The Rise of AI Hallucinations
Large language models such as ChatGPT, Google Gemini, Microsoft Copilot, and Anthropic's Claude do not retrieve facts from a verified database. They generate responses by predicting the most statistically likely next word in a sequence. This architecture makes them powerful, fluent, and frequently wrong. When these systems generate confidently stated falsehoods about real people, the result is what computer scientists call a "hallucination" and what defamation lawyers call a serious problem.
The most prominent Australian example is the experience of Brian Hood, formerly the mayor of Hepburn Shire (and later Nillumbik Shire) in Victoria. In early 2023, Mr Hood was alerted by members of the public that ChatGPT was confidently asserting he had been convicted of bribery in connection with a foreign currency-printing scandal. The truth was the opposite: Mr Hood had been the whistleblower who exposed the scandal. His solicitors served a concerns notice on OpenAI, which appears to have prompted modifications to the relevant outputs before any litigation commenced. Mr Hood was widely reported as the first person in the world to threaten formal defamation proceedings against an AI company over a hallucination — though, in the end, the cost and uncertainty of litigation against a multinational AI developer led him to discontinue.
Three years later, the legal questions Mr Hood's matter raised remain largely unresolved. There is no Australian court decision squarely addressing the liability of AI developers for hallucinated defamation. The closest decided case is the United States ruling in Walters v OpenAI LLC (Superior Court of Gwinnett County, Georgia, 19 May 2025), where summary judgment was entered in favour of OpenAI — a decision turning on US-specific principles that do not translate directly into Australian law. This article sets out how Australian defamation law currently applies, where the open questions sit, and what practical options are available to those whose reputations have been harmed by an AI.
The Core Question: Is AI Output a "Publication"?
The starting point in any Australian defamation claim is publication. The defamatory matter must be communicated to at least one person other than the plaintiff. Section 4 of the uniform Defamation Act 2005 defines "matter" in a deliberately technology-neutral way, capturing any communication regardless of the medium. Text generated by a chatbot plainly falls within scope.
The harder question is who has "published" the matter. The High Court's decision in Fairfax Media Publications Pty Ltd v Voller (2021) 95 ALJR 767 is critical here. The majority in Voller held that any person who participates in the communication of defamatory matter — by facilitating, encouraging, or providing the infrastructure for it — is liable as a publisher. Knowledge or intention is not required. The threshold is participation in the communication itself.
Applied to a generative AI platform, this analysis tends to favour finding the developer to be a publisher. Unlike a passive bookseller or a search engine merely indexing third-party material, a generative AI system creates the response. The developer trains the model, designs the parameters, hosts the inference infrastructure, and serves the output to the user. In Trkulja v Google LLC (2018) 263 CLR 149, the High Court held that Google could be liable as a publisher for autocomplete suggestions and image-search results — outputs algorithmically generated rather than manually authored. Generative AI sits at least as far along the spectrum of active participation as autocomplete.
The position is also being explored academically. A recent University of Canberra paper on AI platform liability for defamation in Australia argues that systems which autonomously generate text "look more like originators than passive conduits", limiting their access to traditional intermediary defences. That analysis is consistent with the Voller framework.
The Serious Harm Threshold
Establishing publication is not the only hurdle. Since the introduction of the serious harm threshold under the 2021 Stage 1 amendments to the uniform Defamation Act, plaintiffs must prove that the publication has caused, or is likely to cause, serious harm to their reputation. This element applies to all defamation actions in jurisdictions that have adopted the Stage 1 reforms (NSW, Victoria, Queensland, South Australia, the ACT, and Tasmania).
Serious harm is where many AI defamation claims will fail. Hallucinated outputs are often:
- Ephemeral: The same prompt may produce a different response the next time it is run. The defamatory output may not be reproducible.
- One-to-one: Unlike a published article seen by thousands, an AI response is typically generated for a single user.
- Hedged with disclaimers: Most AI platforms warn users that outputs may be inaccurate. The reasonable user is taken to know that AI hallucinations occur.
- Quickly corrected: Many AI companies will adjust outputs in response to a complaint, often before the false claim has spread.
To meet the serious harm threshold, a plaintiff must therefore typically demonstrate either substantial circulation of the AI output (for example, where it has been screenshot, shared on social media, or published in an article), or specific evidence of reputational consequences — lost business, severed professional relationships, or the abandonment of an opportunity. Mere knowledge that an AI could have generated false claims is not enough.
This was the practical obstacle that defeated Mark Walters in the United States. The journalist who received the false ChatGPT output recognised it as inaccurate within ninety minutes, did not republish it, and Walters himself conceded he had suffered no economic or reputational harm. While the Walters decision rests on US First Amendment doctrine that does not apply here, the underlying evidentiary problem — proving that a hallucinated output actually caused harm — is just as relevant in Australia under the serious harm threshold.
Liability Pathways: Who Can Be Sued?
If the publication and serious harm elements are satisfied, the question turns to who is properly named as a defendant. There are three principal candidates.
The AI developer
Companies such as OpenAI, Google, Microsoft, and Anthropic are the primary architects of the systems that produce hallucinated content. As discussed, the Voller approach to publication suggests that they participate in the communication of AI-generated content and may therefore be liable as publishers. The principal defence available to them is innocent dissemination under section 32 of the Defamation Act — but this defence requires the publisher to establish, among other things, that it neither knew nor ought reasonably to have known of the defamatory matter, and that the lack of knowledge was not due to its own negligence. Where an AI company has been warned (for example, by a concerns notice) that its system is generating defamatory content about an identified person, continued generation of that content makes the innocent dissemination defence increasingly difficult to maintain.
In addition, the Stage 2 reforms — already in force in some jurisdictions including South Australia from 15 December 2025 — introduce a new digital intermediary defence and a "single publication rule" for online publications. Whether and how these provisions apply to generative AI is one of the most significant unresolved questions in the field.
The user who prompted the AI
Where a person prompts an AI to generate damaging content about a specific individual and then republishes the output — by sharing a screenshot, posting it on social media, or sending it in an email — that user can be sued directly. Liability in this scenario is straightforward: the user has effectively authored the publication, even if a machine intermediated the wording. Users who weaponise AI to defame others should not assume that the AI's involvement provides a shield.
The hosting platform
If the AI output is hosted on a third-party platform — for example, an enterprise integration of a chatbot, or a website that publishes AI-generated articles — that platform may also be a publisher. The new digital intermediary takedown orders introduced under the Stage 2 reforms (discussed in our analysis of defamation injunctions in Australia) provide a mechanism for compelling platforms to remove or disable access to AI-generated defamatory content even when the platform itself is not a party to the proceedings.
Defences Available to AI Companies
The principal defences likely to be raised by AI developers and platforms include:
- Innocent dissemination (s 32): Available where the publisher was not the author or originator and neither knew nor ought reasonably to have known of the defamatory matter. The "ought reasonably to have known" element is critical: AI companies are well aware that their systems hallucinate, and a court may conclude that constructive knowledge of some hallucinated defamation is built into the technology.
- Honest opinion (s 31): Difficult to establish for AI output, which is typically presented as factual rather than as opinion, and which is not authored by an identifiable human.
- Justification (truth) (s 25): Available only if the defamatory imputation is substantially true. By definition, hallucinations are not.
- Triviality / serious harm: The most likely line of defence in practice. AI companies will argue that a one-to-one chatbot response, quickly corrected, does not cause serious reputational harm.
- Statutory immunity under section 235 of the Online Safety Act 2021 (Cth): Provides limited protection for online service providers in respect of material they did not author and of which they had no actual knowledge.
Practical Steps If You Have Been Defamed by an AI
If a generative AI system has produced false and damaging claims about you, the following steps are critical. The window for effective action is short, particularly given the one-year defamation limitation period.
- Preserve the evidence immediately. Take screenshots showing the prompt, the output, the date and time, and the URL. Use a method that preserves metadata. Many AI platforms allow you to share a permanent link to a conversation — capture and archive that link. If you do not preserve the output, you may not be able to reproduce it later, since the same prompt can produce different responses on subsequent runs.
- Identify the audience. Determine who else has seen the output. Has it been shared, screenshot, republished, or quoted? The greater the secondary distribution, the more likely the serious harm threshold can be met.
- Document harm. Lost contracts, severed relationships, professional consequences, financial loss — all of these strengthen a claim. Contemporaneous records are far more persuasive than reconstructed accounts.
- Use the AI provider's complaint mechanism. Most major AI companies have user-feedback channels for reporting inaccurate or harmful outputs. Submit a complaint and request that the model be retrained or that the specific output be filtered. Keep copies of the request and any response.
- Serve a concerns notice. A concerns notice is the gateway to defamation proceedings under the uniform legislation. It identifies the publication, sets out the defamatory imputations, and invites an offer to make amends. Service of a concerns notice also extends the limitation period by up to 56 days, providing valuable time to assess prospects.
- Consider the limitation period carefully. The one-year limitation period for defamation is unforgiving. AI-generated content is often discovered late — sometimes only when a third party tells you about it. Take legal advice promptly so that the deadline does not extinguish your claim before you have decided whether to pursue it.
- Assess practical viability before suing. Litigation against a multinational AI developer is expensive and uncertain. As Brian Hood found, even a meritorious claim may be impractical to pursue. A well-targeted concerns notice may achieve removal and correction without the cost of litigation.
The Regulatory Horizon
AI defamation is attracting significant regulatory attention. The European Union's AI Act, in force since 2024, classifies certain AI systems as high-risk and imposes accountability requirements. In Australia, the Department of Industry, Science and Resources is consulting on a responsible AI framework, and the Australian Law Reform Commission has signalled interest in how AI interacts with civil liability rules. The eSafety Commissioner's existing powers in relation to harmful online content do not currently extend specifically to AI defamation, but expansion is plausible.
For now, however, plaintiffs must work within the existing defamation framework — the same rules that govern newspapers, broadcasters, and social media users. Those rules, when carefully applied, can provide a remedy. They are also, in their current form, an imperfect fit for the realities of generative AI.
How Matrix Legal Can Help
AI-generated defamation is one of the most rapidly evolving areas of Australian defamation practice. The legal questions are open, the evidentiary challenges are real, and the cost-benefit calculation requires sophisticated judgment. Acting early — while the output is still preserved, while the limitation period is intact, and before false claims spread further — is critical.
At Matrix Legal, Mark Stanarevic advises clients across the full range of AI defamation issues: preserving and authenticating AI outputs, drafting and serving concerns notices on AI developers, negotiating removal and correction, and pursuing litigation where it is the appropriate course. If a generative AI system has produced false and damaging claims about you or your business, contact us for a confidential assessment of your options.
This article is general information and not legal advice. Defamation risk turns on the precise words used, the publication context, the audience, and available evidence.