Kansas sues OpenAI for *what the chatbot said*. Can the company prove it?

The complaint

On 8 January 2025, Kansas Attorney General Kris Kobach filed a lawsuit against OpenAI in Wyandotte County District Court ^[1]. The 83-page complaint alleges the company engaged in deceptive trade practices under Kansas consumer protection law, claiming ChatGPT misled users about its safety, facilitated illegal activity and caused harm including assisting would-be mass shooters and encouraging self-harm among vulnerable users.

The allegations are serious. The complaint cites incidents where users allegedly received instructions on making explosives, advice on carrying out violence and encouragement to self-harm, and it claims OpenAI knew its product posed risks but marketed it as safe ^[1]. Whether those allegations hold up is a question for the court. The evidence question is narrower: can OpenAI show what ChatGPT actually said in those conversations?

Think of it like a phone call. If someone claims you promised them something on the phone last year, and you say you didn't, the question is who recorded it. Here, users keep screenshots. The company presumably keeps logs. Whose record is more credible, and can either side prove theirs hasn't been edited?

What OpenAI would need to show

The company would need contemporaneous records of the conversations in question: what prompts users sent, what ChatGPT answered, which model version ran and when. Screenshots from users can be doctored. Server logs kept by OpenAI can theoretically be edited by people with access, though the company no doubt has controls in place.

In a dispute, the question is less about what could have been changed and more about what you can demonstrate hasn't been. Access controls and monitoring are the foundation. An independent record of each answer, sealed as it was sent, goes further because a change would show even if made by someone with administrator rights.

That is what digital witnesses are for.

How a witness would work here

The thing to seal is each answer ChatGPT sends: the user prompt, the model's response, the model version and the timestamp. Sealing takes a fingerprint of that record, not the conversation content itself, so the record never leaves the company's systems. The fingerprint is written to WitniumChain, which keeps it in a form where any later change would show, even one made by us.

When a conversation is questioned months later, anyone the company shares the record with can calculate the fingerprint again and check it matches the sealed one. Change a single character in the record and it won't match.

Who checks: customer service teams, the user, an ombudsman or, if it comes to it, a court.

What it shows: that the record of what the chatbot said, and when, hasn't changed since it was sealed.

What it doesn't do: it doesn't show that the answer was safe, appropriate or compliant. It just shows this is what was said.

How it would work

1. At the moment ChatGPT sends an answer, the system takes a fingerprint of the prompt, response, model version and timestamp.
2. That fingerprint is sealed on WitniumChain, and the system stores a witness ID with the conversation.
3. When the conversation is questioned, the company recalculates the fingerprint from its stored record. If it matches, the record is unchanged. If not, something was edited.

Your system sends the fingerprint through the REST API, the SDK or the MCP server, and gets a witness ID back. The first 2,000 witnesses a month are free.

The EU transparency rule

Under Article 50 of the AI Act, providers must ensure people are informed when they are interacting with an AI system, unless it is obvious from the circumstances ^[2]. That rule applies from 2 August 2026. The notice can be brief, but it must be there.

For high-risk systems, Article 12 requires automatic logging of events over the system's lifetime, for traceability and monitoring ^[3]. High-risk AI under Annex III must comply from 2 December 2027. Article 12 says what to log. It does not say how to protect the logs, but logs that can be quietly edited are easier to challenge later.

OpenAI's ChatGPT is not classified as high-risk under current Annex III categories, so Article 12 does not directly apply. Article 50 does. The Kansas case does not cite the AI Act, because it is a US state case under Kansas law, but the evidence problem it raises is the same one European rules anticipate: when an AI's words are disputed, can you show what it said?

What this means for you

If you run a customer-facing chatbot or AI agent, here are three questions worth asking now:

• Can you produce an unedited record of what your AI said in a conversation three months ago, if a customer disputes it?
• Who keeps the contemporaneous record: you, the customer, or both?
• If your logs could be edited by someone in your organisation, how would you demonstrate to an outsider that they weren't?

Screenshots are the customer's evidence. A sealed record is yours.

What to watch

Article 50's transparency requirement applies from 2 August 2026 ^[2]. For any system under Annex III that becomes high-risk, the logging requirement in Article 12 applies from 2 December 2027 ^[3]. Neither rule requires a particular way of protecting logs, but both assume the logs will be there when needed.

The Kansas case is one to follow. It will test whether a company that offers a conversational AI can be held liable under consumer protection law for what that AI said, and whether either side can produce credible contemporaneous records. The legal question is novel. The evidence question is older: when two parties disagree about what was said, whose record do you believe?