Ontario court to formalise *AI evidence rules* in civil procedure

A court deciding what counts

Ontario's Superior Court of Justice is writing formal civil procedure rules on AI-generated evidence. When they come into force, a party that wants to introduce evidence created by an AI system will have to identify the system, describe its training data, disclose the inputs and prompts given to it, and provide supporting evidence of the output's validity and reliability ^[1].

The draft rules follow cases where courts have rejected AI-generated evidence because no one could explain how the system worked or whether its output was trustworthy. One judge described it as like asking the court to trust a black box ^[1].

Think of it like a receipt from a chemist. If someone turned up in court with test results, the other side would want to know which lab ran the test, which method it used, what the sample was and whether the lab is accredited. The new rules ask the same questions about AI: which system, trained on what, given which inputs, and why should we believe the answer.

What the rules will require

The draft rules set out what a party must disclose when it wants to rely on AI-generated evidence ^[1]:

• the name and version of the AI system,
• a description of its training data,
• the inputs and prompts given to it,
• supporting evidence of the output's validity and reliability.

The rules are aimed at evidence that an AI system has created, such as an accident reconstruction, a financial analysis or a translation. They are not about logs of what an AI system did, such as a chatbot's answers or a screening model's decisions.

That distinction matters. A log is a record of the AI's own actions. AI-generated evidence is something an AI has been asked to produce for the case.

Why courts are cautious

Courts have good reason to be careful. Generative AI can produce plausible-looking outputs that turn out to be wrong, and it is often hard to tell why a system reached a particular answer.

In one Canadian case, a lawyer submitted research that included citations to cases that did not exist, generated by ChatGPT ^[2]. In a US case, a court rejected an accident reconstruction created by AI because the expert could not explain the system's methodology or verify its accuracy ^[1].

The Ontario rules are a response to that problem. They say, in effect, that if you want the court to rely on something an AI created, you have to be able to explain where it came from and why it is trustworthy.

The training data question

One detail stands out: the requirement to describe the AI system's training data.

For most commercial models, that is easier said than done. Training data sets for large language models can include billions of documents, and providers rarely publish full lists of what went in. A party trying to meet the rule may have to rely on whatever the provider has disclosed, which may not be much.

That puts the burden on the side that wants to use the evidence. If you cannot describe the training data, the court may decide the output is not reliable enough to admit.

What this means for you

If you are in Ontario and you plan to use AI-generated evidence in a civil case, or if you advise someone who does, the draft rules give you a checklist:

• Can you name the AI system and its version?
• Can you describe, even in broad terms, what it was trained on?
• Can you show what inputs and prompts were given?
• Can you provide supporting evidence that the output is valid and reliable, such as expert testimony or independent verification?

If the answer to any of those is no, the evidence may not get in.

The rules also matter for anyone building or selling AI tools that might be used to generate evidence. Transparency about training data, methodology and limitations is no longer just good practice. In Ontario, it may soon be a procedural requirement.

Sealing what the AI did

The Ontario rules are about evidence an AI has created. They are not about proving what an AI system itself did.

Say a party wants to show that a chatbot gave misleading advice, or that a screening model rejected an application. The question there is not whether the AI's output is valid, but whether the record of what it said or did is accurate and unchanged.

That is where a digital witness fits. The thing to seal is the AI's output at the time it was generated: the answer, the decision, the score. Sealing takes a fingerprint of the record, not the record itself, so the content never leaves your systems. If the record is questioned later, anyone you share it with can check that it is exactly what was recorded at the time.

A digital witness does not show that the AI's answer was correct or that the decision was fair. It shows that the record has not been altered since it was sealed. That matters when the dispute is about what the AI actually said or did, not about whether what it said was right.

You can seal outputs as they happen, through the REST API, the SDK or the MCP server that connects to AI agents. WitniumChain keeps the fingerprint, you keep the record, and anyone can check them against each other. The first 2,000 witnesses a month are free.

How it would work

1. Your system generates an output (a chatbot answer, a model decision, a document).
2. It sends a fingerprint of that output to WitniumChain and gets a witness ID back.
3. When the output is questioned, you share the record and the witness ID. Anyone can check that the fingerprint still matches.

What to watch

The Ontario rules are not yet in force, and the final wording may change. Once they are adopted, expect other jurisdictions to look at similar measures.

In the EU, nothing yet requires the kind of disclosure Ontario is proposing for AI-generated evidence. But the courts' reluctance to accept unexplained AI outputs is the same everywhere. If you cannot show how an AI reached its answer, a judge is unlikely to rely on it.

For logs of what an AI did, Article 12 of the AI Act requires high-risk systems to allow for automatic recording of events, and applies from 2 December 2027 for Annex III systems. Article 12 says what to log. It does not say how to protect the logs or prove they have not changed.

That is the evidence question the Ontario rules do not address, and the one a digital witness answers.