The Liability Gap: When Public AI Systems Fabricate Facts About Your Organisation

2026-03-30

In early March, a $10 million lawsuit landed against OpenAI. The allegation: ChatGPT had generated legal documents so convincing that a claimant used them to reopen a dispute that had been settled years earlier. The filings cited arguments and procedural motions that had no legitimate basis.

It is one of the first cases to directly challenge an AI provider for the real-world consequences of fabricated outputs. But buried in the legal filings is a detail that should concern every organisation, not just the ones in court: OpenAI's terms of service explicitly disclaim responsibility for the accuracy of what its systems produce.

This is not an oversight. It is the industry standard.

The Disclaimer Economy

Every major public AI system — ChatGPT, Perplexity, Google's Gemini — includes some version of the same caveat: outputs may contain errors, and the provider is not responsible for how they are used.

From a legal standpoint, this makes sense for the AI providers. From a reputational standpoint, it creates a vacuum. When a public AI system tells a potential investor that your organisation is under regulatory investigation — and it is not — there is no correction mechanism. No editorial standards board. No right of reply.

The New York Times has already argued in court that Perplexity generates fabricated statements and falsely attributes them to the newspaper using its registered trademarks. If a major media organisation struggles to get corrections from these systems, consider what it looks like for a mid-market company with no legal team dedicated to AI oversight.

The Numbers Are Not Reassuring

The best public AI systems today hallucinate between 3% and 18% of the time, depending on the complexity of the query. On straightforward summarisation tasks, error rates hover around 0.7%. But on nuanced questions — the kind stakeholders ask about organisations, such as regulatory history, competitive positioning, or leadership track records — rates climb significantly.

A 2024 Deloitte survey found that 38% of business executives had made incorrect decisions based on hallucinated AI outputs. The global financial cost of AI-generated misinformation reached $67.4 billion that same year.

These are not theoretical risks. They are the operating conditions of the information layer that is rapidly replacing traditional search.

No Correction, No Recourse

Here is where the gap becomes dangerous. When a newspaper publishes an error about your organisation, you can request a correction. When a regulator cites inaccurate data, there are formal dispute mechanisms. When a competitor makes a false claim, there are legal remedies.

When a public AI system fabricates something about your organisation, none of these mechanisms exist.

There is no inbox to email. No editorial policy to invoke. No ombudsman. The AI provider's terms already disclaim liability. The person who received the false information may never visit your website to encounter the truth. And the fabricated claim may persist across millions of future queries, reinforced every time someone asks a similar question.

Researchers have demonstrated that public AI systems can be influenced by false content published online, causing them to repeat and amplify misinformation on critical topics. The information does not need to be true to spread. It just needs to exist in the training data — or in the sources the system retrieves.

The Regulatory Lag

Governments are beginning to pay attention, but the frameworks are not designed for this problem. The EU AI Act, with its transparency obligations taking effect in August 2026, focuses primarily on ensuring users know when they are interacting with an AI system and on regulating high-risk applications. It does not address the accuracy of what general-purpose AI systems say about specific organisations.

The White House AI framework, released in March, is similarly focused on developer liability, child safety, and data infrastructure. Neither framework contemplates a world in which public AI systems function as de facto information authorities about real organisations — with no accuracy obligations.

What This Means for Organisations

The liability gap is not going to close quickly. Regulation moves slowly. AI providers have no commercial incentive to guarantee accuracy about individual organisations. And the volume of AI-generated responses is scaling far faster than any correction mechanism could.

This leaves organisations in a position that would have been unthinkable five years ago: the most influential information channel for stakeholders is one where you have no visibility into what is being said, no mechanism to correct errors, and no legal recourse when those errors cause harm.

The first step is not to wait for regulation to catch up. It is to know what is being said — systematically, across every major public AI system, on an ongoing basis. Because the cost of a fabricated claim is not measured in a single query. It is measured in the decisions that follow.