How US courts handle AI evidence: the procedural mess nobody talks about

The problem with proving what happened

When AI goes wrong, the legal fight often turns on a surprisingly basic question: what did the system actually do?

Reed Smith's guide to AI disputes, published in late 2024, walks through the procedural tangles US courts are facing ^[1]. It covers copyright battles, deepfakes and chatbot disputes, but the thread running through all of it is evidence. Not whether the AI was biased or faulty, but whether anyone can prove what it said, what data it saw or which version was running.

Think of it like a car crash where neither side has dashcam footage. You can argue all day about who had the right of way, but without the recording the court has to rely on witness memory, skid marks and whatever else is left. AI disputes are often worse, because the system's workings are opaque and the logs, if they exist, are kept by the party being sued.

The guide is aimed at US litigators, but the questions it raises matter in the EU as well, where high-risk AI systems will soon have to log events under Article 12 of the AI Act ^[2]. Logging is one thing. Proving later that the log is accurate is another.

Discovery: who has to hand over what

In US litigation, discovery is the stage where each side can demand documents and data from the other. In AI cases, that can mean asking for training data sets, model weights, prompts, fine-tuning logs and internal communications about how the system was built ^[1].

The guide notes that there's no settled law yet on what must be disclosed. One judge might order a defendant to hand over the full training data; another might say it's protected as a trade secret. The outcome often depends on how the requesting party frames the demand and whether they can show the data is relevant and not available any other way ^[1].

For the side operating the AI, that uncertainty is a risk. If you can't show what your system did without revealing everything about how it works, you're in a corner. If you kept good logs, you can at least offer a record of inputs and outputs without exposing the model itself.

Authentication: is this evidence what you say it is

US courts have rules for authenticating evidence. Before a document, recording or data file can be admitted, someone has to show it's genuine: that it is what the party claims it is, and hasn't been altered ^[1].

AI-generated content makes that harder. Say a party offers a screenshot of a chatbot conversation, or a video that turns out to be a deepfake. The other side can challenge it, and if nobody can authenticate it the court might exclude it ^[1].

The guide points to a case where authentication was central: a party claimed an AI system had made a particular recommendation, but couldn't produce logs or a reliable record of what the system had output. The court was sceptical, and without corroboration the claim didn't hold up ^[1].

In the EU, the AI Act will require certain systems to identify their outputs as AI-generated, under Article 50 ^[3]. That's a marking duty, not an authentication mechanism. A mark can be stripped or faked, so it doesn't by itself prove the content is unaltered. That's where a sealed record at the point of generation would matter: you'd have a fingerprint of what was created, when, that anyone can check later.

Privilege and confidentiality

Another open question: when is a conversation with an AI protected by legal privilege ^[1]?

In the US, communications between a client and a lawyer can be privileged, meaning the other side can't force you to disclose them. But what if the communication goes through a chatbot, or an AI assistant that stores the exchange on a server? Some courts have said privilege can apply; others have been more cautious, especially if the AI provider can see the conversation ^[1].

The guide also notes that confidential business information shared with an AI tool might lose its protected status if the tool's terms allow the provider to use it for training or improvement ^[1]. That's a contracts question as much as an evidence one, but it affects what records exist and who can see them.

Jurisdiction: where do you sue

AI systems are often built in one jurisdiction, trained in another, hosted in a third and used in a fourth. US courts are still working out where a claim can be brought ^[1].

The guide points to cases where defendants have argued that the court lacks jurisdiction because the AI provider has no physical presence in the state, or because the harm occurred somewhere else. Plaintiffs have countered that the system was marketed or used locally, or that the provider does business through the internet ^[1].

There's no clear rule yet, and the uncertainty adds risk for both sides. For a company operating AI, it means thinking early about which courts might have a say if something goes wrong.

What this means for you

If your organisation uses or builds AI, Reed Smith's guide suggests a few practical steps:

• Decide now who keeps the record. If you can't show what your AI did on a given day, the other side's version might be the only one the court hears.
• Think about what discovery could demand. If you're asked for training data or model internals, do you have a way to show what the system did without giving away how it works?
• Check your terms with AI providers. If you're using a third-party tool, does the contract say what happens to your data, and whether conversations are privileged?
• Assume authentication will be challenged. If you're relying on AI-generated outputs as evidence, can you show they haven't been altered?

How a digital witness handles the record

Authentication and records are exactly the problems a digital witness solves. Here's how it works for the scenarios in Reed Smith's guide.

Say your customer service chatbot answers a question, and six months later the customer disputes what it said. You seal a record of each conversation as it closes: the messages, the timestamps and the model version. The content stays in your system. What gets sealed is a fingerprint.

If the conversation is challenged in court, you produce the sealed record and show that the fingerprint still matches. Change one character in the transcript and it won't. The court, the other side or an expert can check it, because the verification is public.

Or say you're facing discovery and the other side wants your training data. You can offer a sealed manifest of the data set: what it contained and when it was frozen for training. That gives the court a checkable record without handing over the data itself.

The witness doesn't prove the AI was right, fair or compliant. It shows that a record is unchanged since it was sealed. When the other side's argument is "you could have edited that log", a sealed fingerprint cuts the argument short.

You seal records through WitniumChain using the REST API, the SDK or the MCP server. The first 2,000 witnesses a month are free, and you can read the full documentation to see how it fits your case.

How it would work

1. Your system sends a fingerprint of each conversation, log entry or output to WitniumChain as it happens.
2. The fingerprint is sealed on a chain where any later change would show.
3. When the record is questioned, anyone can recalculate the fingerprint and check it against the sealed one.

What to watch

The AI Act's logging requirement in Article 12 applies to high-risk systems under Annex III from 2 December 2027 ^[2]. That's the date to have logging in place. How those logs will be authenticated in EU courts is still an open question, and Reed Smith's guide shows that US courts are working through the same problem now.

For systems that generate content, the marking duty in Article 50 applies from 2 August 2026 for general-purpose AI models ^[3]. Marking is about disclosure. Proving the content hasn't changed is a separate question.