The UK’s newest product safety statute, the Product Regulation and Metrology Act 2025, defines a product as “a tangible item that results from a method of production”. In our view, that phrase matters more to the call from MPs and peers for product-style checks on chatbots than anything said so far. On our reading of the text, a chatbot built into a toy sits on one side of it and a chatbot reached through a browser sits on the other.
Strategic Insight: “Treat chatbots as products” bundles three questions: must they be tested before release, who can stop their supply, and who pays when they cause harm? UK law answers each through a different statute, and none of the reforms reviewed here would by itself require a standalone chatbot to pass safety testing before release.
What are MPs and campaigners actually asking for?
Two requests landed in the same week, and they are easy to run together. The Observer reports that MPs are examining whether companies should have to test chatbots for risks and comply with safety standards ahead of release. The report names no committee, bill or timetable. Separately, a letter to the Business Secretary, Jonathan Reynolds, asked ministers to “prohibit the supply of toys with built-in generative AI chatbots to children under 16”, Sky News reports. We covered that letter separately.
The first request reaches chatbots generally and, on our reading, would need new law. The second concerns physical goods, and its signatories say existing law already allows it.
The analogy covers two very different regimes
The comparisons in the Observer’s report point in different directions. David Davis reached for medicines, and Beeban Kidron for the air fryer, saying there are “more rules about building and distributing an air fryer” than a chatbot.
Those regimes work differently. Under the Human Medicines Regulations 2012, “A person may not sell or supply, or offer to sell or supply, an unauthorised medicinal product.” Permission comes first. Mains electrical goods fall under the Electrical Equipment (Safety) Regulations 2016, whose assessment procedure is “internal production control”: the manufacturer declares conformity “on the manufacturer’s sole responsibility” and keeps its documentation available to enforcing authorities for 10 years.
So one model is a licence, and the other is a self-declaration backed by enforcement. The Observer’s headline puts both under the word “tested”. The toy letter asks retailers for something else again, “independent testing”, which sits closer to the medicines end. In our view, “like an air fryer” and “like a medicine” are different policies, and the reports we read do not choose between them.
Reality Check: An air-fryer-style regime would not mean that a regulator checks a chatbot before launch. It would mean the provider assesses the system against set requirements, declares that it conforms, and answers to enforcers afterwards.
Three routes, three different answers
The Observer calls the current position a “patchwork” of data protection and online safety law. For the product question, we focus on three statutes, with medicines as the comparison point.
| Regime | Its test | Position of a standalone chatbot | When it applies |
|---|---|---|---|
| Product Regulation and Metrology Act 2025 | ”a tangible item”; software counts as an “intangible” component | Outside the definition, on our reading | Marketing or use, through regulations |
| Consumer Protection Act 1987, Part 1 | ”any goods or electricity” | No direct authority, the Law Commission says; reform proposed | After harm, as compensation |
| Online Safety Act 2023, new s.216A | ”AI service”: an internet service able to generate AI content | Can be brought in by regulations | Illegal content and priority offences |
| Human Medicines Regulations 2012 | ”medicinal product” | Not applicable; shown for comparison | Before sale |
The online route is limited by content, not by type of harm
The Crime and Policing Act 2026 inserted a power for ministers to amend the Online Safety Act to tackle “illegal AI-generated content” and the use of AI services for priority offences. It came into force at Royal Assent. The government’s own explanation in March was that generative AI services which only draw on their underlying model and have no content-sharing function “remain unregulated under the OSA”.
The Observer reports that ministers have promised to make AI firms assess the risk of generating illegal content, but that this would not necessarily reach emotional dependency or delusions being reinforced. The reason is worth stating precisely. The Online Safety Act already defines harm as “physical or psychological harm”, so psychological injury is not foreign to it. The limit is that the new power is tied to illegal content and offences.
Strategic Reality: The toy letter’s concern is “psychological, developmental or safeguarding harm”. In principle, a product regime would widen the trigger from what a chatbot says that breaks the law to whether the system is safe. That is the substantive gain, and also the hard part to specify.
Why the toy falls on the other side of the line
A toy is a tangible item. The 2025 Act lets regulations set requirements on a product’s “components (whether tangible or intangible)”, and section 2 says intangible components include software. On that wording, a chatbot inside a toy can be regulated through the toy. That is a power rather than a rule in force: discussing marketplace rules, the Commission says “no relevant regulations have been passed” under the Act as of its publication (para 3.82). Without the toy, the same software has no tangible item to attach to.
The current General Product Safety Regulations 2005 are less clear-cut. Their definition of a product has no tangibility test, and they do not mention software. We have found nothing showing whether they reach a standalone chatbot. The government’s March reform proposals are made under the 2025 Act, so that open question sits in the regime being overhauled.
The older powers point the same way. The Consumer Protection Act 1987 lets ministers make regulations so that goods that “would be unsafe in the hands of persons of a particular description” are not made available to those persons, wording that appears to fit an age-based restriction. Its definition of personal injury includes “any other impairment of a person’s physical or mental condition”. So when Baroness Finlay says the government “already has the power”, the text gives her argument something to stand on.
There are limits. A prohibition notice under section 13 cannot be given on any safety aspect the General Product Safety Regulations 2005 already cover. The 2025 Act also provides for Part 2 of the 1987 Act to be omitted, although that repeal is not yet in force. The report we read does not say which power the letter relies on. The government’s reported reply says it is “urgently investigating” through its call for evidence, without addressing whether the power exists.
Critical Context: The toy safety call for evidence closed on 6 October with no outcome shown on its page. It asks about toys “which use AI”, but its document never uses the words “chatbot” or “psychological”. The letter complains of a category of toy that “no UK regulator tests for psychological, developmental or safeguarding harm before it reaches children”.
What the Law Commission’s proposal would and would not do
The Law Commission’s consultation is the only one of these routes that would treat AI systems and software as standalone products, “whether supplied on or embedded in a device, downloaded from the internet or accessed remotely”. Its 258-page paper has chatbots in view: it argues the mode of delivery should not determine classification, using an AI chatbot reached by app, browser or vehicle as its example (para 2.23).
The paper also takes on the services argument directly. It says AI systems are “generally more appropriately characterised as products than as services”, because users “are entitled to expect them to operate to a reliable and consistent standard” for their purpose (para 2.42). It says the services label the Online Safety Act gives social media platforms “should therefore not dictate their classification” for liability, since that Act regulates systemic risk whilst the 1987 Act compensates people harmed (para 2.106). In our view, the same reasoning reaches chatbots. Free tiers could be caught too: the Commission provisionally proposes treating a product as supplied, when a business offers it or it is offered for profit, even if the user pays nothing (para 2.177). The paper also cites a Florida ruling, Garcia v Character Technologies, that a chatbot app could be a product where claims concern its features rather than its ideas or expression, and quotes the court’s remark that “courts [are] split” on virtual platforms (footnote 132).
That is compensation law. The review concerns Part 1 of the 1987 Act, which the Commission says exists to ease compensation claims by those hurt by defective goods, and none of its key proposals is a duty to test before release. A producer’s exposure to claims could still shape what it chooses to test. The paper draws the same line, describing the 2025 Act as having a “preventive function” and the 1987 Act as providing redress (para 3.83). Nor would passing a test end the matter: where a product has cleared a formal risk-benefit assessment, the Commission says a court “would not be bound to reach the same conclusion” (para 4.53). The Commission can only recommend reform for England and Wales, leaving wider application to the UK Government and devolved administrations (para 1.69). Responses close on 14 January 2027, final recommendations are expected in late 2027, and Lewis Silkin notes the government then decides whether to take them forward.
Hidden Cost: If the Commission’s proposal became law and the 2025 Act stayed as written, a standalone chatbot could be a product when someone sues for harm, yet not a product when ministers set safety requirements. The Commission says compliance with product safety rules “may provide compelling evidence” on whether a product was safe enough, though that does not automatically defeat a claim (para 4.6). A chatbot provider might have no such product rules to point to, beyond whatever online safety duties applied.
Who is affected, and what is still unknown
| Group | What changes for them | What is still unknown |
|---|---|---|
| Chatbot providers | Possible Online Safety Act duties by regulation; possible product liability if the proposals are enacted | Whether any testing duty before release follows |
| Toy makers, importers and marketplaces | Pressure to delist, a proposed commodity code, and proposed marketplace duties and liability | The government’s answer to the letter and to its call for evidence |
| Firms building on others’ models | Lewis Silkin says “producer” could extend to whoever developed software or AI systems, with clearer rules on substantial modification after supply | How liability divides along the supply chain, which Freshfields reported the Commission calling “presently unclear” in earlier work |
| People harmed and their families | A clearer route to psychiatric injury claims, including for people with a close tie to someone harmed, without having to be present at an incident | Whether the proposals are enacted, and in what form |
These documents do not settle who would own the regime either. Under the new power, ministers could give Ofcom functions over AI services. The joint statement on personal AI agents from the CMA, FCA, ICO and Ofcom says product safety sits outside their remits, and the toy call for evidence comes from the Office for Product Safety and Standards. The letter looks to Trading Standards, HMRC and Border Force. A product regime for chatbots would have to decide who runs it, and we have seen no proposal that does.
The open challenges
Testing something that changes after release
A declaration of conformity describes a product model. Lewis Silkin notes that a producer could make a product defective by issuing a faulty software upgrade or failing to provide a necessary update, so liability thinking is already adjusting. The Observer reports Sam Altman’s argument that releasing AI to the public and fixing faults as they appear is how systems get safer, which Davis likens to trying out drugs on the population at random. On that argument, a chatbot keeps being changed after release, so a testing regime would need to decide what triggers a fresh assessment. The disagreement is over when the testing happens, and a one-off check before launch would not fully settle it.
No named standard to test against
The toy regime leans on standards. Under the Toys (Safety) Regulations 2011, a toy conforming with designated standards is presumed to meet the essential safety requirements, to the extent those standards cover them. Neither the Observer’s report nor the toy letter as reported names a standard for how a chatbot should respond to a distressed user or a child. Until one is named, “meet safety standards” describes an aim rather than a test a provider could pass.
Whether a chatbot’s answer is a product feature or speech
This is where the air fryer comparison stops helping. The Commission says psychiatric harm claims over “outputs generated by an AI system or chatbot” may bring in the free-expression right in Article 10 of the human rights convention (para 5.41). Its answer is that meeting lawful but upsetting material should not by itself make a product defective, and that what counts is whether harm is attributable to design or functioning rather than to the meaning of the material (para 5.47). Relevant factors could include targeting known vulnerabilities and an absence of effective safeguards (para 5.48). It does not treat Article 10 as a bar, saying the right “does not make such claims impossible” (para 5.42) and regarding liability “confined to serious, foreseeable and properly attributable psychiatric injury” as compatible with it (para 5.49). Under its provisional proposal, purely psychiatric claims would need a recognised or recognisable illness within the “natural range of reactions of a person of ordinary fortitude”, with the vulnerabilities of groups such as children taken into account when a product is aimed at them (paras 5.32 to 5.37). Chatbot output raises a further boundary: the paper’s view for now is that information with no further function should not generally count as a product (para 2.138), whilst accepting that software and AI can “blur the distinction between functionality and information” (para 2.52). A testing regime aimed at lawful but harmful responses would have to draw a similar line in advance, and we have seen no proposal for how.
Whether unforeseeable behaviour counts as a defence
The Commission proposes to keep the “development risks” defence, which protects a producer from liability where the state of knowledge meant a defect could not be found. But where a producer keeps the ability to update software and fails to, it expects the defence to be “rarely, if ever, available” (para 6.60), because that producer can improve safety “in almost real time”. A chatbot its provider can keep updating looks to us like that case where the defect comes from a failure to update, although the defence would survive if the defect was undiscoverable at that point (para 6.63). It is a provisional proposal, so how it would apply to behaviour nobody predicted remains open.
Diverging from the EU
The government’s own document says the EU AI Act prohibits certain AI systems, giving the example of “an AI-enabled toy that encourages children to engage in risky behaviours”. Freshfields noted that the revised EU Product Liability Directive includes software as a standalone product. The new EU toy regulation will also apply in Northern Ireland through the Windsor Framework. In our view, UK choices will be read against those rules, whichever way they go.
What to watch
Five checkable events will show which route is gaining ground.
- The reply to the toy letter. If ministers restrict sales under existing powers, the product route has real force for AI in physical goods. If they point to the call for evidence, it remains a future reform.
- The toy and product safety responses. Osborne Clarke reports that the March proposals would add cybersecurity, AI and machine learning as considerations in deciding whether a product is safe. Whether the eventual response mentions psychological harm would indicate if the toy letter’s complaint has landed.
- The first regulations under section 216A. A title search of legislation.gov.uk on 5 November found none. Regulations confined to illegal content would confirm that ministers see the online route as sufficient for now. Any move towards testing or design duties would suggest the two routes are converging, within the limits of that power’s purposes.
- 9 December 2026 in Northern Ireland. The Commission says the new EU directive on product liability must be implemented there by then, so far as it falls within the Windsor Framework (para 1.71). If that brings standalone software within product liability in Northern Ireland, part of the UK would get there before any reform in England and Wales.
- The Law Commission’s 14 January 2027 deadline and late 2027 report. Watch whether the final recommendations keep AI systems accessed remotely inside the product definition, and whether government accepts them.
The word “product” is doing three jobs in this debate, and compensation is the only one where a live proposal would treat standalone chatbots as products. MPs asking for safety testing before release are asking for something none of those routes yet provides. The useful question for ministers is whether “tangible item” should remain the boundary of product safety law now that some of the products that worry Parliament have no physical form.
Source: AI chatbots should be tested like medicine and air fryers, say MPs (The Observer, 2026). Legal readings are our own from the published text of the legislation, and are not legal advice.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching how ministers answer the toy letter, and whether the word “tangible” survives the next round of product safety reform. Read more analysis at Insights, or get in touch.