The neutral witness: how parties can agree on what happened without trusting each other

Why we invented neutral witnesses

Two parties disagree about what was agreed. One says the price list sent in March had one figure, the other says it had another. Both are certain. Neither trusts the other's file.

This is an old problem. Societies invented notaries, witnesses and escrow agents to solve it. The idea was simple: bring in a third party who has no interest in the outcome, who writes down what happened and whose word both sides will accept later.

In Swedish law, courts assess evidence freely ^[1]. Everything can be presented, and the weight it carries depends on how reliable it is. A record kept by one of the parties can be challenged. A record kept by someone neutral carries more weight, because that person has less reason to lie.

The same principle runs through the EU's eIDAS Regulation. An electronic timestamp cannot be denied legal effect just because it is electronic or not qualified, and a qualified timestamp has a presumption that the date, time and integrity are correct ^[2]. The law recognises that a neutral third party, in this case a qualified trust service provider, makes a record more credible.

Why digital records made it harder

Digital files are easy to edit. Change a character in a document, adjust a figure in a log, or alter a timestamp in a photo's metadata, and the file looks just as real as the original. The other side has no way to tell.

Screenshots get questioned. Emails can be doctored. Logs kept by the organisation that benefits from them are easy to challenge. In the Air Canada chatbot case in 2024, the airline offered no evidence to contradict the customer's screenshot, and the tribunal used that against it ^[3].

When both sides keep their own records and both records differ, somebody has to decide whose to believe. Courts do that by weighing credibility. But before it gets that far, the parties need a way to agree on a fact. If they can't, the dispute drags on.

What a neutral digital witness does

A digital witness is a third party that records a fingerprint of something at a point in time. It doesn't keep the content, only a short string calculated from it. If the content changes later, even by one character, the fingerprint won't match.

Say a delivery driver photographs a damaged pallet at handover. The consignee later claims the damage happened in transit. The driver's phone sealed the photo with a digital witness at capture, together with the time. When the claim is disputed, anyone the driver shares the photo with can check the fingerprint. If it matches, the photo is exactly what was captured that day. If someone has edited it since, the check fails.

The witness doesn't take sides. It doesn't say the photo shows what the driver claims it shows. It says the file hasn't changed since it was sealed. That's all, and that's often enough. The parties can argue about what the photo means, but they don't have to argue about whether it's been edited.

WitniumChain is one such witness. It has been running in production since 2018. The content never leaves the user's systems. Only the fingerprint is sent, together with a reference the user chooses. The witness returns an ID, and from that point the record is sealed. Checking is public: anyone with the file and the ID can see whether they match. The first 2,000 witnesses a month are free, and it works through a REST API, an SDK or an MCP server ^[4].

What it doesn't do

A digital witness shows two things: that something existed at a certain time, and that it hasn't changed since. It does not show that the content is true.

Say a surveyor seals a damage report. The witness confirms the report existed on that date and hasn't been altered. It doesn't confirm that the damage described in the report actually occurred, or that the surveyor's assessment was correct. It doesn't authorise anything, and it doesn't make anyone compliant with any rule.

If the sealed content is a lie, the witness will faithfully record that lie, unchanged. The value lies in removing one argument: whether the record has been tampered with. What the record says, and whether to believe it, are separate questions.

Where it matters

Deliveries and claims

Insurance claims often turn on photos and documents captured long before the claim is filed. A photo of a damaged container sealed at the quay can be checked a year later by the claims handler, the insurer and, if it comes to it, a court. The marine insurance industry's body, IUMI, has a short film about an app called ASD.DigitalWitness, which uses this idea ^[5].

How it would work:

1. The surveyor photographs the damage and the app seals the photo with the time and location.
2. The sealed record is shared with the insurer when the claim is filed.
3. Anyone reviewing the claim can check that the photo is the one taken on the day, unaltered.

Contract versions and approvals

Disputes about which version of a contract, offer or price list was agreed are common. Each version sealed when it is sent gives both parties a record they can check later. If the buyer says version 2 was agreed and the seller says version 3, the sealed fingerprints settle which version existed when.

AI decisions and logs

High-risk AI systems under the AI Act must log events over their lifetime ^[6]. The logs record what the system did, and they can be questioned months or years later. A log the operator can edit is easy to challenge. Sealing log records as they are written, or sealing log segments at short intervals, gives auditors and supervisors a way to check that the log is what was recorded at the time.

Say a lender's credit-scoring model declines an application and the applicant asks for an explanation six months later. The lender's log shows the input, the model version, the score and the reviewer. If that record was sealed when it was written, the lender can show it's the same record, unchanged. If it wasn't sealed, the applicant's representative can argue it was written after the fact.

Chatbot answers

From 2 August 2026, people interacting with an AI system must as a rule be told they are doing so ^[7]. When a customer questions what a chatbot said, the customer's screenshot is often the only evidence. Sealing each answer, or each conversation when it closes, gives the organisation its own record. Both sides then have evidence, and neither has to trust the other's version.

Software versions

Under the new EU Product Liability Directive, software placed on the market after 9 December 2026 counts as a product ^[8]. When something goes wrong, the question may be which version was running on a given day. Sealing each release and each deployment event gives the developer a record that can be checked later, by customers, notified bodies, insurers or courts.

What this means for you

Before choosing any form of digital witness, ask:

• Is it independent? Can the party keeping the record also change what the witness saw? If so, it isn't neutral.
• Is it permanent? Will the sealed record still be checkable in five years, or ten?
• Can anyone check it? A witness that only the provider can verify forces you to trust that provider.
• What exactly does it seal? The content, or a fingerprint of it? If the content, where does it go?
• What does it cost after the free tier? Predictable pricing matters when sealing becomes part of daily operations.

For your own systems, think about what you would need to show if someone questioned a record a year from now. Could you prove the log, the document or the decision is exactly what was recorded at the time? If the answer is no, or if it depends on the other side trusting you, a neutral witness may be worth considering.

What to watch

The AI Act's logging requirements for high-risk systems in Annex III apply from 2 December 2027 ^[9]. For systems under Annex I, the date is 2 August 2028 ^[10]. Neither article says how to protect the logs, but the question of their evidential value will come up in audits and disputes.

The new Product Liability Directive gives courts the power to order defendants to hand over relevant evidence under certain conditions ^[11]. When that evidence is a log or a version history kept by the defendant, being able to show it hasn't been edited strengthens the defence.

DORA's technical standard on ICT risk management, which has applied to financial entities since January 2025, requires logs to be protected against tampering, deletion and unauthorised access ^[12]. Sealing doesn't replace access control or monitoring, but it answers a question those measures can't: whether someone with the right access changed the log anyway.