EU publishes free AI labelling icons ahead of August 2026 deadline

The European Fee has revealed a set of free icons that publishers, advertisers and platforms can use to mark AI-generated or manipulated content material, a sensible step towards compliance with transparency duties underneath the EU AI Act that start making use of on 2 August 2026.

The icons, launched via the Fee’s Shaping Europe’s Digital Future portal with a final replace dated 10 June 2026, kind a part of Part 2 of the Code of Follow on marking and labelling of AI-generated content material. In accordance with the Fee’s revealed steerage, deployers of generative AI programs can use the icon set to label sure AI-generated content material in accordance with the AI Act’s transparency guidelines, and the icons help compliance with Article 50(4) of Regulation (EU) 2024/1689, which requires disclosure of deepfakes and of AI-generated textual content revealed on issues of public curiosity.

Three icons make up the set: a primary icon for basic disclosure, a “Totally AI-Generated” icon for content material created totally by AI with out human editorial management, and a “Partially AI-Modified” icon for pre-existing human-made content material altered with AI. Every is accessible in 4 variations masking black, white, and 50 p.c transparency variations, downloadable in SVG and PNG codecs. In accordance with the Fee, the icons underwent user-testing, and efficiency improved throughout all measures when the fundamental icon was accompanied by a textual content label corresponding to “modified.”

The publication follows months of parallel legislative exercise. The AI Act itself, formally Regulation (EU) 2024/1689, entered into pressure on 1 August 2024, with its transparency provisions underneath Article 50 among the many final main obligations to take impact underneath the unique timetable. These obligations require deployers of AI programs that generate deepfakes to reveal the factitious origin of that content material, and require comparable disclosure for AI-generated textual content revealed to tell the general public on issues of public curiosity, until a pure or authorized individual has exercised editorial management over the content material.

What the icons cowl, and what they don’t

Not every bit of AI-generated content material wants a label underneath the Fee’s framework. In accordance with the revealed steerage, the disclosure requirement covers solely two classes: deep fakes, outlined as AI-generated or manipulated picture, audio, or video content material that resembles current individuals, objects, locations, entities or occasions and would falsely seem to an individual to be genuine or truthful; and AI-generated or manipulated textual content revealed to tell the general public on issues of public curiosity, the place that textual content didn’t bear human assessment or editorial management and the place no authorized or pure individual assumed editorial duty for it.

The Fee’s steerage attracts an additional distinction for inventive and satirical content material. The place deep pretend materials varieties a part of an evidently inventive, inventive, satirical, fictional or analogous work, the transparency obligation narrows to disclosure “in an applicable method that doesn’t hamper the show or enjoyment of the work.” That carve-out doesn’t prolong to promoting. A separate Fowl & Fowl evaluation of the Fee’s draft Article 50 pointers, revealed 11 June 2026, states plainly that an “AI-manipulated video involving deep pretend simulation of people promoting a product in an AI-generated scene depicting the usage of the product by the simulated shoppers with the goal of persuading viewers to purchase the product” doesn’t qualify for the lighter artistic-content regime. The identical draft pointers exclude AI-generated imagery of celebrities implying involvement in actions they didn’t participate in, until that imagery serves a real fictional, satirical or analogous function.

Two additional exemptions apply. Deepfakes and public-interest textual content licensed by legislation for detecting, stopping, investigating or prosecuting felony offenses fall exterior scope. So does AI-generated textual content that underwent human assessment or editorial management, offered a named authorized or pure individual holds editorial duty for the publication.

The 200-token threshold and the marking-versus-labelling cut up

The icons tackle labelling duties underneath Article 50(4), which fall on deployers, the organizations that truly publish or distribute content material to the general public. A separate and technically distinct responsibility underneath Article 50(2) falls on suppliers, the businesses that construct the underlying generative AI programs, requiring them to mark outputs in machine-readable codecs so the AI origin of content material could be detected robotically.

In accordance with an in depth analysis revealed 22 June 2026 by legislation agency Fowl & Fowl, the finalized Code of Follow on Transparency of AI-Generated Content material dropped an possibility that had existed in earlier drafts: a “provenance certificates,” or signed manifest, that suppliers may as soon as use as an alternative choice to watermarking for textual content outputs. The ultimate textual content now requires imperceptible watermarking for any free-form textual content longer than 200 tokens, with solely very brief textual content falling under that carved out. Suppliers should additionally preserve no less than two machine-readable marking layers for many content material: digitally signed and time-stamped metadata plus an imperceptible watermark, in response to the Fowl & Fowl abstract of Dedication 1 underneath Part 1 of the Code.

That marking-and-detection regime for suppliers is distinct from the icon-based labelling regime for deployers described within the Fee’s newly revealed steerage, although the 2 work collectively. A generative AI system embeds a machine-readable mark; the group publishing the ensuing content material then applies a human-visible icon, no less than for deepfakes and qualifying public-interest textual content.

The Fee’s steerage units out the place the icon should seem, and the principles are stricter than a passing point out deep in a doc. In accordance with the revealed abstract, the icon should be clearly perceivable and distinguishable on the newest on the time of first publicity of a pure individual to the content material, should be positioned the place no intervening overlay parts exist, and should be straight embedded into the deep pretend or revealed textual content, aside from inventive works, until equal alternate options corresponding to a consumer interface overlay can be found. The icon should stay seen when content material is reshared or downloaded.

A separate practitioner guide by product supervisor Gianna Brachetti-Truskawa, revealed 2 July 2026, states the position requirement much more straight for textual content: the label “goes on the high, within the byline space or earlier than the headline,” and “a footer disclosure doesn’t fulfill ‘at first publicity.'” The identical information notes that the icon alone doesn’t set up compliance with Article 50(4); deployers stay chargeable for assembly the underlying authorized requirement no matter whether or not they use the Fee’s icon or an equal of their very own design.

The Fee’s personal steerage reinforces this level straight. The usage of the EU icons is non-compulsory, however the labelling necessities underneath Article 50 of the AI Act will not be, and the usage of the icons doesn’t by itself set up authorized compliance. Signatories of the Code of Follow on marking and labelling of AI-generated content material are required to duly implement the measures the Code comprises.

An editorial-responsibility exemption, with a legal responsibility commerce

Article 50(4) comprises a narrower path for publishers who conduct real editorial assessment of AI-assisted textual content: content material that underwent a strategy of human assessment or editorial management, the place a pure or authorized individual holds editorial duty for the publication, falls exterior the disclosure obligation totally.

That exemption comes with a catch, in response to the identical practitioner information by Brachetti-Truskawa. The exemption “removes the duty however transfers legal responsibility.” A choice-tree useful resource accompanying the information frames the selection starkly: claiming editorial duty means a company is “able to personal full legal responsibility if an error slips via,” as a result of “the AI hallucinated” stops being a legitimate protection as soon as that duty is formally claimed. The choice, making use of the disclosure label as an alternative, retains legal responsibility connected to the AI system itself and doesn’t require documentation of a assessment course of. The information recommends that organizations counting on the exemption doc, earlier than publication, who reviewed the content material, when, and why they had been certified to take action, describing this because the group’s protection if a dispute arises later.

The excellence between a basic editor and a reviewer with substantive, independently verifying experience within the particular content material area issues underneath this framework. In accordance with the information, proofreading for fashion, with out independently verifying whether or not a selected factual declare is correct, doesn’t meet the edge; the exemption requires substantive assessment able to catching factual errors and hallucinations in that exact topic space.

Analysis favors temporary labels over prolonged disclosures

One knowledge level within the practitioner information is more likely to affect how deployers truly phrase their labels. In accordance with analysis the information cites as revealed in 2026, lengthy disclosures, together with particular instrument names and descriptions of oversight workflows, carried out worse on reader belief than temporary labels did. That discovering aligns with the Fee’s personal icon design alternative: solely the capitalized acronym “AI” is necessary, with further element corresponding to “GENERATED” or “MODIFIED” merely inspired moderately than required, in response to the Fowl & Fowl evaluation of the finalized Code.

The Fowl & Fowl evaluation notes an additional purpose the Fee might have stored the deeper disclosure non-compulsory moderately than necessary. Requiring each deployer to reveal whether or not content material was “generated” versus merely “modified” may have an effect on a piece’s copyright protectability, for the reason that time period “generated” implies doubtlessly no human authorship, a standing copyright legislation doesn’t shield. Leaving that second-layer disclosure non-compulsory permits a deployer to explain content material as “modified” the place the info help it, a characterization extra favorable to copyright protectability and, in response to the Fee’s personal user-testing, higher understood by audiences in any case.

Signing window closes earlier than the icons take authorized impact

Organizations that signal the Code of Follow by 22 July 2026 achieve what the practitioner information describes as a presumption of compliance, shifting the main target of enforcement towards adherence to the Code’s particular measures moderately than towards the underlying authorized textual content of Article 50 itself. Signatories are listed publicly. The Code itself states, in response to the Fowl & Fowl evaluation of its personal textual content, that adherence “doesn’t represent conclusive proof of compliance,” that means that even signing doesn’t absolutely insulate a company from scrutiny, although it does shift the place that scrutiny concentrates.

For organizations that don’t signal, the Code’s affect stays oblique however actual. Fowl & Fowl’s evaluation states that market surveillance authorities might deal with the Code because the benchmark for what sufficient compliance appears to be like like, that means a non-signatory firm that adopts its personal various technical method carries the burden of demonstrating that its various is no less than as efficient, interoperable, strong, and dependable because the Code’s method.

A parallel deadline shift for high-risk programs doesn’t contact this obligation

Article 50 transparency duties are continuing on their authentic schedule whilst different components of the AI Act have shifted. In a provisional political settlement reached on 7 Might 2026, the Council of the EU and the European Parliament settled on new fastened dates for a separate class of obligations, these masking high-risk AI programs underneath Chapter III of the Act, corresponding to programs utilized in employment screening, credit score scoring, and biometric identification, shifting compliance for Annex III programs to 2 December 2027 and for Annex I programs to 2 August 2028. That settlement varieties a part of the Digital Omnibus package deal the European Fee launched on 19 November 2025 and doesn’t contact the Article 50(4) labelling responsibility or the two August 2026 date on which it takes impact. A separate associated change underneath that very same settlement provides suppliers of generative AI programs already available on the market earlier than 2 August 2026 till 2 December 2026 to convey their machine-readable marking underneath Article 50(2) into conformity, in response to reporting on the Digital Omnibus settlement.

Why this issues for promoting and advertising and marketing organizations

The icon publication lands squarely inside a compliance query that promoting businesses, in-house advertising and marketing groups, and publishers utilizing generative AI for marketing campaign property have had purpose to trace for the reason that Fee opened its Article 50 session in September 2025. The excellence the Fowl & Fowl evaluation attracts between promoting content material and genuinely inventive or satirical work removes any expectation {that a} persuasive product demonstration that includes AI-generated shopper avatars may depend on the lighter creative-content disclosure regime; unusual full labelling applies.

The editorial-responsibility exemption is more likely to matter most for organizations publishing AI-assisted commentary, market evaluation, or information content material on issues of public curiosity, classes the Fee’s steerage defines broadly sufficient to incorporate present affairs, public well being, environmental points, politics, shopper safety, financial coverage and science, in response to a decision-tree useful resource accompanying the practitioner information. Pure product advertising and marketing, in response to the identical useful resource, sits exterior the Article 50(4) public-interest scope totally, although different Article 50 obligations should still apply relying on how the content material is generated and disseminated.

For advertising and marketing organizations counting on a number of distributors, freelance contributors, or company companions to supply AI-assisted content material, the uniformity query the Fowl & Fowl evaluation raises is a sensible one: a number of events feeding one model might every apply their very own compliant however in another way designed label, leaving a single advertiser with a patchwork of inconsistent marks throughout its content material. Standardizing on the Fee’s EU icon moderately than a bespoke equal, and constructing contractual necessities across the Code’s placement guidelines, is a method organizations are addressing that danger, in response to the evaluation.

Timeline

• 13 June 2024 — The European Parliament and Council undertake Regulation (EU) 2024/1689, the EU AI Act.
• 1 August 2024 — The AI Act enters into pressure.
• 4 September 2025 — The European Fee opens a session on pointers and a Code of Follow for Article 50 transparency obligations.
• 8 Might 2026 — The Fee publishes its draft Article 50 pointers, opening a stakeholder session interval working till 3 June 2026.
• 7 Might 2026 — The Council and European Parliament attain a provisional settlement underneath the Digital Omnibus package deal, fixing new 2027 and 2028 deadlines for high-risk AI programs with out altering the Article 50(4) labelling deadline.
• 10 June 2026 — The European Fee publishes the free EU icons for labelling AI-generated content material, final up to date on this date.
• 11 June 2026 — Fowl & Fowl publishes an evaluation of the Fee’s draft Article 50 deep pretend pointers.
• 22 June 2026 — Fowl & Fowl publishes its evaluation of the finalized Transparency Code of Follow.
• 2 July 2026 — A practitioner decision-tree information on the Article 50(4) disclosure-versus-editorial-responsibility alternative is revealed.
• 22 July 2026 — Deadline for organizations to signal the Code of Follow to acquire a presumption of compliance.
• 2 August 2026 — Article 50 transparency obligations, together with the deepfake and AI-text labelling responsibility, turn into legally relevant.
• 2 December 2026 — Deadline for generative AI suppliers already available on the market earlier than 2 August 2026 to convey Article 50(2) machine-readable marking into conformity.
• 2 February 2027 — Deadline for suppliers to implement a watermark-detection interoperability answer.

Abstract

Who: The European Fee revealed the icons; they apply to deployers of generative AI programs, together with advertisers, advertising and marketing businesses, publishers and platforms working in or serving the European Union.

What: A free set of three icons, every with 4 visible variations in SVG and PNG format, that organizations can use to label deepfakes and AI-generated textual content underneath Article 50(4) of the EU AI Act. The icons kind a part of the Code of Follow on marking and labelling of AI-generated content material and had been revealed alongside detailed placement guidelines distinguishing embedded labels from overlay alternate options.

When: The icons carry a last-update date of 10 June 2026. The underlying authorized obligation they help, Article 50 of Regulation (EU) 2024/1689, turns into relevant on 2 August 2026. Organizations searching for a presumption of compliance via the Code of Follow should signal by 22 July 2026.

The place: The duty applies throughout the European Union and European Financial Space, and extends to organizations established exterior the EU if their AI-generated content material is made accessible to readers throughout the bloc.

Why: The labelling requirement addresses the chance that AI-generated or manipulated content material, significantly deepfakes and AI-written materials on issues of public curiosity, may deceive audiences about its genuine origin. Promoting content material is explicitly excluded from the lighter disclosure regime obtainable to genuinely inventive or satirical work, that means advertising and marketing organizations utilizing AI-generated shopper avatars or artificial spokespeople in persuasive campaigns face the total labelling obligation. Non-compliance with Article 50 falls underneath the AI Act’s basic penalty construction, which allows fines of as much as 15 million euros or 3 p.c of worldwide annual turnover, whichever is greater, for violations apart from the extra severely penalized prohibited practices underneath Article 5.