Friday, 14 August 2026
Businesses that bought generative artificial intelligence systems on the strength of overstated claims about their capabilities could eventually seek compensation through fraud-based legal actions, according to an analysis by CMS lawyers Lee Gluyas and Jonathan Gardner.
The issue is emerging as organisations gain more experience with large language models and discover that systems capable of producing highly convincing material can still generate incorrect information, invented references and unsupported conclusions.
The legal profession has provided some of the most visible examples. Lawyers have faced criticism and, in some cases, judicial action after AI-assisted submissions included fictitious cases or authorities. Gluyas and Gardner point to nearly 1,900 reported instances globally involving hallucinated legal cases, arguing that similar problems are likely to become more visible in other professional sectors as AI adoption expands.
The potential commercial dispute, however, may increasingly concern not simply whether an AI system produced an incorrect answer, but what customers were told about the technology before buying it.
Generative AI systems are designed to generate responses based on patterns learned from large quantities of data. Their ability to produce authoritative-sounding language does not mean every statement has been independently checked. Citations can be incorrect, sources may not exist and apparently confident conclusions can require external verification.
For corporate buyers, that distinction could become significant where software was purchased after suppliers or advisers made ambitious claims about productivity, reliability or the range of work that could safely be automated.
The CMS analysis identifies several areas that could create legal exposure, including overstating a product’s capabilities, minimising known limitations, recommending applications beyond what the technology can reliably perform or failing to correct an important misunderstanding during the sales process.
Such questions could become particularly relevant where companies made substantial operational decisions based on anticipated AI productivity gains. Some businesses have already reconsidered workforce strategies after finding that automation did not deliver the expected results or still required significant human supervision.
However, poor AI performance by itself would not normally establish that a vendor is liable.
Software agreements commonly contain extensive limitations of liability, disclaimers and contractual provisions governing statements made before an agreement was signed. These protections can make conventional claims against technology suppliers difficult, particularly where sophisticated companies negotiated the contracts.
Fraud represents a potentially different route because English law does not allow a party to contract out of liability for its own fraud.
If a customer could demonstrate that it entered an agreement because of fraudulent representations concerning an AI product, the potential financial consequences could therefore extend beyond ordinary contractual remedies. Depending on the circumstances, damages could potentially include licence expenditure and additional losses caused by reliance on the misleading representation.
Proving such a case would nevertheless be difficult.
A claimant would generally need evidence showing that a representation was false, that the relevant person knew it was false or was reckless as to whether it was true, and that the customer relied upon it. Marketing language alone would therefore not automatically establish fraud.
One area likely to receive greater attention is implied representation — where the dispute concerns what a vendor’s conduct or statements effectively communicated rather than an explicitly false sentence in marketing material.
Gluyas and Gardner draw a comparison with litigation involving complex financial products, where claimants have sometimes relied on implied representations when contractual protections made other avenues for compensation difficult.
They highlight the English High Court’s 2023 decision in Loreley Financing (Jersey) No 30 Ltd v Credit Suisse Securities (Europe) Ltd and others. The proceedings included allegations concerning implied representations about honesty and the characteristics of loans underlying residential mortgage-backed securities. The case also illustrates that constructing such an argument does not guarantee that it will succeed.
For the AI industry, the broader issue could therefore become the difference between legitimate promotion of rapidly developing technology and representations that materially exceed what vendors knew their products could reliably deliver.
Warnings supplied with AI systems will also matter. Providers increasingly state that generated information can contain errors and should be independently checked. Where customers received clear warnings but nevertheless relied on AI output without verification, establishing responsibility on the supplier’s side could become considerably harder.
The question is particularly important for professional services, finance, property, healthcare and other industries where AI-generated information can influence decisions involving significant financial or legal consequences.
As enterprise adoption grows, due diligence on AI procurement may consequently need to move beyond functionality and price. Buyers may increasingly need records of demonstrations, sales presentations, capability statements, accuracy claims and discussions about intended applications alongside the final software contract.
The next phase of AI litigation may therefore focus less on whether large language models sometimes make mistakes, something now widely recognised, and more on whether customers were given an accurate picture of those limitations when they decided to buy and deploy the technology.
Source: CMS
