AI in law firms raises a new evidence question: what did the system say?

Lawyers get their time back

Legal AI has moved from trial to routine. A 2026 article in Tidningen Advokaten describes how Swedish law firms now use AI to draft agreements, summarise cases and speed up research, allowing lawyers to spend more time on the work only they can do ^[1]. The technology has, as one lawyer puts it, made lawyers into lawyers again.

That shift is welcome. It also brings a practical question that matters more as reliance grows: when someone later asks what the AI said or produced, can the firm show it?

What the AI Act says about legal AI

The AI Act distinguishes between systems used in general legal work and those used to materially influence legal outcomes or assess people. Most legal AI falls into the first category and is not classified as high risk ^[2].

Article 6(2) sets out the conditions for high-risk classification in the legal field. A system must be intended to assist a judicial authority in researching and interpreting facts and the law, and in applying the law to a concrete set of facts, or be intended to influence the outcome of an election, a legislative or rule-making process, or a judicial proceeding ^[2]. Systems used for purely administrative tasks, such as anonymising decisions or managing court scheduling, are explicitly excluded ^[2].

In practice, tools used by law firms to draft contracts, summarise documents or conduct research do not meet that threshold. They remain general-purpose AI, and the transparency requirements under Article 50 apply instead. From 2 August 2026, anyone using a general-purpose AI system must inform people when they interact with it, unless it is obvious from the circumstances ^[2].

When the record matters

The absence of high-risk classification does not mean the output is unimportant. A contract clause drafted by AI, a summary of a case file or a research note can all shape decisions. When a client questions the advice, or a counterparty points to a term they say was omitted, the firm needs to show what its system produced.

A screenshot is one side's evidence. So is an email forwarded later. Neither proves the text has not changed since the moment the AI generated it. If the firm kept no independent record at the time, it starts a dispute on the back foot.

The new Product Liability Directive, which in Sweden is proposed to enter into force on 9 December 2026, treats software as a product ^[3]. Courts can under certain conditions order defendants to disclose relevant evidence, and liability is strict for death, personal injury and certain property damage ^[3]. The directive applies to products placed on the market after that date. While law firms do not typically sell software, they do rely on it, and the principle holds: evidence that can be independently verified is harder to challenge than evidence kept only by the party relying on it.

Whose version stands?

Law is built on records. Agreements are signed, emails are dated and court filings are stamped. When a dispute turns on what was said or done, each side produces its version, and someone decides whose account is more credible.

AI complicates that pattern, because the record is often incomplete. A lawyer asks a system to draft a clause, reviews the output, edits it and sends it on. The final version is clear. What the AI originally suggested is not always kept, and when it is kept, it can be altered by anyone with access to the system.

That matters most when months have passed. A client says the advice was different from what the file shows. A counterparty claims a term was discussed but not included. If the firm has no sealed record of what the AI produced at the time, the question becomes which story sounds more plausible.

What a digital witness does

A digital witness takes a fingerprint of the AI's output at the moment it is generated. The content stays with the firm. The fingerprint is sealed on a record where any later change would show, even one made by the firm itself.

When the output is later questioned, the firm calculates the fingerprint again. If it matches, nothing has changed. Alter a single character and it will not match. The check can be done by anyone, so it does not require the other side to trust the firm.

That approach does not prove the content is true. It shows the content has not changed since sealing, and it shows the date and time of sealing. For a law firm, that is enough to answer the question: what did our system produce, and when?

What this means for you

• If your firm uses AI to draft, summarise or advise, decide now who keeps the record of what the system produced, before a client or counterparty does.
• Article 50 of the AI Act requires transparency about AI use from 2 August 2026. Check whether your templates and client communications make that clear.
• A sealed record per output gives you a fixed point when something is questioned. Consider whether your workflows capture that.
• The new product liability rules apply from 9 December 2026. While they do not typically affect law firms directly, they change the evidential landscape for any organisation that relies on software in its work.

What to watch

The AI Office is expected to publish further guidance on the scope of high-risk classification in the legal field during 2027. The Swedish government's draft bill on the new Product Liability Act is with the Council on Legislation, and the law is proposed to enter into force on 9 December 2026 ^[3]. For general-purpose AI, Article 50's transparency requirement has applied since 2 August 2026 ^[2].