ChatGPT Hallucinations as Defamation: Who Is Liable When AI Lies About You?

Large language models like ChatGPT, Google Gemini, and Meta's Llama are generating false statements about real people at an unprecedented scale. These "hallucinations" — confident-sounding but entirely fabricated outputs — can include false criminal records, invented misconduct, and fictional professional failures. When these outputs are published or relied upon, they can constitute defamation under UK law.

What Are AI Hallucinations?

AI hallucinations occur when a language model generates text that is factually incorrect but presented with apparent confidence. Unlike human lies, hallucinations are not intentional — they result from statistical pattern matching rather than factual understanding. Common defamatory hallucinations include:

• Falsely stating that a person has been convicted of a crime
• Inventing sexual harassment allegations against named individuals
• Fabricating professional misconduct or regulatory sanctions
• Attributing statements to people who never made them
• Confusing individuals with others who share similar names

Who Is Liable for AI-Generated Defamation?

The question of liability for AI hallucinations is developing rapidly. Potential defendants include:

• The AI developer — companies like OpenAI, Google, and Meta that create the models. Under section 1 of the Defamation Act 2013, the developer may be a "publisher" if the hallucination is generated and displayed to users
• The platform operator — websites and apps that deploy AI chatbots may bear responsibility as secondary publishers
• The user who republishes — anyone who copies AI-generated defamatory content and shares it (e.g., in a blog post, email, or social media) becomes a primary publisher
• The employer — organisations that deploy AI tools for customer-facing communications may be vicariously liable for defamatory outputs

The Section 5 Defence for Operators

Website operators that host AI-generated content may seek to rely on the section 5 defence for website operators. However, this defence requires the operator to demonstrate that it did not post the statement — which is problematic when the AI is the operator's own tool generating content in real-time.

Proving Publication and Serious Harm

AI hallucinations present unique challenges for the standard defamation elements:

• Publication — each time a user receives a hallucinated defamatory response, this constitutes a separate publication. If millions of users can generate the same false output, the scale of publication may be enormous
• Identification — the hallucination must identify the claimant. Full names, job titles, or contextual details that make the person identifiable are sufficient
• Serious harm — harm may be inferred from the scale of potential publication and the gravity of the false allegation (e.g., fabricated criminal convictions)

Practical Remedies

1. Notice and takedown — notify the AI developer that their model generates defamatory content about you. Document the hallucination with screenshots and exact prompts
2. Pre-action protocol — send a formal letter before action to the developer requiring model correction
3. Injunctive relief — seek a court order requiring the developer to filter or correct the defamatory output
4. Damages claim — pursue compensation for reputational harm, particularly where the hallucination has been widely reproduced
5. Right to erasure — under UK GDPR, request deletion of inaccurate personal data from the model's training data