When a fake viper photo sparked a real search: the evidential cost of unlabelled AI

The incident

In March 2025, an image reportedly showing a viper was shared on social media in France ^[1]. According to the AI Incident Database, the image prompted a police alert and a search operation. The photograph was later identified as likely generated by artificial intelligence, not taken with a camera ^[1].

The case is instructive not because it was dramatic—nobody was harmed and the search was called off—but because it shows what happens when an authority has to act on digital content without knowing where it came from. The police had a plausible image and a location claim. Verifying whether the image was real or machine-generated took time, and resources were spent before that verification was complete.

The transparency rules since August 2026

Since 2 August 2026, the European Union's AI Act has required transparency when AI systems interact with people or generate content ^[2]. Article 50 sets out three main duties.

First, providers must ensure that people are informed when they are interacting with an AI system, unless that is obvious from the circumstances ^[2]. Second, AI-generated content must be marked in a machine-readable format, and systems that generate such content must be designed to allow that marking ^[2]. Third, users must disclose when image, audio or video content is a deepfake, and when AI-generated text is published to inform the public and no human has reviewed it and taken responsibility for it ^[2].

The rules say what must be disclosed. They do not say how to prove later that the disclosure was made, or that the content has not been altered since.

The evidential question

When a court or an authority tries to establish what happened, it faces two separate questions: what does the evidence say, and can the evidence be trusted.

Under Swedish procedural law, courts assess evidence freely ^[3]. Anything may be submitted, and its value is judged by how reliable it is ^[3]. A sealed document is not automatically more persuasive than an unsealed one, but it can be checked. If someone presents a photograph and claims it was taken on a given day, and a digital witness shows the same fingerprint was sealed on that day, the claim becomes easier to verify. If the photograph has changed, the fingerprint will no longer match.

The viper incident illustrates the inverse. Without a way to verify the image's origin, the authorities had to treat it as potentially real until they could determine otherwise. The time and effort required to make that determination is the cost of missing provenance.

What happens when deepfakes reach authorities

Deepfakes have been used to manipulate share prices, impersonate executives and spread disinformation during elections ^[1]. Most incidents documented so far involve financial fraud or reputational harm. The viper case is unusual because it reached a public authority and triggered an operational response.

The AI Act's transparency requirements aim to reduce such incidents by making it clear when content is machine-generated. Disclosure helps people decide how much weight to give an image or a statement. It does not, however, prevent someone from removing the label or sharing the content without it.

That is where the evidence question returns. A machine-readable label can be stripped away. A fingerprint sealed at the time of creation cannot be removed from the record, and any change to the content shows up when the fingerprint is recalculated. The label tells you what the content is. The fingerprint lets you show it has not been altered.

The cost of uncertainty

The French police did what they had to do: they treated the report as credible until they could verify otherwise. The cost was measured in hours and resources. In a legal dispute, the cost might be measured in settlement terms or in the strength of a party's position.

Consider a scenario. A company uses an AI system to generate contract summaries for clients. One client later claims the summary was different from the one they received. The company has a copy of the summary. The client has a screenshot. Neither side can show that their version has not been altered since the date in question. The court assesses both freely, but neither carries much weight.

Now consider the same scenario with a digital witness. The company sealed the summary when it was sent, and the client received the same summary with the same fingerprint. If the client's screenshot matches the sealed version, both parties can agree on what was sent. If it does not match, the sealed version provides a reference point. The dispute may still go to court, but at least the content of the document is no longer in question.

What this means for you

If your organisation generates content with AI, or relies on digital content as evidence, ask:

• Do we mark AI-generated content in line with Article 50, and can we show we did so?
• If someone questions an image, a document or a log, can we show what we had at the time, and that it has not changed?
• Do we have a process for sealing content when it is created or received, so that we hold the record, not just the file?

The AI Act requires disclosure. Whether that disclosure holds up as evidence is a separate question, and one worth answering before it is tested.

What to watch

The European Commission is expected to adopt implementing acts on the machine-readable marking of AI-generated content. The technical standards will specify formats and methods. When those acts are published, review your systems to confirm they can produce the required markings and that you can demonstrate compliance.

For high-risk AI systems under Annex III, the logging requirements in Article 12 apply from 2 December 2027. Those logs must allow for the automatic recording of events over the system's lifetime. The rules say what to log, not how to protect the logs, but their evidential value is a reason to consider sealing them.