AI is now used routinely in the production of sports and entertainment content. The EU AI Act (the “Act”) requires production companies, rights holders and others “deploying” AI systems to generate content to disclose when the content they have generated or manipulated is a “deep fake”. The rules, which have applied from 2 August 2026, reach into everyday production: virtual presenters, synthetic set extensions and AI-assisted graphics may all amount to a deep fake. While the Act is EU legislation, it can catch any business around the world which generates content using AI where it foresees that content being disseminated in the EU, including by posting it on the web.
The question of what a deep fake is, and the requirements for when and how a label must be displayed, are nuanced. The European Commission’s Guidelines (the “Guidelines”) explain how labelling is meant to work, but the examples in the guidance are limited and fairly obvious. So, what does this all mean in practice?
When realism becomes a deep fake
The Act defines a deep fake as AI-generated or manipulated content that resembles existing persons, objects, places, entities or events and “would falsely appear to a person to be authentic or truthful”. “Existing” is read broadly: it is enough that the subject could plausibly exist. As a result, the assessment usually turns on false authenticity. Considering some examples:
Likely to be a “deep fake”
A virtual travel show host – the Guidelines expressly treat realistic AI-generated avatars and personas as “persons”. Putting an AI-generated presenter in real locations, “reporting” in a format the audience expects to be genuine, creates exactly the false impression of authenticity that the definition of deep fake targets.
A comedy show depicting real politicians doing something they actually did, but in a fake, photorealistic environment– although the action is real, the fabricated setting means this is likely to be a deep fake. The assessment looks at the whole foreseeable audience, so content can qualify even if most viewers would see through it, but some, such as children, might be misled. It also makes no difference that the same effect could be achieved by non-AI production methods: if AI is used, the labelling obligation applies. As evidently satirical content, however, it should benefit from the lighter-touch disclosure regime discussed below.
A motorsport series’ photorealistic audio-visual rendering of a proposed new street circuit in a real city which is on the race calendar next year – again this would fall under the definition of a deep fake. Even though the new street circuit hasn’t yet been used or tested for the sport, the labelling obligation applies as it would appear real to many in the audience. Only the biggest fans of the sport are likely to know that the image isn’t real (because the sport has never raced on that city’s streets before) without more. The Guidelines state that a high degree of photorealism renders it more likely that such content should be considered a deep fake but this needs to be assessed in the context in which it is deployed. An “artist’s impression” label or other clear on-screen framing as an “AI visualisation” is likely to be enough.
Unlikely to constitute a “deep fake”
An AI-generated fantasy world appearing in a dream sequence in a soap opera – this is obviously fake, so no labelling is required. The Guidelines also indicate that films in which real actors perform against AI-generated backgrounds are unlikely to require labelling. Context is key here: the Guidelines make the potential to mislead, judged against the context and audience expectations, the core element of the test.
A “ghosting” graphic in head-to-head sports coverage – it’s now common to see graphics of two F1 drivers who raced at different times on the same track shown simultaneously, overlaying each other. Viewers understand this as an analytical composite of real footage, so this is unlikely to need labelling. The Guidelines make clear context and audience expectations are central to the assessment.
An AI-generated cartoon depicting real people – as no one mistakes a cartoon for real life, this would not be a deep fake of those people. A more interesting scenario might be where a production “translates” live, real action into a cartoon, for example, to appeal to a younger audience (as the NFL has experimented with) or to create an “alt-cast” (as with the Australian Open’s player avatar broadcast). If these are positioned as showing the real, live event, just in a novel format, they may fall within the definition of a deep fake as they “resemble” an existing event and may falsely appear to be truthful due to the manner in which they are positioned. If the content is not then accurate (for example, it shows the wrong player getting a red card), is this a deep fake? It’s unlikely that these and other sports-specific scenarios, including the use of virtual advertising, were on the minds of the drafters of the legislation. It seems sensible to take a risk-based approach in such situations – if people are likely to be misled, it is cheap insurance to include a label.
When, where and how to label
The Act requires disclosures to be clear and distinguishable, and accessible (including to vulnerable groups such as children and persons with disabilities), at the latest at the time of first exposure to the content. However, it doesn’t specify the form of the label. However, the official, but voluntary, Code of Practice on Transparency of AI-Generated Content (the “Code”) does provide solutions which signatories agree to adopt. The Code has been approved as adequate by the Commission and the AI Board, which means that signatories who follow it will be able to rely on it to demonstrate compliance with the Act.
For most deep fakes, the Code requires that signatories use an icon or label at the start of the content. It should also appear at regular intervals where possible, and after interruptions such as ad breaks at a minimum. This reflects the reality of live viewing, screenshots and clipped fragments. The default position of the label/icon is somewhere clear of other on-screen graphics, for example, the top right corner.
However, the Act permits a lighter-touch approach for evidently artistic, creative, satirical or fictional works. The Act provides that disclosure for these works should not hamper enjoyment or “normal exploitation” of the work, simply requiring disclosure in an appropriate manner. The Code indicates that this may be done in accompanying notes, or the opening or closing credits. On digital platforms, the Code allows the label to sit next to the frame or within the interface, including hover or click icons. But there is a limit to this flexibility; the disclosure must be noticeable without the viewer having to do anything.
Sports content, it would seem, will rarely qualify as a creative or fictional work. As such, the use of deep fakes within a sports context is likely to require the standard labelling approach for non-creative works.
Who has to do the labelling?
Under the Act, the obligation falls on the “deployer”. The Guidelines describe this as the party with authority over the decision to use AI and how it is used, which does not have to come with technical control. Businesses that only passively disseminate content, which can include broadcasters, are not deployers, though they are encouraged to preserve the labels applied by others.
- Production company to broadcaster or streamer: the production company is typically the deployer. A broadcaster or platform that directs or signs off the AI use may well be a deployer too. As we have noted in the advertising context, deployer status can sit with more than one party in the same chain.
- Rights holders and federations distributing direct via OTT or social channels: they are plainly deployers, including those outside the EU whose content foreseeably reaches EU audiences.
- Live real-time graphics: the host broadcaster or production entity commissioning the graphics will usually be the deployer, even where an outside vendor develops and/or operates the tools on its behalf. Note that vendors that supply an AI system for such uses may separately be a “provider” with obligations under the Act to “mark” outputs using machine-readable techniques.
The Guidelines expect deployers, such as the production companies and rights holders, to make sure their labels survive distribution, including through contractual conditions with distribution partners. Commissioning and licensing agreements should therefore cover who assesses each asset, which label is used, where it sits, who preserves it downstream, and who bears the risk. This is serious stuff. Fines can reach €15 million or 3% of global turnover, whichever is higher. Of course, a label does not cure other potential liabilities that can arise as a result of deep fakes and, in particular, IP, image, personality and data protection rights will always need to be considered when generating deep fakes.
