuring until you consider what the other 28% of cases looked like, and that real patients presenting to chatbots rarely match the clean framing of a clinical vignette.

    The more concerning pattern is directional confidence. Large language models, by design, generate fluent, authoritative-sounding text. A patient asking about chest tightness and fatigue might receive a response that confidently prioritises anxiety or acid reflux, without adequately surfacing the possibility of cardiac involvement. The model is not lying. It is doing exactly what it was trained to do: produce plausible, coherent output weighted by probability. But medicine is full of low-probability, high-consequence events that probability-weighted systems systematically underweigh.

    Research from University College London published in early 2025 looked specifically at patients who had used AI tools to interpret symptoms before a GP visit. Around 34% had formed a strong prior belief about their diagnosis before the consultation, and GPs reported that challenging those beliefs added meaningful time pressure to already stretched appointments. This is a practical problem with systemic consequences.

    Why certain symptom types carry higher risk

    Not all AI symptom checking carries equal risk. For common, self-limiting conditions, a chatbot telling someone they likely have a cold or mild gastroenteritis is probably harmless, and might even reduce unnecessary demand on NHS services. The risk concentrates in two areas.

    First, rare or atypical presentations of serious conditions. Symptoms of conditions like ME/CFS are notoriously diffuse and overlap with dozens of other diagnoses. An AI trained predominantly on mainstream clinical literature may push patients toward more common explanations and away from the right one. Second, mental health presentations. A patient describing low mood, cognitive fog, and fatigue might be told by a model that they sound stressed or sleep-deprived, when what they are experiencing is a prodrome of something requiring clinical assessment.

    GP consultation desk representing the clinical encounter following AI self-diagnosis in the UK
    Photo by Thirdman on Pexels

    There is also a particular problem with supplement and lifestyle recommendations that chatbots often attach to their diagnostic suggestions. A patient convinced they have a magnesium deficiency based on an AI conversation might spend weeks self-treating before discovering the real cause of their symptoms. The information itself is not necessarily wrong; the sequencing and framing almost always is.

    Where UK regulation currently stands

    This is where things get genuinely complicated. The MHRA’s updated framework for software as a medical device, which I’ve covered in the context of AI-driven health apps, applies to tools that are intended to be used for a medical purpose. The word “intended” is doing enormous work here. A general-purpose large language model, such as ChatGPT or Google’s Gemini, is not marketed as a diagnostic tool. Its developers explicitly disclaim medical use. That puts it largely outside the MHRA’s current regulatory perimeter, even when patients are using it for exactly that purpose.

    The Care Quality Commission, which regulates health services in England, has similarly limited reach here. CQC oversight applies to registered providers. A software product that refuses to call itself a health service sits in a gap that existing frameworks were never designed to address. The CQC’s 2025 annual report on innovation in health acknowledged the issue without offering a concrete regulatory pathway.

    NHS England has issued informal guidance encouraging patients to use NHS 111 or the NHS App as first points of contact for health concerns, and to treat AI chatbots as general information tools rather than diagnostic ones. That guidance is reasonable. The gap between reasonable guidance and actual patient behaviour is, as usual, wide. You can find the NHS’s current position on digital health tools at nhs.uk.

    What this means for patients navigating it now

    I’d be dishonest if I said AI chatbots have no value in a health context. They can help patients organise and articulate symptoms more clearly before an appointment. They can provide useful background on how a diagnosis works, what questions to ask a GP, or what a prescribed medication does. Used in that framing, as a preparation tool rather than a diagnostic oracle, they are genuinely useful.

    The problem is that the interface encourages a different behaviour. Asking a chatbot “what is causing my headaches?” and receiving a structured, confident three-paragraph response does not feel like reading a general information article. It feels like receiving a personalised assessment. That experiential difference matters, because it shapes how strongly a person holds the belief that follows.

    My honest take: treat any chatbot response about personal symptoms the same way you’d treat advice from a well-read friend with no medical training. Useful context, possibly, but not a clinical opinion. If you are researching health information online more broadly and want better tools for evaluating what comes up in search results, resources like expert free SEO tools can at least help you understand which sources rank highly and why, though that is a separate skill from evaluating medical credibility.

    The harder structural question

    Underneath the accuracy debate is a more uncomfortable truth: patients in the UK are turning to AI in part because accessing a GP in 2026 is genuinely difficult. NHS England data shows average waiting times for a routine GP appointment still exceeding two weeks in many areas. When someone is worried and can’t get timely access to a clinician, a chatbot that responds immediately feels better than nothing. That is not a technology failure. That is a capacity failure that technology is filling imperfectly.

    Regulatory frameworks that treat AI symptom use as purely a tech problem to be policed will miss this. The more productive frame is: how do we design AI health tools that are accurate about their own limitations, that actively route patients toward appropriate care rather than substituting for it, and that work within NHS pathways rather than around them? That design challenge is solvable. The regulatory appetite to mandate it is still catching up.

  • The Real-World Accuracy of AI Coding Assistants in 2026: What UK Developers Are Actually Finding

    The Real-World Accuracy of AI Coding Assistants in 2026: What UK Developers Are Actually Finding

    There is a lot of noise around AI coding assistants right now, and most of it comes from two camps: developers who think these tools are transformative, and developers who are quietly cleaning up the mess they leave behind. The truth, as ever, sits somewhere between the two. I’ve spent time with the benchmarks, the independent surveys, and the candid developer forums, and the picture that emerges is more nuanced than the marketing suggests. For UK developers in particular, the question of AI coding assistant accuracy in 2026 deserves a straight answer.

    UK developer reviewing AI coding assistant accuracy on a laptop screen in 2026
    Photo by Lukas Blazek on Pexels

    What the benchmarks actually measure

    Most headline figures around tools like GitHub Copilot, Amazon CodeWhisperer, and Cursor come from vendor-commissioned tests or cherry-picked HumanEval scores. HumanEval, developed by OpenAI, tests whether a model can write short Python functions from a docstring. It is a useful starting point, but it bears little resemblance to what a developer at a UK fintech or NHS digital team actually deals with day-to-day.

    The Stack Overflow Developer Survey 2025 found that roughly 76% of respondents were using or planning to use AI coding tools, but satisfaction scores told a different story: just 43% said the output was reliable enough to merge without careful review. That gap matters. Accepting code suggestions is easy. Trusting them is another thing.

    Independent benchmarks like SWE-bench, which tests whether models can actually resolve real GitHub issues in established codebases, show much lower success rates than HumanEval implies. The best-performing models in early 2026 resolve around 40-50% of SWE-bench tasks. That is progress, but it also means that in roughly half of real-world scenarios, the model either fails outright or introduces a solution that compiles but behaves incorrectly under edge cases.

    Where AI coding assistants genuinely save time

    This is not a dismissal. These tools do save time, and I’d be dishonest not to say so clearly. The categories where the gains are real and consistent are fairly specific.

    Boilerplate code is the obvious one. Writing repetitive scaffold code, setting up test files, generating getter and setter methods, or producing the skeleton of a REST endpoint: these tasks are well-suited to autocomplete-style suggestions, and the time savings are measurable. Developers in the Stack Overflow survey reported saving between one and four hours per week on routine tasks, with the higher estimates coming from those working in strongly typed languages like TypeScript where the context is clearer.

    Documentation generation is another genuine win. Getting a model to produce a first-draft docstring or README section is quicker than writing one from scratch, and even if you edit it afterwards, you are starting from something rather than a blank page.

    Learning unfamiliar syntax is where I personally find these tools most useful. If you are moving between languages, or working with an API you have not touched before, having an assistant that can surface working examples in context is faster than cycling between documentation tabs.

    Where the errors creep in and why they are costly

    The errors that AI coding assistants introduce are rarely catastrophic and obvious. They tend to be subtle. That is precisely what makes them expensive.

    Security vulnerabilities are the most serious concern. A 2025 study from Stanford University found that developers using AI assistants were significantly more likely to introduce security flaws than those coding without them, partly because the suggestions feel authoritative and get merged without the scrutiny a human-written block would receive. Common patterns include outdated cryptographic practices, SQL injection vectors in generated query strings, and improper input validation. For UK developers working in regulated sectors, whether financial services under FCA rules or health data environments governed by the ICO, these are not abstract risks.

    Logic errors in conditional branches are another consistent problem. The model may generate code that handles the happy path correctly but silently fails on null inputs, empty arrays, or timezone edge cases. I have seen this described again and again in developer forums: the code runs, the tests pass, and the bug surfaces three weeks later in production.

    Hallucinated library references also remain a genuine issue. Tools occasionally suggest method calls that do not exist in the version of a library the project actually uses, which wastes debugging time and can confuse junior developers who assume the suggestion is reliable.

    There is also a subtler risk around over-reliance. Developers who lean heavily on AI suggestions for code they do not fully understand are producing code they cannot confidently maintain. The growing role of AI in regulated environments makes this a genuine governance issue, not just a craft one. For teams building tools in health or finance, code ownership and auditability matter.

    The skills and habits that change the outcome

    The developers who report the best outcomes with AI coding assistants share a few consistent habits. They treat suggestions as a starting point rather than a final draft. They run the generated code through their existing test suites before accepting it. They stay especially sceptical with any suggestion that touches authentication, data persistence, or external API calls.

    Prompt quality matters more than most vendors acknowledge. A vague instruction produces a vague suggestion. Developers who invest time in writing precise, contextual prompts, including relevant type signatures, examples of existing patterns in the codebase, and explicit constraints, consistently get better output. This is a skill, and it takes time to develop.

    Team culture plays a role too. Organisations where code review is taken seriously, where junior developers are encouraged to question AI-generated suggestions rather than defer to them, tend to catch errors earlier. The AI assistant does not remove the need for good engineering practices; it actually raises the stakes on having them.

    What this means for UK developers practically

    If you are a UK developer or engineering manager deciding how to integrate these tools into your workflow, the honest advice is to use them, but with explicit guardrails. Establish a team norm that AI-generated code requires the same review rigour as code from any other source. Consider adding AI-specific items to your pull request checklist: has this been tested on edge cases, does it reference a real and current API, has anyone checked for security patterns the model might have assumed incorrectly.

    The productivity gains are real enough to justify adoption. The accuracy gaps are real enough to justify caution. Those two things are not in contradiction. And for anyone thinking about the broader pattern here, the same principle applies whether you are assessing an AI coding assistant or thinking about workplace tools that promise more than the evidence supports: the question is never whether a tool does something useful, but whether the specific claims match the specific evidence.

    UK developers are, by most accounts, adopting these tools at roughly the same pace as their counterparts in the US and Germany. The constraint is not enthusiasm; it is informed scepticism, and that is worth holding onto. The tools will improve. The discipline of reviewing what they produce should improve alongside them.

  • The MHRA’s New Approach to AI Medical Devices: What It Means for the Apps Diagnosing and Monitoring You

    The MHRA’s New Approach to AI Medical Devices: What It Means for the Apps Diagnosing and Monitoring You

    If you have ever used an app to log symptoms, interpret a blood glucose reading, or get a personalised health recommendation, there is a real chance you have interacted with something that the MHRA considers a medical device. Most people have no idea that category exists, let alone what it requires of the companies building these tools. The MHRA AI medical device regulation UK framework is one of the most consequential and least-discussed shifts in British healthcare right now, and it is worth understanding in plain terms.

    NHS clinician reviewing a health app on a tablet, relevant to MHRA AI medical device regulation UK
    Photo by Tima Miroshnichenko on Pexels

    What the MHRA actually classifies as a medical device in software

    The MHRA’s Software and AI as a Medical Device (SaMD) framework draws a clear line: if a piece of software is intended to be used for a medical purpose, it falls under regulatory oversight. That includes diagnosing a condition, predicting deterioration, monitoring a chronic illness, or informing a treatment decision. General wellness apps, step counters, and sleep trackers that make no medical claims sit outside this definition. The moment an app says it can detect an arrhythmia, flag early signs of diabetic retinopathy, or suggest a medication adjustment, it crosses the line.

    The MHRA uses a risk-based classification system, running from Class I (lowest risk) up to Class III (highest risk). Most AI-driven diagnostic tools land in Class IIa or IIb, which requires a Conformity Assessment carried out by a UK Approved Body. Developers must maintain a technical file that includes clinical evidence, details of the algorithm’s training data, and a post-market surveillance plan. For Class IIb and above, that plan must be proactive, not reactive. The MHRA’s published guidance on AI as a medical device lays out the current framework in detail, including the 2024 roadmap updates that are now being implemented.

    How the UK framework differs from what came before

    Prior to the UK leaving the EU’s regulatory orbit, software developers could rely on CE marking under the EU MDR. Since then, the MHRA has been building its own regime, and the UK Conformity Assessed (UKCA) mark is now the requirement for Great Britain. The MHRA published its “Software and AI as a Medical Device Change Programme” in stages from 2022 onwards, and the 2026 position is that manufacturers must demonstrate what the regulator calls “safe and effective performance” throughout the entire product lifecycle, not just at launch.

    What makes this genuinely different for AI tools is the acknowledgement of continuous learning. An algorithm that updates itself based on new patient data is not the same product it was six months ago. The MHRA’s framework addresses this with the concept of predetermined change control plans, meaning developers must specify in advance what kinds of changes they anticipate the model making and get those parameters approved. That is a meaningful shift from how software regulation has historically worked, where a product was reviewed once and then largely left alone.

    What developers are required to demonstrate

    In practical terms, a company seeking approval for an AI-driven diagnostic app needs to show several things. Clinical validation is the big one. The algorithm must be tested against the population it will actually serve, not just the dataset it was trained on. This matters enormously because training data bias is a real and documented problem. An ECG algorithm trained predominantly on data from white men in their forties will perform differently on South Asian women in their sixties, and the MHRA expects developers to account for that explicitly.

    Beyond clinical evidence, developers must provide transparency about how the algorithm reaches its outputs. That does not necessarily mean publishing the model weights, but it does mean producing documentation that allows a clinician to understand the basis for a recommendation. There is also a usability requirement: the interface must be tested with real users to ensure the information is presented in a way that supports safe decision-making rather than creating confusion or misplaced confidence.

    Post-market surveillance is where many smaller developers struggle. They are required to collect real-world performance data after launch, track adverse events, and submit periodic safety update reports. For a well-funded medtech company this is manageable. For a startup with a lean team, the administrative burden is substantial, and I would argue that is partly why so many health apps quietly drop medical claims from their marketing to avoid classification altogether.

    Where the gaps still are

    The framework is improving, but several gaps remain. Enforcement is the most obvious. The MHRA does not have the resource to proactively audit every app marketed to UK users. In practice, most enforcement is complaint-driven, which means a poorly validated tool can operate for months or years before anyone formally challenges it. This is particularly concerning given how many AI-driven health apps are marketed directly to consumers via app stores, where the App Store and Google Play review processes have no meaningful medical device checkpoint.

    International products are another complexity. An app developed in the US, Canada, or elsewhere, marketed to UK users, technically requires UKCA marking if it makes medical claims. In reality, many do not have it, and the MHRA’s ability to pursue overseas developers is limited. Given the rapid growth of AI wearables and monitoring platforms, this is not a small edge case.

    There is also a definitional grey area around generative AI tools used in health contexts. A large language model that answers questions about symptoms is almost certainly providing information, not diagnosis, and therefore sits outside SaMD classification. But that line is blurring. If a tool consistently recommends specific courses of action based on symptom inputs, at what point does it become a diagnostic aid? The MHRA has signalled it is watching this space closely, but firm guidance is still in progress.

    What this means if you use health apps

    As a consumer, you have limited visibility into whether a given app has gone through proper regulatory review. A few practical checks help. If an app makes medical claims, look for a UKCA mark or CE mark in the product documentation (many legitimate medical device apps display this in their terms or in app store listings). Check whether the company publishes a clinical evidence summary. If they do not, ask yourself why. Apps that have genuinely validated their algorithms tend to say so clearly because it is a genuine competitive advantage.

    The MHRA maintains a register of certified medical devices, though navigating it is not exactly user-friendly. It is worth knowing that the NHS App and NHS-commissioned digital tools are held to a higher standard, going through NHS England’s Digital Technology Assessment Criteria (DTAC) as well as MHRA requirements. That dual-layer review gives NHS-approved tools more credibility than consumer apps operating solely on the basis of self-declaration.

    The regulatory picture for AI in health is improving, but slowly. The gap between what technology can claim and what it has rigorously demonstrated is still wide in places. For anyone relying on an app to inform decisions about their health, that gap is worth keeping front of mind.

    If you are thinking about how digital health tools intersect with NHS systems and patient data, the questions raised about NHS diagnostic pathways and patient experience are closely connected. And the broader question of how UK adults manage their health with limited NHS access is part of why so many people turn to consumer apps in the first place. Understanding what those apps are, and are not, legally required to prove is a reasonable starting point for anyone trying to make informed choices.

  • Deepfakes, Synthetic Media and the UK Law: What Your Rights Actually Are in 2026

    Deepfakes, Synthetic Media and the UK Law: What Your Rights Actually Are in 2026

    Synthetic media has moved fast. Tools that once required professional equipment and weeks of processing time now produce convincing fake video or audio in minutes, on a standard laptop, sometimes from a single photograph. The results range from harmless novelty to something far more damaging: fake intimate images, fabricated statements attributed to real people, and manipulated footage used to harass or defraud. Understanding what deepfake law UK 2026 actually covers, and where it still falls short, matters more than ever.

    Woman reviewing content on a laptop relating to deepfake law UK 2026 protections
    Woman reviewing content on a laptop relating to deepfake law UK 2026 protections

    What the Law Currently Covers

    The UK’s legal response to non-consensual deepfakes has developed through a patchwork of legislation rather than a single comprehensive statute. That patchwork has grown significantly in the past two years.

    The Online Safety Act 2023 was the first major piece of legislation to target synthetic intimate imagery directly. It created a specific criminal offence for sharing deepfake pornography without consent, carrying an unlimited fine. Importantly, intent to cause distress no longer needs to be proven for sharing; the absence of consent is sufficient. Prosecutors had previously struggled to bring cases under older harassment or malicious communications laws because those required demonstrating deliberate harm.

    Then came the Criminal Justice Bill amendments passed in 2024, which went a step further: creating a new offence for the creation of sexually explicit deepfakes, even if they are never shared. That was a meaningful shift. Prior to it, a person could fabricate intimate imagery of someone they knew, keep it on a device, and face no criminal sanction whatsoever, provided it never left their possession. The creation offence closed that particular gap.

    Beyond intimate imagery, existing law still provides some protection. The Malicious Communications Act 1988 and the Communications Act 2003 can cover deepfakes used to harass. The Fraud Act 2006 applies where synthetic media is used to deceive someone for financial gain. Defamation law, governed by the Defamation Act 2013, remains relevant where a deepfake causes serious reputational harm, though civil defamation claims are expensive to pursue and legal aid is rarely available.

    Where the Gaps Still Exist

    Despite those advances, the current framework has real weaknesses. The law as it stands focuses heavily on intimate imagery. Deepfakes used for other purposes, such as fabricating political statements, manipulating job applicants, or generating fake testimonials from real people for commercial gain, sit in murkier legal territory.

    Smartphone held in hand illustrating platform reporting tools relevant to deepfake law UK 2026
    Smartphone held in hand illustrating platform reporting tools relevant to deepfake law UK 2026

    There is also the question of enforcement. Identifying who created a synthetic image or video is technically challenging. Many tools are accessed through overseas platforms, and perpetrators can use anonymising technology. Even where a suspect is identified, building evidence to the criminal standard of proof remains difficult. The Internet Watch Foundation reported in 2025 that AI-generated child sexual abuse material had risen sharply year on year, illustrating both the scale of the problem and the limits of reactive enforcement.

    Platform liability is another open question. The Online Safety Act places duties on platforms to remove illegal content, but proactive detection of synthetic media is far from perfect. Deepfake detection tools exist but are not foolproof, and platforms vary considerably in how seriously they implement their obligations. Ofcom, which regulates platform compliance under the Act, has powers to issue fines of up to £18 million or 10% of global turnover, but enforcement actions take time.

    Consent and attribution in non-intimate contexts remain largely unaddressed by statute. If someone uses your voice, cloned from publicly available recordings, to record a fake interview or a commercial endorsement, the legal route is uncertain. You might pursue it through data protection law via the ICO, or through passing off under common law if it implies a commercial endorsement, but neither route is straightforward.

    What the Health and Wellbeing Angle Looks Like

    It is worth being direct about why this matters beyond the obvious. Non-consensual deepfakes cause serious psychological harm. Research cited by the mental health charity Mind consistently links online harassment and image-based abuse to anxiety, depression, post-traumatic stress, and in some cases suicidal ideation. The harm is not abstract. For victims, particularly women and younger adults who are disproportionately targeted, the damage to mental health can be lasting and severe.

    That human cost is precisely why getting the legal framework right matters. Legislation that only partially covers the problem, or that exists on paper but is difficult to enforce, does not offer meaningful protection to the people most at risk.

    Practical Steps You Can Take Right Now

    If you believe you are a victim of a non-consensual deepfake, here is what is worth knowing in practical terms.

    Report to the police. If the content is sexually explicit, it may constitute a criminal offence under the Online Safety Act. Keep records of everything: screenshots with timestamps, URLs, any messages you have received. Your local police force can refer cases to specialist units.

    Contact the platform directly. All major platforms operating in the UK are subject to the Online Safety Act’s takedown duties. Report the content using the platform’s reporting mechanism. Follow up if nothing happens. Document your report.

    Use specialist support organisations. The Revenge Porn Helpline (run by SWGfL) handles synthetic media cases and can assist with takedown requests across multiple platforms. Their service is free. Refuge and Galop also provide support where the content is part of a broader pattern of abuse.

    Consider a data protection complaint. If identifiable personal data (including your image or voice) has been processed unlawfully, you can report to the ICO. This route is slower but can result in formal enforcement action against a platform.

    Consult a solicitor. Civil routes, including injunctions and defamation claims, remain available. Some solicitors specialise in online abuse and may take cases on a conditional fee arrangement. The Law Society’s solicitor finder is a good starting point.

    The Direction of Travel

    The UK government has signalled further legislative work on synthetic media, particularly around electoral integrity and commercial fraud. The Law Commission has also been asked to review how existing defamation and privacy law handles AI-generated content. Progress is real but incremental.

    Understanding deepfake law UK 2026 means accepting two things simultaneously: the framework is substantially stronger than it was three years ago, and it still leaves meaningful gaps for victims whose situation falls outside intimate imagery. Knowing where you stand, and what tools exist to help, is the most honest place to start.

    Frequently Asked Questions

    Is it illegal to create a deepfake of someone in the UK?

    It depends on the content. Creating sexually explicit deepfakes of a real person without their consent is now a criminal offence in the UK following 2024 legislation. Creating non-intimate synthetic media, such as a fake speech or fabricated interview, is not automatically criminal, though it may attract liability under fraud, harassment, or defamation law depending on how it is used.

    What can I do if someone has shared a deepfake image of me online?

    Report it to the platform immediately using their reporting tools, as platforms regulated under the Online Safety Act have legal takedown duties. You should also report the matter to the police, particularly if the content is sexually explicit, and contact the Revenge Porn Helpline (run by SWGfL) which handles synthetic media cases and can assist with removals across multiple platforms free of charge.

    Does UK law cover deepfakes used in fraud or financial scams?

    Yes, the Fraud Act 2006 applies where synthetic media is used to deceive someone for financial gain, such as a fake video call impersonating a senior executive to authorise a bank transfer. These cases are increasingly common and are investigated by Action Fraud and specialist police units.

    Can I take someone to court over a non-sexual deepfake that damaged my reputation?

    Potentially, yes. The Defamation Act 2013 applies where false content causes serious reputational harm, and a deepfake fabricating statements attributed to you could meet that threshold. However, civil defamation claims are costly and legal aid is rarely available, so getting specialist legal advice first is strongly recommended.

    What is Ofcom's role in regulating deepfake content on UK platforms?

    Ofcom regulates platform compliance with the Online Safety Act 2023, which includes duties to remove illegal content such as non-consensual synthetic intimate imagery. Ofcom can issue fines of up to £18 million or 10% of global annual turnover for serious breaches, though formal enforcement proceedings typically take several months to conclude.

  • AI in NHS Diagnostics: Where the Technology Is Already Working and Where It Falls Short

    AI in NHS Diagnostics: Where the Technology Is Already Working and Where It Falls Short

    The promise has been repeated so often it risks becoming background noise: artificial intelligence will transform NHS diagnostics, catch cancers earlier, cut waiting lists, and free up clinicians to focus on what humans do best. Some of that is already happening. Some of it remains, at best, a well-funded aspiration. Getting an honest picture requires looking past the press releases and into the actual deployment data, trial results, and independent evaluations that have emerged over the past two to three years.

    NHS radiologist reviewing chest X-rays using AI diagnostic tools in a hospital reading room
    NHS radiologist reviewing chest X-rays using AI diagnostic tools in a hospital reading room

    Where AI diagnostic tools in the NHS are genuinely being used

    The clearest real-world deployment sits in radiology. NHS England’s AI and Digital Transformation programme has funded over 86 AI imaging tools through its Accelerating AI in Imaging initiative, and a significant number of these are now in routine clinical use across NHS trusts. The most mature application is chest X-ray triage. Tools like Annalise.ai and Behold.ai analyse chest radiographs and flag potential abnormalities, prioritising urgent cases in radiologist worklists. At University Hospitals of Leicester NHS Trust, a 2023 evaluation found that AI triage reduced the median time to report urgent chest X-rays by around 11 days. That is not a marginal gain.

    Mammography screening is another area where the evidence is solidifying. A large UK trial published in The Lancet Oncology in 2023 found that AI could safely act as a second reader in breast cancer screening, matching the performance of a human radiologist. The NHS breast screening programme, which handles millions of mammograms annually, is now piloting AI double-reading in several trusts, potentially freeing up radiologist time at a moment when the workforce is under severe pressure. NHS England’s own 2025 report on AI in imaging cited projected capacity gains of up to 30% in some radiology departments where these tools have been integrated properly.

    Pathology and the harder problem of deployment at scale

    Pathology is where the technology is promising but the implementation story is messier. Computational pathology, which uses AI to analyse digitised tissue slides, has shown genuine clinical value in research settings. Paige.ai, for instance, received regulatory clearance in the UK for prostate cancer detection on digital slides. Studies have shown AI can identify clinically significant prostate cancer with sensitivity comparable to experienced pathologists, and in some cases catching cases that were initially missed.

    Pathologist preparing tissue slides for digital scanning as part of AI diagnostic tools evaluation in NHS pathology
    Pathologist preparing tissue slides for digital scanning as part of AI diagnostic tools evaluation in NHS pathology

    The problem is infrastructure. The majority of NHS pathology labs are not yet fully digitised. Whole slide imaging requires significant investment in scanners, storage, and connectivity, and many trusts are still in transition. NHS England’s National Pathology Programme has set targets for full digital pathology rollout, but as of early 2026, adoption remains uneven across the country. Without digitisation, AI pathology tools simply cannot run. The gap between what the technology can do in a well-resourced trial setting and what it can do in the average NHS lab is still wide, and it is mainly a structural problem rather than a clinical one.

    What the independent evaluations actually say

    The National Institute for Health and Care Excellence (NICE) has been publishing evidence reviews on AI medical devices through its Early Value Assessment programme. These assessments are worth reading carefully because they are specifically designed to distinguish between devices that have robust evidence and those that have been deployed on the basis of theoretical benefit. Several AI diagnostic tools that entered NHS use with considerable fanfare have received NICE guidance noting that the evidence base remains limited or that further real-world data is needed before confident recommendations can be made.

    The Topol Review, conducted by Dr Eric Topol and published by Health Education England, laid out a credible framework for AI in clinical practice, but it was also honest that workforce preparation, data governance, and integration with existing electronic health record systems were the real bottlenecks, not the algorithms themselves. That observation has aged well. A 2025 review by the Health Foundation found that NHS trusts frequently struggled to integrate AI tools with their existing IT infrastructure, and that time spent on implementation and staff training often exceeded original projections considerably.

    For context, the NHS’s broader digital maturity varies dramatically by trust. A foundation trust in London with mature electronic patient records and a strong informatics team is in a very different position from a district general hospital still reliant on paper-based processes in some departments. AI diagnostic tools tend to perform best where digital foundations are already solid, which means the hospitals that arguably need capacity relief the most are often the least equipped to benefit quickly.

    The clinical safety and bias questions that still need answering

    One area that deserves more public attention is algorithmic bias. Most AI diagnostic tools were trained on datasets that over-represent certain demographic groups, and there is credible evidence that performance can differ across ethnicities, age groups, and body types. NHS England’s AI and Digital Transformation team has acknowledged this in its guidance and requires suppliers to provide disaggregated performance data. Whether trusts are consistently demanding and reviewing that data in practice is a different question.

    The Medicines and Healthcare products Regulatory Agency (MHRA) regulates AI diagnostic devices as software medical devices in the UK, and its evolving guidance framework is genuinely trying to keep pace with the technology. But post-market surveillance of AI tools, tracking how they perform once deployed in real NHS conditions rather than controlled trials, remains an underdeveloped area. A tool that performs well in a trial with carefully curated data may behave differently when processing the full messy complexity of real clinical practice. The NICE evidence reviews on AI and digital tools provide a useful starting point for anyone wanting to assess specific devices critically.

    The honest picture in 2026

    AI diagnostic tools in the NHS are not a fantasy. In radiology, particularly chest X-ray triage and mammography, there are real deployments with real evidence of clinical benefit. Pathology is following, more slowly, constrained by infrastructure rather than technology. The honest picture is one of genuine but uneven progress, with pockets of meaningful impact sitting alongside a much larger landscape of tools that are approved, funded, or trialled but not yet meaningfully integrated into routine care.

    The workforce question matters enormously here. AI tools are most useful when clinicians trust them enough to act on their outputs, which requires training, familiarity, and time, all things the NHS is chronically short of. There is also a governance dimension that does not get enough attention: ensuring that patient data used to train and run these systems is handled appropriately, that consent is meaningful, and that the commercial interests of AI vendors do not distort clinical decision-making. These are not reasons to be pessimistic about the technology. They are the actual work that needs to happen alongside it.

    On a lighter administrative note, if you work in NHS communications or health tech and you are sending bulk email updates about AI programmes and digital health initiatives, it is worth ensuring your messages actually reach inboxes. A free spam checker can quickly identify whether your outgoing emails are likely to be filtered before they reach clinicians or commissioners who need to see them.

    The technology is capable of making a genuine difference to diagnostic capacity and patient outcomes in the NHS. The evidence from the strongest deployments makes that clear. What is equally clear is that the gap between potential and consistent, equitable, system-wide benefit is still significant, and closing it is largely a leadership, infrastructure, and workforce challenge rather than an algorithmic one.

    Frequently Asked Questions

    Which AI diagnostic tools are currently approved for use in the NHS?

    A number of AI imaging tools are in active NHS use, including tools for chest X-ray triage, mammography second-reading, and diabetic eye screening. NICE’s Early Value Assessments and NHS England’s AI imaging programme list approved and piloted devices, though availability varies considerably by trust and region.

    Is AI replacing NHS radiologists and pathologists?

    No. Current AI diagnostic tools act as decision-support systems, flagging cases for review and prioritising worklists rather than making final clinical decisions. The intent is to support and extend the capacity of clinical staff, not to replace human judgement, particularly given ongoing NHS workforce shortages.

    How accurate are AI diagnostic tools compared to human clinicians?

    In specific, well-studied tasks such as mammography reading and chest X-ray triage, some AI tools have demonstrated performance comparable to experienced clinicians in controlled trials. However, real-world performance can differ from trial performance depending on data quality, patient demographics, and local implementation, so independent evaluation is important.

    Are there concerns about bias in NHS AI diagnostic systems?

    Yes, and this is an active area of scrutiny. Many AI tools were trained on datasets that do not fully represent the diversity of the UK population, which can lead to reduced accuracy for certain ethnic groups or demographic subsets. NHS England requires suppliers to provide disaggregated performance data, but monitoring in routine deployment remains inconsistent.

    Who regulates AI diagnostic devices used in the NHS?

    The MHRA (Medicines and Healthcare products Regulatory Agency) regulates AI diagnostic tools as software medical devices in the UK. NICE also provides clinical guidance and evidence assessments through its Early Value Assessment programme, which helps NHS trusts evaluate whether specific tools have sufficient evidence to justify adoption.

  • What the EU AI Act Means for the Apps and Services You Use Every Day

    What the EU AI Act Means for the Apps and Services You Use Every Day

    The EU AI Act has been quietly coming into force across 2025 and 2026, and most people have no idea what it actually does. That’s a problem, because it directly shapes the apps on your phone, the recommendation algorithms deciding what you watch, and the automated systems that can affect decisions about your finances or healthcare. The EU AI Act consumer impact is real, it’s already here, and it’s worth understanding in plain terms.

    This isn’t just a European issue either. UK readers should pay attention. Although the UK left the EU and has its own emerging AI governance framework, many of the products and platforms used daily by people in Britain are built by companies operating in the EU or selling into the European market. Those companies are now legally required to meet the Act’s standards, which means the changes trickle through to everyone using their products.

    Woman reviewing EU AI Act consumer impact disclosures on a laptop in a modern London office
    Woman reviewing EU AI Act consumer impact disclosures on a laptop in a modern London office

    What the EU AI Act Actually Is

    The EU AI Act is the world’s first comprehensive legal framework specifically for artificial intelligence. It came into force in stages, with the most significant provisions applying to high-risk AI systems from August 2026 onwards. The core idea is straightforward: the more potential harm an AI system can cause, the more oversight and transparency it must demonstrate.

    The Act divides AI into risk tiers. Unacceptable-risk systems, such as social scoring by governments or real-time biometric surveillance in public spaces, are outright banned. High-risk systems, covering areas like healthcare diagnosis tools, credit scoring, and recruitment software, face strict requirements around documentation, human oversight, and accuracy. Lower-risk systems, including most chatbots and content recommendation engines, must meet lighter transparency obligations. General-purpose AI models, like the large language models powering many consumer tools, face their own set of rules around transparency and capability disclosure.

    What Changes for You as a Consumer

    The most immediate change you’re likely to notice is disclosure. Under the EU AI Act, companies must clearly tell you when you’re interacting with an AI system rather than a human. That might sound obvious for a chatbot, but it applies to more subtle situations too: automated customer service systems, AI-generated content, and deepfake detection obligations. If a piece of content is AI-generated and could deceive you, the company behind it must label it as such.

    For health and wellness apps, which increasingly use AI to provide personalised advice, the rules become more significant. Apps making health recommendations that influence medical decisions are likely to fall into higher-risk categories. That means they’ll need to demonstrate their systems are accurate, properly documented, and subject to human oversight. Practically speaking, this should raise the floor on quality for health tech products sold or used across Europe.

    Smartphone showing health app with EU AI Act consumer impact transparency label
    Smartphone showing health app with EU AI Act consumer impact transparency label

    Credit decisions are another area where the EU AI Act consumer impact becomes concrete. If an AI system contributes to a decision about whether you get a loan, insurance, or a rental agreement, you now have a stronger right to an explanation of how that decision was made. This builds on existing rights under GDPR, but goes further in requiring the systems themselves to be auditable and contestable. The UK’s own Financial Conduct Authority has been watching these developments closely, and similar transparency pressures are building domestically.

    High-Risk Classifications and Why They Matter to You

    The high-risk category is where the Act has the most teeth. It covers AI systems used in:

    • Healthcare and medical devices
    • Education and vocational training (systems that assess students)
    • Employment and recruitment
    • Access to essential private and public services, including credit and insurance
    • Law enforcement and border control
    • Administration of justice

    If you’ve ever applied for a job through an automated screening platform, or been assessed by an algorithm for benefits eligibility, those systems now have to meet standards around bias testing, accuracy documentation, and human review. Companies deploying them must register them in an EU database, which is publicly accessible. That’s a meaningful accountability shift. You can actually look up whether a system affecting your life has been properly registered.

    The European Commission’s digital strategy pages maintain updated guidance on the Act’s scope and timelines if you want to go deeper into the official position.

    What About Chatbots and Everyday AI Tools?

    Most of the consumer AI tools people use daily, including writing assistants, productivity apps, and recommendation systems, sit in the lower-risk tiers. They aren’t banned and don’t face the same heavy compliance burden. But they do face transparency requirements. If you’re talking to an AI, it must be clear that you’re talking to an AI. If content has been generated by an AI in a way that could mislead you, it must be labelled.

    The rules around general-purpose AI models (GPAIs) are worth flagging separately. Large-scale models that power many consumer products now have obligations around publishing technical documentation, copyright policies, and, for the most powerful models, full adversarial testing and incident reporting. This means the underlying engines driving the tools you use are subject to transparency requirements you haven’t seen before.

    Does This Apply if You’re in the UK?

    Technically, the EU AI Act does not directly apply in the UK. Post-Brexit, the UK is developing its own approach through the AI Safety Institute and a sector-by-sector regulatory model, rather than a single overarching law. However, the practical EU AI Act consumer impact for UK users is significant for one simple reason: most major tech platforms operate across both markets and are not going to build two entirely separate product versions. They will comply with the stricter standard, which is the EU Act, and that compliance will extend to their UK-facing products too.

    This is sometimes called the Brussels Effect, and it’s exactly what happened with GDPR. UK consumers gained real privacy protections partly because the companies serving them were already complying with EU law. The same dynamic is likely to play out with AI regulation.

    What You Can Actually Do With This Information

    A few practical points worth holding onto. First, if an AI system makes a significant decision about you, whether in hiring, finance, or healthcare, you have growing rights to ask for an explanation. Exercise them. Second, if an app claims to give you medical or mental health guidance, ask how it classifies itself under the Act. Reputable providers should be able to answer. Third, watch for the labelling. When AI-generated content becomes ubiquitous, the labels required by the Act are one of the few honest signals left about what you’re actually looking at.

    The EU AI Act consumer impact is not a silver bullet. Enforcement is complex, and the rules are being tested in real time. But the direction of travel is clear: more transparency, more accountability, and more tools for ordinary people to understand and challenge the automated systems shaping their lives. That’s a genuine step forward, and it’s worth knowing about.

    Frequently Asked Questions

    Does the EU AI Act apply to people in the UK?

    Not directly, since the UK has its own regulatory approach. However, major tech companies serving both the EU and UK markets typically comply with the stricter EU standard, meaning UK consumers often benefit from the same protections in practice.

    What rights do I have if an AI makes a decision about me?

    If a high-risk AI system, such as one used in credit scoring, recruitment, or healthcare, makes or contributes to a significant decision about you, you have the right to request a meaningful explanation of how that decision was reached. Companies must also ensure a human review process is available.

    Which apps or services count as high-risk under the EU AI Act?

    High-risk categories include AI used in healthcare diagnostics, job recruitment screening, credit and insurance assessments, educational evaluation, and law enforcement. These systems face the strictest transparency and documentation requirements under the Act.

    Will companies have to tell me when I'm talking to an AI?

    Yes. The EU AI Act requires that users are clearly informed when they are interacting with an AI system rather than a human. This applies to chatbots, automated customer service, and AI-generated content that could deceive users.

    When did the EU AI Act come into full effect?

    The Act came into force in stages. The ban on unacceptable-risk AI systems applied from early 2025, while the most significant obligations for high-risk AI systems are being enforced from August 2026 onwards.

  • Open Source vs Proprietary AI Models: What the Difference Actually Means for You

    Open Source vs Proprietary AI Models: What the Difference Actually Means for You

    Most people using AI tools day-to-day have no idea whether the model running underneath is open source or proprietary. And honestly, that gap matters more than it might seem. The choice between open source vs proprietary AI shapes who can see your data, how reliable the tool is long-term, what it costs you, and who gets to decide when things change. These are not abstract technical questions. They have real consequences for individuals, small businesses, and organisations across the UK.

    Developer examining open source vs proprietary AI code on a workstation in a UK office
    Developer examining open source vs proprietary AI code on a workstation in a UK office

    What does open source actually mean in AI?

    In simple terms, an open source AI model is one where the underlying code, and often the model weights themselves, are made publicly available. Anyone can inspect it, download it, modify it, and in many cases run it locally on their own hardware. Meta’s Llama models are a well-known example. Mistral, a French AI company, has also released open models that developers across Europe have adopted widely.

    Proprietary AI, by contrast, is a closed system. The company that builds it controls everything: the training data, the model architecture, the safety guidelines, the pricing, and the infrastructure. OpenAI’s GPT-4o, Google’s Gemini, and Anthropic’s Claude are all proprietary. You access them through an API or a consumer-facing product, but you never see what’s underneath.

    Privacy: who can actually see what you type?

    This is where open source vs proprietary AI becomes very concrete. When you use a proprietary AI service, your prompts and outputs typically pass through that company’s servers. Depending on their terms of service, your inputs may be used to improve the model, stored for a defined period, or reviewed by human moderators for safety reasons. The UK Information Commissioner’s Office has made clear that UK GDPR applies to AI systems processing personal data, which means those companies are legally obliged to handle data appropriately. But whether they do, and how you’d verify it, is another matter.

    Open source models that you run locally hand control back to you entirely. Nothing leaves your machine. For a GP surgery, a solicitor’s firm, or any professional handling sensitive client information, that distinction is significant. The catch is that running a capable model locally requires real technical skill and reasonably powerful hardware.

    Close-up of hands at a laptop exploring open source vs proprietary AI privacy considerations
    Close-up of hands at a laptop exploring open source vs proprietary AI privacy considerations

    Reliability and the risk of dependency

    Proprietary AI products can change without warning. Pricing goes up. Features are removed. An API that your business has built a workflow around gets deprecated. This has already happened to developers who integrated early versions of commercial AI models, only to find behaviour shift noticeably between versions as companies quietly updated their systems.

    Open source models sidestep this risk in a meaningful way. Once you have a version of a model, it does not change unless you choose to update it. A small UK legal tech startup, for example, could pin their application to a specific version of an open model and maintain consistent output quality indefinitely. That kind of control is simply not available with closed systems.

    That said, reliability has another dimension: performance and safety guardrails. Proprietary models tend to have more extensive testing, red-teaming, and content filtering. Open source models vary considerably. Some community-maintained models have weak or nonexistent safety filters, which introduces different risks if you are deploying to end users.

    Cost: cheap until it isn’t

    Many proprietary AI tools start free or at low cost, then introduce tiered pricing as you scale. OpenAI’s API pricing, for instance, is charged per token and can accumulate quickly if you are processing high volumes of text. For a small business handling thousands of queries a month, that cost becomes non-trivial.

    Open source models can dramatically reduce those ongoing costs. If you have the infrastructure to run them, the marginal cost per query drops to almost nothing. The upfront investment in compute and engineering time is real, but many UK companies are finding the economics make sense at medium scale. Cloud providers including AWS and Google Cloud now offer hosted versions of some open models, which provides a middle ground between full self-hosting and pure proprietary dependency.

    The power balance between users and tech companies

    This is perhaps the least discussed but most important dimension of the open source vs proprietary AI debate. Proprietary AI centralises enormous power in a small number of companies, most of them based in the United States. They set the terms. They decide what the model will and will not do. They can implement restrictions overnight based on political pressure, regulatory concerns, or business strategy changes that have nothing to do with your needs.

    Open source distributes that power. It allows academic researchers, smaller companies, national governments, and individuals to build on AI capability without depending on a commercial relationship that can be withdrawn. The UK government’s AI Safety Institute has acknowledged the importance of understanding both open and closed frontier models precisely because the governance implications differ substantially.

    None of this means proprietary AI is bad and open source is good. The reality is more nuanced. A sole trader using a consumer AI tool for drafting emails does not need to run a local model. But a hospital trust, a financial services firm, or any organisation handling genuinely sensitive data should at least be asking these questions seriously.

    Which one should you actually use?

    For most individuals and small businesses, proprietary AI products remain the practical choice. They are easier to access, better documented, and more capable on complex tasks. If you are not handling sensitive data and you are comfortable with a company’s data terms, the tradeoffs are manageable.

    If you are in a regulated sector, building a product, or simply value knowing that your data stays where you put it, it is worth exploring what open source models can do. The capability gap between open and closed models has narrowed considerably in 2025 and 2026. Models like Meta’s Llama 3 and Mistral’s recent releases perform competitively on many everyday tasks.

    The honest answer is: understand what you are signing up for before you build your workflows around either. The open source vs proprietary AI question is not a one-time decision. It is something worth revisiting as your needs and the landscape both evolve.

    Frequently Asked Questions

    Is open source AI safe to use?

    It depends on the model and how it is deployed. Well-maintained open source models from reputable organisations can be safe, but they often have fewer built-in safety guardrails than large proprietary systems. Anyone deploying an open source model to end users should conduct their own safety and content moderation review.

    Can I run an open source AI model without a powerful computer?

    Smaller open source models can run on a modern laptop with a decent GPU, but larger, more capable models require significant compute power. Cloud platforms now offer hosted open source options, which removes the hardware requirement entirely while still reducing some of the proprietary dependency.

    Do proprietary AI companies store my data?

    Most proprietary AI providers do retain prompt data to some extent, though enterprise tiers often offer opt-outs. UK GDPR requires these companies to handle personal data lawfully, so you should check the provider’s data processing agreement and privacy policy before using any AI tool with sensitive information.

    What is the main advantage of proprietary AI over open source?

    Proprietary models from companies like OpenAI, Google, and Anthropic tend to be more capable on complex tasks, better tested for safety, and far easier to access without technical expertise. They are generally the right choice for individuals or teams who need reliable, high-quality output without managing infrastructure.

    Is open source AI cheaper than proprietary AI?

    It can be significantly cheaper at scale if you have the technical capability to self-host. For individual users, many proprietary tools have free tiers that make the cost comparison largely irrelevant. The real cost saving from open source comes when you are processing high volumes and would otherwise pay substantial API fees.

  • The Truth About AI Productivity Tools: Do They Help or Just Create More Noise?

    The Truth About AI Productivity Tools: Do They Help or Just Create More Noise?

    There is a version of the story where AI productivity tools transform your working day. Emails answered in seconds, meeting notes summarised before you have even left the room, first drafts produced whilst you get on with thinking. And to be fair, some of that is genuinely happening. But there is another version, one that cognitive scientists are increasingly interested in, where these tools simply add another layer of noise to already overwhelmed brains. The truth, as usual, sits somewhere between the two.

    Assessing AI productivity tools effectiveness honestly means looking at what the research actually says, not what the marketing decks promise. And it means taking seriously the cost that comes with every new tool you add to your workflow.

    Professional reviewing AI productivity tools effectiveness at a modern London office desk
    Professional reviewing AI productivity tools effectiveness at a modern London office desk

    What cognitive science tells us about tools and attention

    The human brain has a finite capacity for what researchers call “executive function”, the mental bandwidth that handles planning, decision-making, and sustained focus. Gloria Mark at the University of California has documented over many years that it takes an average of around 23 minutes to fully regain deep focus after an interruption. That finding holds up across replications, and it has uncomfortable implications for any tool that pings you, nudges you, or asks for a micro-decision.

    The issue with many AI productivity tools is not the AI itself. It is the interface. Notifications, suggested replies, inline prompts, smart compose suggestions, each one is a small cognitive interrupt. The brain registers it, evaluates it, and either acts on it or suppresses it. Both options cost something. Research published by the British Psychological Society has explored how multitasking and digital interruptions correlate with increased cortisol levels and reduced performance on complex tasks. You can read more about their research summaries on the BPS website.

    This does not mean AI tools are inherently bad for your brain. It means that how they are designed and how you use them matters enormously.

    Where AI tools genuinely do improve output

    Let’s be specific, because broad dismissals are just as unhelpful as breathless enthusiasm.

    Transcription and summarisation tools have a strong evidence base for reducing cognitive load. If you spend time in back-to-back meetings, a tool like Otter.ai or Microsoft Copilot’s meeting summary feature can free up the mental effort you would otherwise spend on note-taking. That is not a distraction. It is a genuine offload of routine processing, which leaves more capacity for higher-order thinking.

    Writing assistance tools show similar promise when used in a specific way: as a drafting aid after you have done your thinking, not as a shortcut that replaces it. Studies from the Oxford Internet Institute suggest that people who use AI to sharpen drafts they have already structured report feeling more confident in their final output, without the cognitive shortcut effect that can flatten original thinking.

    Task management and prioritisation tools are more mixed. Some people find that AI-assisted scheduling (tools that automatically block focus time or reorder tasks based on deadlines) reduces decision fatigue at the start of the working day. Others find the handover of control anxiety-inducing. Individual differences here are real and should not be papered over.

    Close-up of someone taking notes while testing AI productivity tools effectiveness
    Close-up of someone taking notes while testing AI productivity tools effectiveness

    When AI tools become the problem

    The pattern that emerges from user research is telling. People who adopt multiple AI productivity tools simultaneously, who are essentially trying to AI their way out of a structural overload problem, tend to report higher stress, not lower. A 2025 survey by Workfront (part of Adobe) found that UK knowledge workers using five or more digital tools simultaneously reported 34% higher feelings of overwhelm than those using fewer than three, even when overall task volume was similar.

    There is also a subtler issue worth naming: the cognitive tax of managing the tools themselves. Every AI assistant you add to your stack requires configuration, prompting, verification, and occasional correction. For complex or sensitive work, that verification step is not optional. Errors in AI-generated content have real consequences, and the mental effort of checking output can negate the time saved in generating it.

    AI productivity tools effectiveness is therefore not a fixed property. It is highly context-dependent. A freelance copywriter who uses one AI tool for research and one for first drafts may genuinely feel sharper and more productive. A project manager who has been given six AI tools by their employer, none of them integrated, is almost certainly experiencing the opposite.

    The wellbeing angle is not separate from the productivity one

    This matters for health reasons, not just output reasons. Chronic cognitive overload is associated with poorer sleep, higher rates of anxiety, and the kind of low-grade mental fatigue that builds up over weeks rather than days. The NHS’s own guidance on workplace stress notes that sustained pressure on attention and decision-making is among the most common drivers of burnout presentations in primary care.

    If your AI tools are adding to that load rather than reducing it, the productivity gains are illusory. You might move faster in the short term. You will pay for it in focus, mood, and recovery time.

    The honest advice, grounded in what research actually shows, is to treat AI productivity tools as you would any supplement or intervention: start with one, give it a genuine trial period, and measure the effect on your actual output and your subjective sense of control. If it helps, keep it. If it adds cognitive weight without clear return, cut it.

    A practical framework for choosing what stays

    Three questions worth asking before adopting any new AI tool:

    Does it remove a task I currently find draining, or does it add a new decision point? Draining task removal is valuable. New decision points are usually not.

    Can I use it in batch mode rather than real-time? Tools that work asynchronously, where you consult them rather than have them interrupt you, tend to fare better in attention research. Real-time suggestions and always-on assistants carry higher cognitive cost.

    Am I adopting this because it solves a real problem, or because it feels like progress? The novelty effect of new technology is well documented. It creates a short-term motivation spike that fades. Give any tool at least four weeks before deciding it has changed your working life.

    AI productivity tools effectiveness is real in the right contexts, with the right tools, used with deliberate restraint. The noise comes when we treat every new product as a solution to a problem we have not clearly defined. That is not a technology failure. It is a human one, and it is entirely fixable.

    Frequently Asked Questions

    Do AI productivity tools actually improve focus or make distraction worse?

    It depends heavily on how they are designed and how you use them. Tools used asynchronously, such as meeting summarisers or batch drafting assistants, tend to support focus. Real-time AI suggestions and notification-heavy interfaces often fragment attention, based on what cognitive research consistently shows.

    Which AI productivity tools are most effective for knowledge workers in the UK?

    Tools with clear, single-purpose functions tend to outperform all-in-one platforms. Meeting transcription tools like Otter.ai or Microsoft Copilot summaries, and focused writing assistants, receive the strongest user satisfaction scores in UK knowledge worker surveys. The key is avoiding tool sprawl.

    Can AI tools cause burnout or mental fatigue?

    Indirectly, yes. Research links high tool volume and constant digital interruption to elevated cortisol and reduced cognitive performance. If AI tools increase the number of micro-decisions you face rather than reducing them, they can contribute to the conditions associated with burnout over time.

    How long should I trial an AI productivity tool before deciding if it works?

    Most productivity and cognitive researchers suggest a minimum of four weeks to see past the novelty effect. Track something specific during that period, such as tasks completed, time spent on deep work, or your own sense of mental clarity at the end of each day.

    Are AI productivity tools worth paying for?

    Only if they solve a clearly identified problem. Many free tiers of tools like Notion AI, Copilot, or ChatGPT are sufficient for personal use. Paid tiers make more sense for high-volume, time-sensitive workflows where the time saving is measurable. Avoid paying for tools you have not trialled properly first.

  • Agentic AI Explained: What It Means When Your Software Starts Making Decisions for You

    Agentic AI Explained: What It Means When Your Software Starts Making Decisions for You

    Most people have now used a chatbot of some kind. You type something, it responds, and the exchange ends there. Agentic AI is a fundamentally different proposition. With agentic AI explained properly, the distinction becomes clear: these are software systems that don’t just respond to prompts but pursue goals, make decisions, and take sequences of actions across multiple tools and platforms, often without a human approving each step. That shift, from reactive assistant to autonomous actor, is one of the most consequential changes happening in technology right now.

    Understanding what these systems actually do, and where they fall short, matters whether you run a business, work in a regulated industry, or simply want to know what is being built into the software you already use every day.

    Person studying autonomous workflow systems on monitors, illustrating agentic AI explained in a modern tech workspace
    Person studying autonomous workflow systems on monitors, illustrating agentic AI explained in a modern tech workspace

    What Makes an AI System “Agentic”?

    A standard large language model responds to a single input and produces a single output. It has no memory between sessions, no ability to take action in the world, and no plan beyond answering the immediate question. Agentic systems are built differently. They combine a language model with persistent memory, tool access (web search, code execution, APIs, file systems), and a planning loop that allows them to break a goal into subtasks, attempt those subtasks, evaluate the results, and adjust their approach accordingly.

    The key word is autonomy. An agentic AI might be given a goal such as “research our three nearest competitors, summarise their pricing, and draft a report” and then complete that task end-to-end without further instruction. It decides which tools to use, in what order, and how to handle unexpected results along the way. This is categorically different from asking a chatbot to summarise a document you have already pasted in.

    Where Agentic AI Is Already Being Deployed in 2026

    Deployment is further along than most people realise. In software development, agentic systems now write, test, debug, and refactor code across entire projects, not just single functions. In customer operations, agents handle multi-step support queries by pulling account data, processing refunds, and updating records without routing the customer through a human at each stage. In legal and compliance work, agents review contracts, flag clauses against regulatory frameworks, and generate variance reports.

    Healthcare is one of the more significant frontiers. Agentic systems are being piloted to monitor patient data across multiple sources, identify early warning patterns, and generate clinical summaries for review by practitioners. The NHS and several private health networks in the UK have begun structured trials, with human oversight remaining mandatory at decision points. According to research published via the Lancet Digital Health, AI systems operating in a monitoring capacity can reduce the time clinicians spend on documentation by up to 40 percent, freeing capacity for direct patient care.

    Hands interacting with a decision-pathway interface, a detailed visual of agentic AI explained through connected automation nodes
    Hands interacting with a decision-pathway interface, a detailed visual of agentic AI explained through connected automation nodes

    The creative and manufacturing sectors are also seeing genuine adoption. Companies managing complex production workflows, including print and fulfilment businesses such as Print Shape, a UK-based online printing service, are using agentic tools to automate order routing, production scheduling, and quality checks across interconnected systems. The appeal is consistency and speed at scale, tasks that would require a team of coordinators handled by a system that runs continuously.

    The Genuine Risks You Should Understand

    The risks are not hypothetical. They are already being observed in early deployments and are worth taking seriously rather than dismissing as science fiction.

    The first is compounding errors. Because agentic systems act across multiple steps, a mistake made early in a task can propagate and amplify before any human sees the result. A standard chatbot error is contained; an agentic error can trigger a chain of consequential actions based on a flawed premise.

    The second is goal misalignment. When you specify a goal rather than a process, an agentic system optimises for the stated goal, sometimes in ways that satisfy the letter of the instruction while missing the intent entirely. This is not malice; it is the natural result of the system doing exactly what it was told to do, narrowly interpreted.

    The third is accountability. When an automated system makes a decision that causes harm, financial or otherwise, questions of liability become genuinely complex. UK regulators, including the Information Commissioner’s Office, have begun issuing guidance on how agentic AI activity intersects with GDPR obligations, particularly around automated decision-making that affects individuals.

    The fourth risk is over-reliance. Organisations adopting agentic tools without adequate human review processes risk degrading the internal expertise needed to catch errors when the system gets things wrong. This is a structural concern rather than a technical one.

    What the Benefits Actually Look Like in Practice

    When deployed in appropriate contexts with proper oversight, the productivity gains from agentic AI are real and measurable. McKinsey’s 2025 State of AI report found that organisations using agentic systems for knowledge work tasks reported median time savings of 25 to 35 percent on complex multi-step processes. The benefits are not evenly distributed, and they depend heavily on how well the system is scoped and supervised, but they are not illusory.

    For individuals, the most immediate benefit is cognitive offloading. Research tasks, administrative coordination, report drafting, and data collation can all be delegated in ways that free up time for judgement-intensive work. For businesses, the compounding effect of automating dozens of routine workflows can be transformative at the operational level.

    Businesses like Print Shape, operating in high-volume, process-driven environments, are among those positioned to extract genuine efficiency gains from agentic tooling, particularly where workflows are well-defined and measurable. That clarity of process, knowing exactly what success looks like, is also what makes agentic AI easier to supervise and correct in sectors like fulfilment and production.

    How to Think About Agentic AI as a Non-Technical Person

    The most useful mental model is this: treat an agentic AI system the way you would treat a capable but new member of staff. You would not give them unrestricted access to every system on day one. You would define the scope of their responsibilities clearly. You would check their work before it went out under your name. And you would expect to spend time teaching them what good looks like in your specific context.

    That framing, combining genuine capability with appropriate supervision, is where the most responsible and effective deployments of agentic AI currently sit. The organisations getting this right are not those handing over the most autonomy; they are those who have thought carefully about where human judgement remains non-negotiable and built their systems accordingly.

    Agentic AI is not a future concern. It is live in systems you likely already interact with, and understanding it clearly, its mechanics, its limits, and its risks, is now genuinely useful knowledge for anyone working in or around technology.

    Frequently Asked Questions

    What is the difference between agentic AI and a regular chatbot?

    A regular chatbot responds to a single prompt and produces a single output, with no ability to take independent action. Agentic AI pursues multi-step goals autonomously, using tools like web search, APIs, and file systems, and adjusts its approach based on results without needing human approval at each stage.

    Is agentic AI already being used in the UK?

    Yes. Agentic AI systems are actively deployed in UK sectors including software development, legal compliance, customer operations, and healthcare. The NHS and several private health networks have begun structured pilots, with mandatory human oversight at key decision points.

    What are the biggest risks of agentic AI systems?

    The main risks include compounding errors (early mistakes amplifying across multiple steps), goal misalignment (the system optimising for the literal instruction rather than the intent), accountability gaps when automated decisions cause harm, and organisational over-reliance that erodes internal expertise. UK regulators including the ICO have begun issuing guidance on these concerns.

    How is agentic AI regulated in the UK?

    There is no single dedicated agentic AI law in the UK as yet, but existing frameworks apply. The Information Commissioner’s Office has issued guidance on automated decision-making under UK GDPR, and the AI Safety Institute continues to publish risk assessments. Regulation is evolving rapidly as deployment scales.

    What kind of businesses benefit most from agentic AI?

    Businesses with high-volume, well-defined, repeatable workflows tend to see the clearest gains: fulfilment operations, legal document review, software development pipelines, and customer service at scale. The more precisely a success outcome can be defined and measured, the more effectively an agentic system can be scoped and supervised.

The Latest