Can AI chatbot conversations be used as evidence?
For years, investigators have treated email, text messages, browser history and messaging applications as potential sources of electronic evidence. A new category is now joining that list: AI chatbot conversations. A Washington Post investigation published on 27 August 2026 identified a growing number of civil and criminal cases in which conversations with AI systems became part of the evidential record. The significance goes beyond privilege. AI conversations are becoming a new class of digital evidence.
Yes, they can. In the United States, AI chat histories are already appearing in civil discovery, criminal investigations and privilege disputes, and in February 2026 a federal court held that a defendant’s documents memorialising conversations with a public AI platform were not protected by attorney-client privilege or the work-product doctrine in the circumstances before it. Privilege and discovery remain fact-specific and jurisdiction-specific, but users should not assume a chatbot conversation carries the protections of a conversation with a lawyer. The evidence may sit in browser artefacts, application data, screenshots, exports, account archives or provider systems. A prompt still needs context: who controlled the account, whether the conversation is complete, and what happened before and after. Hong Kong courts apply their own rules; take advice.
What courts are already seeing
The Washington Post identified a growing number of cases in which AI conversations became evidence, and court filings already show requests seeking documents created with public large-language models. In a US federal criminal case in the Southern District of New York, a judge considered whether documents memorialising a defendant’s conversations with Anthropic’s Claude were protected by attorney-client privilege or the work-product doctrine. The court held, in a written opinion dated 17 February 2026, that they were not protected in the circumstances before it. The defendant had not been tried when the ruling was reported; we do not name him.
Users may treat the conversation as private even when the law does not
A person may type something into a chatbot that they would never send to another person. They may ask whether particular conduct was lawful, how an investigation might find something, whether deleted information can be recovered, how to respond to an allegation, how to prepare for litigation, or what a message or contract means. The interface feels conversational, and that can create an assumption of privacy. From an evidential perspective, that assumption is dangerous.
In that case, the defendant had used the AI system while facing a federal investigation and later shared AI-generated documents with his lawyers. The court concluded that the communications were not privileged because the AI system was not counsel, and rejected the work-product claim on the facts before it. The important point is not that every AI interaction is automatically discoverable; privilege and discovery remain fact-specific and jurisdiction-specific. It is that users should not assume an AI conversation carries the legal protections that attach to communications with a lawyer.
The evidence may exist in more than one place
From a forensic perspective, an AI chat is not one artefact. Depending on the platform and how it was used, investigators may encounter:
- browser history and cached web content
- local application data
- screenshots
- exported conversations and downloaded account archives
- notification records and email notifications
- clipboard artefacts
- cloud account data
- linked documents or uploaded files
- provider-side records
Each source can have a different evidential value. A screenshot may show what appeared on a screen; it may not establish the complete conversation. A browser artefact may indicate access to the service; it may not show who typed a particular prompt. An account export may provide conversation content; it may need interpretation to establish timestamps, account identity and completeness. The same principles familiar from messaging and cloud evidence apply. The novelty is the content and context.
A prompt is not necessarily a confession
A user may ask a question that appears incriminating. That does not automatically establish conduct, intent or authorship. People test hypotheticals, research other people’s problems, write fiction, ask questions after an event, copy text from elsewhere and ask systems to argue both sides. A prompt therefore requires context. An examiner or lawyer should consider:
- who controlled the account
- which device generated the interaction
- whether multiple people had access
- whether the timestamp is reliable
- whether the conversation is complete
- whether prompts were edited or regenerated
- whether uploaded documents influenced the response
- whether the user was asking hypothetically
- what happened before and after the interaction
- whether other evidence corroborates the apparent meaning
The existence of a chat can be highly relevant. Its interpretation still requires care.
AI responses are also evidence of a different kind
The user’s words may be evidence of what the user typed. The model’s response is not evidence that the response was factually correct. Generative AI systems can produce inaccurate statements, misunderstand the user and give different answers to similar prompts, and their output can depend on the model version, configuration and conversation history. An investigator may need to distinguish between evidence of the user’s input, evidence of what the system displayed, evidence of what information was available to the user, evidence of whether the user relied on the answer, and the factual truth of the answer itself. Those are different propositions.
Discovery requests are already moving in this direction
In Aly v. Apple Inc., a discovery request sought documents created using public large-language models concerning the plaintiff’s complaint, allegations or the defendant’s defences; the plaintiff objected, including on privilege, work-product and proportionality grounds. Separately, litigation involving OpenAI has required courts to grapple with very large samples of ChatGPT conversation logs.
These cases involve different legal issues, but together they show that AI chat data is no longer outside the normal discovery conversation. Lawyers should expect opponents to ask whether relevant AI systems were used. Investigators should expect AI artefacts to appear on devices. Organisations should expect their employees’ AI usage to become relevant in disputes.
Corporate AI governance now has an evidential dimension
Most corporate AI policies focus on confidentiality, cybersecurity, intellectual property and accuracy. They should also consider evidence. If employees use AI systems for business decisions, internal investigations, HR matters, regulatory responses, transaction planning or disputes, those interactions may later become relevant. Which AI systems are approved? Are users authenticated? Are conversations retained, and can the organisation export them? What happens when an employee deletes a chat? Are uploaded documents retained separately? Can legal hold procedures preserve relevant AI interactions? Are enterprise and consumer accounts treated differently? These are potential discovery and investigation questions, not merely technology-governance ones.
Forensic acquisition needs to evolve with user behaviour
Digital forensics has always followed where users communicate and store information: first email, then mobile messaging, then cloud collaboration. AI conversation history is simply the next environment. The challenge is not to treat it as exotic. The fundamentals remain familiar: preserve the original source where possible, document acquisition, establish provenance, correlate independent artefacts, understand platform behaviour, separate content from interpretation, identify limitations, and avoid claiming more than the evidence supports.
What organisations should assume now
A sensible working assumption is straightforward: if an employee, litigant or subject of an investigation uses an AI chatbot to discuss a matter, that conversation may eventually become relevant evidence. That does not mean every chat should be retained forever. It does mean organisations should understand how AI usage fits into existing evidence-preservation and information-governance processes. Lawyers should also consider whether clients are using public AI tools to discuss sensitive matters, creating records that may not carry the protections the user expects. Privilege in Hong Kong is a matter of Hong Kong law; take advice.
The private feel of a chatbot interface should not be confused with legal privilege or evidential invisibility. The questions now are familiar ones: where is the evidence, how was it preserved, is it complete, who created it, and what does it actually support?
See computer forensics and eDiscovery.
Sources
- American Bar Association: Federal Court Rejects Privilege Assertion for AI-Generated Research
- Aly v. Apple Inc., No. 1:26-cv-03262-AS: discovery materials concerning documents created using public large-language models
- Authors Guild et al. v. OpenAI Inc. et al.: orders concerning discovery of ChatGPT conversation logs
- The Washington Post, 27 Aug 2026: They confided in ChatGPT. Their secrets ended up in court
General commentary on public reports, not legal advice and not a finding about any organisation or person named in those reports. Facts are as reported at the time of writing.
Frequently asked questions
Can AI chatbot conversations be used as evidence?
Yes. AI chat histories are already appearing in US civil discovery, criminal investigations and privilege disputes, and one federal court held a defendant's AI conversations were not privileged in the circumstances before it. The evidence can sit on devices, in account exports or with the provider, and needs context to interpret. Hong Kong courts apply their own rules; take advice.
Are my conversations with an AI chatbot privileged?
Do not assume so. A US federal court held that a defendant's conversations with a public AI platform were not protected by attorney-client privilege, because the AI system was not counsel. Privilege is fact-specific and jurisdiction-specific; take advice.
Where is AI chat evidence found?
In browser history and caches, local application data, screenshots, exported conversations and account archives, notification records, clipboard artefacts, cloud account data, uploaded files and the provider's own records.
Does an incriminating-looking prompt show what someone did?
Not on its own. People test hypotheticals, research other people's problems, write fiction and copy text. Investigators need to know who controlled the account, whether the conversation is complete, and what other evidence shows.
Should companies include AI tools in legal holds?
They should understand how AI use fits their preservation processes: which tools are approved, whether conversations are retained and exportable, what happens on deletion, and whether hold procedures can preserve relevant AI interactions.
Facing something like this?
The first discussion is confidential and free, and usually changes what you decide to do next. Call +852 5808 1071 or use the case review form.