AI labels on social media: EU AI Act and Instagram rules 2026

If you post AI-generated images or video for clients, labelling stopped being a matter of taste in 2026. Since 2 August the EU AI Act's transparency rules apply, and every major platform has its own label and its own idea of when to use it. The two overlap but aren't the same, and the gap between them is where agencies get caught out.

Here is what each asks for, and a checklist to run on every AI asset. This is not legal advice: for a borderline campaign, ask a lawyer.

The EU AI Act: what Article 50 says

When it applies

The transparency rules in Article 50 of the AI Act (Regulation (EU) 2024/1689) apply from 2 August 2026 (European Commission). The Digital Omnibus on AI (Regulation (EU) 2026/1744) added a grace period until December 2026, but only for the machine-readable marking that generative AI tools already on the market must build into their output. That's a duty for the companies that make the tools, and it doesn't delay the labels on your posts. Deep fakes made before 2 August 2026 don't need labelling retroactively, though the Commission encourages it (same source).

Is it you or your client?

The labelling duty falls on the "deployer": whoever uses an AI system under their authority, outside purely personal use. In guidelines published on 20 July 2026 (European Commission), the Commission says authority means deciding to use the system and how its output is used; technical control isn't needed (Commission guidelines, para. 12). Two of its examples speak directly to agencies (para. 14):

• An advertising company is the deployer of the AI it uses. Its employees, and freelancers working under its control, are not separate deployers.
• A company that "merely commissions an advertising agency to produce an advertisement, without taking decisions and exercising control over whether and how the advertising agency uses AI" is not a deployer.

So if you decide to make a client's visual with an AI tool, the duty is most likely yours. If the client decides how AI is used, the client may be a deployer too. Write down who decides.

The rules also reach deployers outside the EU who direct AI content at people in the EU, including by posting deep fakes on the open internet (para. 13). An agency in New York posting AI visuals for a Milan client is very likely covered.

What has to be labelled: deep fakes

The Act defines a deep fake as AI-generated or manipulated image, audio or video content "that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful" (guidelines, para. 113). Two points in the Commission's reading surprise most marketers:

• The subject doesn't have to exist. It's enough that it "can plausibly exist", and "persons" includes "realistic AI-generated human avatars or personas" (para. 113). A realistic AI model wearing your client's coat counts.
• Intent doesn't matter. The test is whether the content could mislead, not whether you meant it to (para. 114).

From the guidelines' examples, these are deep fakes:

• an AI video of a celebrity influencer in an advertising context;
• a realistic synthetic avatar of a company's CEO thanking staff;
• an AI image of a product that makes it look different, more appealing or better than it really is.

These are not:

• an AI video of mice arguing about the best cheese, for a cheese brand;
• a real product, such as a car, shown against an AI-generated background, as long as the ad doesn't mislead about the product.

Routine edits usually don't count either. Colour correction, removing passers-by, noise reduction or an aesthetic background swap in a product ad are "likely to have only a minor impact" on how authentic people find the ad (para. 116).

Creative work gets a lighter touch. Most ads don't

Deep fakes in evidently artistic, creative, satirical or fictional work still need disclosing, but "in an appropriate manner that does not hamper the display or enjoyment of the work" (Article 50(4)). The Commission reads this narrowly: its examples that don't qualify include a "realistic synthetic influencer testing out a sponsored real product" (guidelines, section 6.1.3).

AI-written text

Text is covered only when it's published to inform the public on matters of public interest, and not when it "has undergone a process of human review or editorial control" with someone holding editorial responsibility (Article 50(4)). The Commission's examples put ad copy and product descriptions outside the rule, unless they make claims about things like health, consumer safety or sustainability (guidelines, section 6.2).

So most client captions need no EU label. A post giving health information or commenting on public affairs is different: have a person really review it and own it. Spell-checking, automated review or "cursory editorial approval without substantive engagement" don't count (para. 135).

How and when to label

The disclosure must be "clear and distinguishable" and given no later than the first exposure (Article 50(5)). The Commission's own example is "when encountering AI-manipulated deep fakes when scrolling on social media"; for video, it's the start of the clip (guidelines, section 7.2). Metadata alone isn't enough, because people can't see it (para. 117).

Where a very large platform offers a labelling tool that gives a clear disclosure, deployers "can rely on such tools" there, though the responsibility stays theirs (para. 126). In practice: use the platform's AI label, and add your own words where it might be missed.

A voluntary Code of Practice, published on 10 June 2026, includes an EU set of icons deployers may use. Signing it is optional, but "the transparency requirements under article 50 of the AI Act are legal obligations" (European Commission). Fines can reach €15 million or 3% of worldwide annual turnover, whichever is higher; for small and medium-sized businesses, whichever is lower (guidelines, para. 152).

What each platform asks for (as of October 2026)

The law sets a minimum in the EU. Platform rules apply wherever your client posts.

Instagram, Facebook and Threads. Meta shows an "AI info" label when it detects AI from industry-shared signals or when the person posting disclosed it. It requires people to use its disclosure tool "when they post organic content with a photorealistic video or realistic-sounding audio that was digitally created or altered", and "may apply penalties" if they don't (Meta). Still images aren't in that requirement, but under the AI Act a realistic one may still need a label.

Instagram profiles built around an AI person. Since 31 August 2026 Instagram's label for profiles that present an AI-generated person is called "AI-generated profile" (it was previously the "AI creator" label). Profiles that should carry it and don't can have their reach limited. Creators who simply use AI tools in their process don't need it (Instagram).

TikTok. Creators must label "AI-generated or significantly edited content that shows realistic-looking scenes or people", and "unlabeled content may be removed, restricted, or labeled by our team" (Community Guidelines in force since 24 September 2026, TikTok). TikTok also labels uploads automatically when it reads C2PA Content Credentials (TikTok Newsroom), and creators can't remove an automatic label (TikTok Support).

LinkedIn. Media carrying C2PA Content Credentials shows a small icon revealing its origin, and in select countries you can self-attest AI use on upload. LinkedIn admits "it's not yet possible to identify and label all AI-generated and modified content" (LinkedIn Help).

Pinterest. An "AI modified" label appears when Pinterest reads IPTC metadata, when its classifiers detect AI, or when the content owner indicates AI use (Pinterest Help).

X. X may add a "Manipulated media" label to media it judges significantly and deceptively altered or fabricated (X Help). For realistic AI media on X, the safe habit is to say it in the post itself.

A checklist for every AI asset you post for a client

1. Could this pass as real? A realistic person, place, product or event gets a label. Clearly impossible or cartoonish content usually isn't a deep fake, though platforms may label it anyway.
2. What did the AI change? Colour, clean-up or a new backdrop behind a real product: usually fine. A product that looks better than it is, or a person doing something they never did: label it, and ask whether to post it at all.
3. Switch on the platform's label: Meta's AI disclosure, TikTok's AI-generated content setting, LinkedIn's self-attestation where offered. If your scheduling tool can't pass that setting, publish that post directly in the app.
4. Say it in words too, where people look first: "The model in these photos is AI-generated." For video, an on-screen note in the opening seconds.
5. Keep the metadata. Stripping C2PA or IPTC data to dodge an automatic label doesn't remove your obligation.
6. Have a person properly review AI text on public-interest topics, and use Instagram's AI-generated profile label for any account fronted by an AI persona.
7. Keep a log and agree it in writing: which tool, which posts, which label, who signed off, and who decides when AI is used.

A caption that does the job without spoiling the post:

The Harbour coat in navy, styled three ways for autumn. The model is AI-generated; the coat is real, photographed in our studio.

And a note to the client before the campaign:

Hi Marco, a quick note before the autumn campaign. Since August, EU rules require realistic AI images and video, such as an AI model wearing your coats, to be clearly labelled, and Instagram and TikTok have their own labels too. We'll mark every realistic AI visual with the platform's label and a line in the caption, and we won't use AI to make a product look different from the real one. If you'd prefer real people only, tell us and we'll plan a shoot.

Build the label check into sign-off, so it happens when the client approves the post. There's more on that in how to get client approval on social media posts.

Julexis can generate images and video inside the post editor, which is exactly why this checklist matters: making the asset is the easy part. Julexis doesn't label AI content automatically, so the caption line and each network's label stay on your list for every AI-made image or video you publish. The features page shows what the editor does.

Key takeaways

• Article 50 of the EU AI Act applies from 2 August 2026. The December 2026 grace period covers tool makers' machine-readable marking, not your labels.
• An agency that decides to use AI for a client is usually the deployer. Realistic AI people, products, places or events need a clear label at first sight, including in the feed.
• Ordinary AI ad copy isn't covered by the EU rule. Public-interest text is, unless a person genuinely reviews it and takes responsibility.
• Meta, TikTok, LinkedIn, Pinterest and X each have their own labels. Use them, add a caption line, and don't count on automatic detection.
• Agree the approach with each client in writing and keep a log of AI-made posts.

Cover image generated with AI.