Is AI-generated advertising allowed? The 2026 rules

On 2 August 2026 the disclosure rules of the EU AI Act came into force, and the question changed shape. Before that date, whether you labelled AI work was a matter of each platform's house rules, and you could brief around them. Article 50 puts the duty on the deployer. If you use AI to generate or manipulate content that counts as a "deep fake," you disclose that it is artificial. That duty sits with the advertiser and the agency, not with the platform the work runs on.

The term "deep fake" sounds narrow, and this is where most campaign plans go wrong. The Commission's final Guidelines read it broadly: realistic AI depictions of products, places, animals, events and human avatars, even where no real person appears. The Guidelines give an AI product image that flatters the product as an example. They give a synthetic influencer demonstrating a sponsored product as another. There is a reduced-disclosure carve-out for artistic and creative work, and it generally does not rescue advertising, because the commercial character of the piece prevails. Obviously fantastical content falls outside. So does routine colour and noise correction. A photoreal AI hero shot of a real SKU does not.

Disclosure has to be clear, distinguishable, and given at first exposure, which rules out burying it in a caption that loads below the fold. Penalties run up to €15m or 3% of worldwide annual turnover. Full summary of the guidance here.

YouTube

YouTube's test is realism, not AI usage. Creators must disclose AI-generated or AI-altered content that seems realistic: a real person appearing to say or do something they didn't, altered footage of a real event or place, a realistic scene that never occurred, or AI music as the main focus of the video. Non-realistic content is exempt, and so is a long list of "minor edits" that covers a great deal of normal post: beauty and colour filters, captions, upscaling, audio repair, cloning your own voice for voiceover, and using AI for scripts, thumbnails or ideation.

Labels appear either on the player or in the expanded description, and YouTube may apply them automatically to content that arrives carrying C2PA metadata. Worth telling a nervous client: disclosure by itself does not limit reach or monetisation. Persistent non-disclosure is what risks penalties. YouTube's policy.

TikTok

The clearest advertiser mandate of the four. TikTok Ads Manager documentation lists AI-generated, synthetic or manipulated media as a case where a disclaimer is mandatory, covering images, video or audio that are "completely AI-generated" or real source material "significantly modified by AI." In practice it is a disclaimer toggle in the ad setup that adds a clickable textual label. The page was last updated in September 2025 and has not moved since. The disclaimer documentation. TikTok was also first to attach C2PA Content Credentials and auto-label AI content arriving from other platforms, back in May 2024.

Meta

Read this one carefully, because it is the most commonly misquoted. Mandatory advertiser disclosure at Meta applies only to ads about social issues, elections or politics. For ordinary commercial ads there is no declaration duty. The mechanism is automatic labelling. Meta applies an AI info label when advertisers use its own Background Generation, Image Generation or Add Animation tools, and it detects third-party AI through C2PA metadata, surfacing that under "About this ad." From 1 June 2026 it also adds information to ads it detects were made with third-party AI tools. Meta's political ad rules; reporting on the commercial tags.

Google Ads

Also mostly tooling. The only mandate remains the political one: election ads containing synthetic or digitally altered content that inauthentically depicts real or realistic-looking people or events must have the "Altered or synthetic content" box checked, effective July 2024. On 9 July 2026 Google added a "How this ad was made" panel in My Ad Center across Search, YouTube and Discover. For ads made with Google's own generative advertising tools, that disclosure is automatic. For advertisers using other tools, it is a voluntary control. Google points to its SynthID watermarking underneath. Defaults and plumbing, not a general duty. Policy; the announcement.

The distinction marketers keep collapsing

A duty to disclose and automatic labelling are not the same instrument, and confusing them is how campaigns get pulled.

• A duty means you must act. The EU AI Act and TikTok's advertiser disclaimer both work this way. If you do nothing, you are exposed.
• Automatic labelling means the platform acts, if it can detect the work. Meta's commercial ads and most of Google's new panel sit here. If detection misses, no label appears, and no label is not the same as compliance with Article 50.

The practical failure case is a campaign that relies on a platform to surface provenance it cannot actually read. Provenance metadata is fragile. It is stripped by re-encodes, crops and third-party editors, and platform rendering of third-party C2PA is mainly operational for content made on-platform.

Meta made you responsible for what its own AI does

Meta's updated Commercial Terms, effective 30 October 2026, added an "AI Features" clause making advertisers "solely responsible" for what Meta's AI features create or alter. That includes features Meta switches on itself. Bookshop.org found an AI illustration added to an ad it had made by hand. Exverus pulled live ads to strip changes it had not enabled. Meta described the update as housekeeping and noted that its AI "may make mistakes." The reporting. If your creative is subject to a legal disclosure duty, you now also carry the risk of AI you did not commission being added to it after approval. Audit live placements. Turn the enhancement features off where you can.

What the voluntary layer recommends

The IAB published its AI Transparency & Disclosure Framework v2 on 18 August 2026. It is voluntary, with no enforcement authority behind it, and it is the most usable document of the set because it is materiality-based: "not every use of AI needs a label."

“A duty to disclose and automatic labelling are not the same instrument, and confusing them is how campaigns get pulled.”

It recommends disclosure for prompt-generated images and video even where a human edited afterwards, synthetic voices of deceased people even with estate permission, synthetic voices of living people in events that never happened, digital twins of the deceased in any advertising, digital twins of living people in fabricated scenarios, fully AI-generated avatars mistakable for humans, and conversational agents simulating humans. It exempts routine post-production, clearly fictional or stylised characters, generic narration voices, background music, and standard copy and translations.

On placement it asks for a visual indicator on the first frame and sustained through the piece, and for audio, a verbal disclosure before or after the AI segment, repeated past sixty seconds. A sparkle icon or plain text such as "AI-generated image" both qualify. It names C2PA content credentials as the machine-readable layer with watermarking as complementary, and then says the thing worth printing out: where platform rendering cannot be relied on, burn a visible disclosure into the creative. The advertiser remains accountable even when a platform renders the label. The framework.

Three other laws are worth a line. The California AI Transparency Act (SB 942) took effect the same day as Article 50, but its obligations sit on AI developers with over a million monthly users rather than on advertisers. It shapes what metadata travels with the outputs you receive. New York's synthetic performer law took effect in June 2026. South Korea has had labelling mandates since early 2026.

What this means when you brief a studio

Settle four things before anyone generates a frame.

• Where it runs, and in what order. EU distribution pulls in Article 50 regardless of what the platform asks. A film cut for TikTok needs the disclaimer toggle. The same film on Meta may get labelled automatically, or not at all. Decide the disclosure per placement, at briefing, not at trafficking.
• Which uses of AI are in scope. A photoreal synthetic product shot and an AI-assisted grade are not the same decision. Ask the studio to list what the model generated versus what it cleaned up. We write that down per shot, because you cannot disclose accurately what nobody tracked.
• Who owns the visible mark. If you cannot verify that a platform will render your provenance metadata, put the disclosure in the frame. It costs a design decision at concept stage and nothing at all later.
• Who checks the live ad. After the Meta clause, approval is not the end of the job.

One caution on the business case for labelling. A NYU/Emory field study on the Google Display Network is widely reported, including in the IAB framework, to have found that labelling an ad "AI-generated" or "AI-edited" cut click-through by around 31.5% against an unlabelled human-made ad. We could not locate the primary paper, the full author list or a sample size, so treat it as trade reporting rather than an established finding. We mention it because it is the number a client will raise, and the honest answer is that we do not know how solid it is. What we do know is that a pulled campaign costs more than a label.