Not the satire: the supervisor challenges how the notice was designed
7 min read
If your organisation generates or alters, with AI, the image or voice of a real person — a campaign, a demo, an internal video, an assistant speaking with someone’s voice — who decided how viewers get told? And does that notice work for someone who tunes in halfway through, or only for someone paying attention from the start?
Only after that comes the ruling: in recent days Italy’s data protection supervisor, the Garante, published a decision answering exactly these two questions, finding a breach of GDPR articles 5 and 25 against R.T.I. Reti Televisive Italiane S.p.A., on a complaint under GDPR article 77 by the journalist who anchors an evening news programme, over segments of “Striscia la notizia”.
What the Garante is NOT challenging
Before the breach, it is worth stating as plainly as the decision itself: the right to satire is not at issue here. The panel — Stanzione, Cerrina Feroni as rapporteur, Ghiglia, Montuori — states that the processing “must be treated as falling among those carried out in the exercise of freedom of expression”, so articles 136-139 of the Privacy Code apply to journalistic processing and other forms of expression, “including the right to satire”, a regime that “provides specific safeguards and cautions … confirming their lawfulness, even where carried out without the data subjects’ consent, provided they respect the rights, fundamental freedoms and dignity of the persons” portrayed (our translation; the original is in Italian). No consent required, then: the Garante is not saying a deepfake is unlawful in itself. It is saying something else.
The facts
The complaint “concerns the allegedly improper use of the complainant’s professional image as a journalist to introduce, during several episodes of the programme ‘Striscia la notizia’, segments built around events involving various public figures, accompanied by comments of various kinds spoken by the complainant’s deepfake, dubbed using artificial intelligence systems.” In the segments “the complainant’s real image was used, filmed, moreover, inside the television studio from which he anchors an evening news programme, altering its authenticity through dubbing carried out with tools such as not to make the existence of that alteration clearly perceptible.”
The defence
R.T.I. defended itself on two fronts. On the content: the implausibility and hyperbole “as in this case” would be “immediately perceptible to the public”; the only activity carried out: “dubbing Mr Mentana with a voice different from his own, using computer systems, on footage acquired under licence and duly paid for”. On article 25: “television viewers had, from the outset, all the information needed to understand that the videos broadcast were not showing the real Mentana, but only his deepfake.”
Two defences to keep separate: whether the public recognises the fiction, and whether it knows in time. The Garante did not challenge the first. It challenged the second.
The test: the least attentive viewer
The heart of the decision: the on-air notices were not “sufficiently clear, at least with reference to the technological knowledge of an average or otherwise inattentive public”; “the contextual data, together with the content spoken by a well-known journalist’s deepfake, would have required more explicit communications, made more prominent at the various points when the segments aired, so as to reach even the least attentive users, or those who had tuned in after the programme had already started.”
Read that twice: it overturns the standard most organisations use to design a notice. It is not enough that it exists at the start, for someone watching from minute zero. It has to work for someone tuning in halfway through, half paying attention. The reference audience is not the ideal user: it is the least attentive one, the one who arrives late.
Article 25, and why it matters beyond television
On this basis the Garante found “that the processing described amounts to a breach of the principles of lawfulness, fairness and transparency in data processing under article 5 of the Regulation, as well as of article 25 on data protection by design and by default, in relation to the failure to adopt adequate technical and organisational measures” to protect those rights.
Outside television: article 25 calls for a notice designed for the real audience, not the ideal one. A banner shown only in the first seconds, small print at the foot of a page, a disclosure shown once and never repeated: measures that exist but were not designed for someone arriving later or scrolling fast. The breach is not in the synthetic content — it is in the failure to design the measure that should have accompanied it, and it applies equally to a campaign made with AI, a demo using a cloned voice, a voice assistant using a familiar voice.
How we check it
We do not start from a view on what satire is lawful: we start from a check on how the notice is designed, run against synthetic content already published or about to be. The first session produces a dated inventory of the organisation’s synthetic content — what it is, who approved it, how it is marked, and whether the marking holds up for someone arriving halfway through — and it stays with you even without going further. No cost, no other commitment.
See the service · Talk to an engineer
The measures, and why a reprimand
Three measures. The Garante “orders the prohibition of any further processing” of the complainant’s data as described, “save for their mere retention for possible use in judicial proceedings” (article 58(2)(f)); “the measure of a reprimand” “for failure to comply with the provisions on data processing, with particular regard to the measures to be adopted to ensure compliance with the general principles of lawfulness, fairness and transparency” (point (b)); and “an entry in the Authority’s internal register” (article 57(1)(u)). No fine: thirty days to report steps taken.
Why a reprimand? The Garante is explicit: “this being a legal question characterised by novelty, arising from the application of recently introduced technological tools whose contours, especially with reference to their use within entertainment programmes and hence traceable to artistic expression.” A leniency with an implicit deadline: the same novelty will not excuse the next breach. If the prohibition is not observed, “the criminal penalty under article 170 of the Code may apply, in addition to the administrative fine under article 83(5)(e) of the Regulation.”
From 2 December, it won’t just be the GDPR
The standard drawn from article 25 — a notice designed for the least attentive viewer — is about to become a product obligation outside the GDPR too. Article 50(2) of the AI Act requires marking synthetic content in a machine-readable format. For systems already on the market before 2 August 2026, the transitional period ends on 2 December 2026 (article 111(4), Regulation (EU) 2026/1744), covering systems generating “synthetic audio, image, video or text content” (we’ve already written about this).
The Garante used article 25 to say, four months early, what the AI Act will say by product obligation from 2 December: technical marking, not a notice left to the writer’s discretion. Whoever prepares now does it once.
The two pillars, applied
Comply. Marking synthetic content, and proving it was designed for the least attentive viewer, becomes a check on the client’s publishing systems — what was marked, with what technique, who approved it — with a dated trail for an inspection, not an opinion filed away.
Decide. The same set-up brings together archives, business systems, documents and publishing systems into a single operating model, on which AI agents execute decisions with a human operator in command, for large enterprises, defence, the public sector and healthcare: a new marking rule turns into an answer, within hours, to “which content is affected, and who approved it.”
Always in two modes: on-premises, on self-contained machines without deep integration into the client’s network, or a dedicated cloud with a data centre in Italy — always with joint management, as for a public-sector body publishing AI content.
Could you say, right now, how many pieces of synthetic content your organisation has published in the last twelve months, who approved them and how they are marked? Let’s talk: the first session, at no cost, produces a dated inventory of your synthetic content — what it is, who approved it, how it is marked, and whether the marking holds up for someone arriving halfway through.