Insights

    AI Regulation

    Telling users they are talking to an AI is now a product requirement, not a courtesy

    Devence Lab

    · 2 min read

    Share
    Telling users they are talking to an AI is now a product requirement, not a courtesy
    Photograph · Unsplash

    Transparency rules applying from 2 August require interactive AI systems to disclose themselves and generated content to be labelled. The engineering is trivial; the product consequences are not.

    The transparency rules that began applying on 2 August require that certain AI systems tell users they are interacting with AI, and that generated or altered content be labelled — chatbots must disclose, deepfakes must be marked.

    The implementation is a line of copy. The interesting part is everything that follows from it.

    Disclosure changes behaviour, which is the point

    People interact differently with a system they know is automated. They phrase requests differently, trust outputs differently, and escalate sooner. Any product metric measured before disclosure was measured under different conditions and should be expected to move.

    For teams that quietly replaced a human channel with an automated one and watched satisfaction hold, disclosure is the test of whether it held because the system was good or because nobody knew.

    Disclosure does not change what the system does. It changes what the user believes, and the metrics were measuring the belief.

    Labelling has a harder edge

    Content labelling raises a question most pipelines cannot currently answer: what counts as generated or altered. A fully synthetic image is easy. A photograph with AI-assisted retouching, a human-written article edited by a model, a summary of a real document — these sit on a spectrum, and the obligation does not resolve where the line falls for every case.

    The engineering implication is provenance tracking. To label reliably you must know, for any given artefact, what touched it — which means capturing that at creation time rather than inferring it later. Most content pipelines record no such thing, and retrofitting provenance is considerably harder than adding a label.

    What we would do now

    Find every interactive AI surface and confirm disclosure is present, unmissable and at first contact rather than buried in terms. Then look at content pipelines and ask whether you could state, for anything published, what generated or modified it.

    Where the answer is no, that is the actual project. The label is a five-minute change; knowing what to put on it is an architecture question.

    Sources

    1. Commission starts enforcing AI Act rules and new transparency requirements on 2 AugustEuropean Commission
    2. Safer and more transparent AIEuropean Commission
    3. EU AI Act 2026 Updates: Compliance Requirements and Business RisksLegal Nodes

    Written by the Devence Lab research team.

    Share

    Collaborate

    We share findings with partners operating in the same constraint space.

    Get in touch