EU AI Act for Media and Publishing

Industry Guide20 July 2026· 16 min read

Most editorial AI is not high-risk under the EU AI Act. The real exposure is Article 50 disclosure, deepfake marking, Article 5 limits and GPAI copyright.

Not high-risk. A newsroom running automated summaries, headline testing, a personalised feed, synthetic explainer graphics and a reader chatbot is not deploying high-risk AI under Regulation (EU) 2024/1689 — the EU AI Act — because none of those uses appears in Annex III, and Article 6(2) is what makes a stand-alone system high-risk. So the conformity stack under Articles 8–15, 16 and 43 does not attach to ordinary editorial work. The publisher's real exposure is concentrated in four narrow surfaces: Article 50 transparency, the Article 5 prohibitions, GPAI duties when you build on a foundation model, and Article 4 AI literacy.

This guide classifies each editorial system, names the one realistic high-risk path, and handles the duties that genuinely apply rather than over-engineering a programme the estate does not need. Start with how limited-risk classification works, because that is where the publishing stack sits.

Most Media and Publishing AI Is Not High-Risk — Lead With That

High-risk status is set by Annex III via Article 6(2), not by how visible, expensive or editorially important a system feels. A generative headline tool that touches every article on the site is no more high-risk than a spell-checker if it does not land in an Annex III use case. The expensive mistake is treating reach as risk.

Why Annex III Decides It — Article 6(2)

Content generation, automated journalism, article summarisation, headline testing, news recommendation, ad targeting and synthetic image or video production for editorial use are not, by themselves, high-risk AI systems — they appear on no Annex III list. Annex III's eight areas are biometrics (point 1), critical infrastructure (point 2), education (point 3), employment (point 4), essential private and public services (point 5), law enforcement (point 6), migration and border control (point 7), and the administration of justice and democratic processes (point 8). Routine publishing sits in none of them.

The One Realistic High-Risk Path

A newsroom is only pulled into the high-risk regime through a side door it opens as an employer, not as a publisher. The clearest example is an AI hiring or CV-screening tool: AI used for recruitment and selection is high-risk under Annex III point 4(a), and a media business that buys or builds one becomes a deployer of a high-risk system. That obligation lives entirely in the HR domain. Dragging editorial systems into the same programme is how publishers waste a year of compliance effort.

Setting Honest Expectations

The real EU AI Act workload for publishers is Article 50 transparency, the Article 5 prohibitions, GPAI duties under Articles 51–55, and Article 4 AI literacy — not the high-risk conformity stack under Articles 8–15, 16 and 43. Map each surface to its owner and the programme becomes a handful of disclosure and literacy duties, not a documentation marathon.

The Four AI Surfaces a Newsroom Actually Has to Watch

Four exposures, and only four, run across a normal publishing estate. Name them so editorial, product, legal and commercial leads each know which one is theirs.

Article 50 Transparency — the Load-Bearing Duty

Article 50 is the obligation that touches a newsroom again and again: disclosure that a reader is interacting with an AI system, and marking of AI-generated or manipulated text, image, audio and video. It is light per system but pervasive across the estate, and it carries a fixed content-marking date to plan against.

Article 5 Manipulation and the New Deepfake-Adjacent Bans

Article 5 sets the outer limit. Manipulative or deceptive techniques and synthetic content used to materially distort behaviour and cause significant harm are prohibited, not merely regulated. For most publishing this line is never crossed; the point is to know where it sits.

GPAI Duties When You Build on a Foundation Model — Articles 51–55

GPAI obligations under Articles 51–55 have applied since 2 August 2025 and bite on whoever places a general-purpose model on the market. A publisher using a third-party model is normally a downstream deployer, not a GPAI provider — but the role analysis still has to be run.

Article 4 AI Literacy Across Editorial and Product

Article 4 AI literacy, in force since 2 February 2025, requires that staff operating these systems have a sufficient understanding of them — a light but real, organisation-wide duty. Not a certification scheme, but evidence that the people running the tools know what they do.

Article 50 Transparency: Disclosing AI Content and Marking Synthetic Media

This is the load-bearing obligation for publishers, read here through the editorial workflow — wire-photo manipulation, AI-written copy, synthetic anchors — rather than the abstract rule the Article 50 transparency obligations pillar covers in general terms.

Chatbots and AI Interaction Must Be Disclosed — Article 50(1)

Article 50(1) requires that AI systems intended to interact directly with people — a reader-facing chatbot or a conversational news assistant — inform users they are dealing with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person from the context. A short in-widget label at the start of the conversation is the safe pattern; no technical-documentation or conformity duty attaches.

Deepfakes Disclosed as Artificially Generated — Article 50(4)

Where a publisher generates or manipulates image, audio or video that constitutes a deepfake, Article 50(4) requires it to disclose that the content has been artificially generated or manipulated. For a newsroom this reaches synthetic news anchors, AI-altered photography and reconstructed footage — the cases where a reader could mistake the synthetic for the real.

Generative-AI Providers Mark Outputs Machine-Readably — Article 50(2)

Article 50(2) requires providers of generative AI systems to mark their outputs in a machine-readable format detectable as artificially generated or manipulated. This is a provider-side duty: it matters when a publisher builds its own generative tooling rather than consuming a third party's.

The Content-Marking Date and the Editorial Carve-Out

Article 50 content-marking duties apply from 2 December 2026 under the Digital Omnibus package — now adopted. Until adoption and Official Journal publication, the operative reference is the Article 50 statute as enacted; plan now, do not assume slippage. Article 50(4) carries an editorial and artistic carve-out: where AI-generated or manipulated content forms part of an evidently artistic, creative, satirical or fictional work, the disclosure is limited so as not to hamper the work's display or enjoyment. This is not a licence to deceive: a synthetic clip presented as real footage falls outside the carve-out entirely.

Article 5: The Manipulation Line and the New Deepfake-Adjacent Bans

If Article 50 is the recurring duty, Article 5 is the rare red flag — the only place in this guide that touches the top penalty tier. These are prohibitions, in force since 2 February 2025.

Subliminal, Manipulative and Deceptive Techniques — Article 5(1)(a)

Article 5(1)(a) prohibits AI that deploys subliminal techniques beyond a person's consciousness, or purposefully manipulative or deceptive techniques, that materially distort behaviour and cause — or are reasonably likely to cause — significant harm. For publishing this bites where synthetic content or an engagement engine crosses from persuasion into engineered deception: a covert synthetic persona, an undisclosed deepfake designed to mislead.

Exploiting Vulnerability — Article 5(1)(b)

Article 5(1)(b) prohibits AI that exploits vulnerabilities of age, disability or a specific social or economic situation to materially distort behaviour and cause significant harm. For a publisher this is a real limit on AI-driven targeting of vulnerable audiences — a recommendation system tuned to push distressing or exploitative content at a cohort it has identified as susceptible.

Where the Risk Concentrates

Editorial illustration, clearly-labelled satire and standard recommendation are persuasion, not prohibited manipulation. The threshold is material behavioural distortion plus significant harm — both limbs, not one. Covert synthetic personas and undisclosed deepfakes designed to deceive are where the risk concentrates. The Digital Omnibus package also adds child-protection prohibitions — CSAM and "nudifier"-type tooling — as part of the same Article 5 family. Breaches sit in the top tier under Article 99(3): up to EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher. See the Article 5 prohibitions already in force.

Building on Foundation Models: GPAI, Roles, and When a Publisher Becomes a Provider

When an editorial team builds generative tooling on a general-purpose model, the GPAI obligations in Articles 51–55 come into view — the question is who carries them.

Downstream Deployer Versus GPAI Provider

A publisher fine-tuning or wrapping a third-party general-purpose model for headline generation or summarisation is usually a downstream deployer. The GPAI obligations in Articles 51–55 — including the Article 53 duties to maintain a copyright-compliance policy and publish a sufficiently detailed summary of training content — fall on the model provider, in force since 2 August 2025. Accessing a model through an API does not inherit those duties.

The Triggers That Flip the Role — Article 25(1)

That split is not permanent. Under Article 25(1), a deployer, distributor or importer becomes a provider where it puts its own name or trademark on a high-risk system, makes a substantial modification, or changes the intended purpose so the system becomes high-risk. For media use the trigger rarely produces a high-risk system, because the underlying editorial use is not Annex III — but run and document the role-shift analysis whenever you move from consuming a model to customising it.

Keeping GPAI Scope Honest

GPAI provider compliance is a partial, evolving area, not a box a deployer can simply tick — at Confir it stays on the roadmap rather than being marketed as complete. Copyright and training data are a live, adjacent question: the Article 53 copyright-policy and training-summary duties shape vendor due diligence even though the deployer is not the provider. The substantive copyright position for training and outputs is set by EU copyright law, which sits alongside, not inside, the AI Act.

Worked Example: How a Mid-Sized European Publisher Classifies Its AI Stack

Consider Meridian Media, an illustrative regional digital news publisher with roughly 350 staff and 8 million monthly readers, running five distinct AI surfaces. Here is how each lands.

Walking Each System to a Tier

System 1 — an LLM-based article-summary and headline tool built on a third-party foundation model. Not Annex III, so not high-risk; Meridian is a deployer; Article 50 marking applies to any reader-visible generated text. System 2 — a personalised news-feed recommendation engine. Not high-risk, no Annex III hook; the Article 5 manipulation limits set its outer boundary. System 3 — synthetic image generation for explainer graphics. Not high-risk, but Article 50(4) deepfake disclosure applies, subject to the artistic carve-out. System 4 — a reader-facing support chatbot. Limited-risk; Article 50(1) disclosure of AI interaction. System 5 — an AI CV-screening tool used by HR. The only high-risk system: Annex III point 4(a), pulling in full deployer obligations separate from editorial work.

Running the Article 25 Role Analysis

Meridian is a deployer of a third-party foundation model accessed via API for Systems 1 and 3 — not a GPAI provider. That holds unless it fine-tunes and rebrands the model as its own, triggering an Article 25(1) reassessment. Even then, because the editorial use is not Annex III, the flip rarely produces a high-risk system.

The Practical Compliance Outcome

Four of the five systems are limited- or minimal-risk, carrying Article 50 and Article 4 literacy duties. One — the HR CV-screening tool — is an isolated high-risk system handled by HR under Annex III point 4(a). Meridian's total editorial AI Act exposure is transparency and literacy work, not a full conformity programme.

The Media and Publishing Use-Case-to-Risk Map

The table collapses the guide into one view: find your system, read its likely tier, then the obligation and Article that follow.

Publishing AI use	Likely tier	Key obligation and Article
Automated journalism / summaries	Limited	Article 50 marking of generated text
News recommendation	Minimal–limited	Article 5 outer limit; no high-risk duty
Synthetic editorial image / video	Limited	Article 50(4) disclosure + artistic carve-out
Reader-facing chatbot	Limited	Article 50(1) AI-interaction disclosure
Ad targeting	Minimal–limited	Article 5 limit + GDPR profiling
Covert synthetic persona to deceive	Prohibited	Article 5(1)(a); top tier Article 99(3)
AI CV-screening for hiring	High-risk	Annex III point 4(a) deployer obligations

The only prohibited row is the covert-deception one, the only high-risk row is the HR tool, and the only recurring AI Act duty across the editorial estate is Article 50 transparency.

The Penalty Tiers, Inline

Article 5 prohibited-practice breaches sit at the top: up to EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher, under Article 99(3). Most other obligation breaches — including Article 50 transparency duties — are up to EUR 15,000,000 or 3% under Article 99(4). Supplying incorrect, incomplete or misleading information to authorities is up to EUR 7,500,000 or 1% under Article 99(5). For SMEs and start-ups, Article 99(6) caps the fine at the lower of the two.

Deadlines, the Digital Omnibus Caveat, and How Confir Helps

The Dates That Bind Media and Publishing

Several already apply. Article 5 prohibitions and Article 4 AI literacy have been in force since 2 February 2025; GPAI duties under Articles 51–55 since 2 August 2025. The Article 50 content-marking duties apply from 2 December 2026. For context only — since editorial use is not high-risk — the stand-alone high-risk Annex III statute date under Article 6(2) is 2 August 2026, with a deferral to 2 December 2027 that is now adopted, and the Annex I product-embedded route under Article 6(1) applies from 2 August 2027.

The Omnibus Caveat in One Paragraph

The Digital Omnibus was adopted in June 2026: the European Parliament passed it in plenary on 16 June 2026 and the Council formally adopted it on 29 June 2026, with entry into force on Official Journal publication. The high-risk dates are settled — plan against the 2 December 2027 stand-alone Annex III date (2 August 2028 for product-embedded Annex I systems) and the 2 December 2026 content-marking date. A standards-contingent "stop the clock" was rejected, and not everything is delayed — treat the fixed calendar dates as live.

Where Confir's Rule-Based Engine Fits

Confir maps each AI surface to its tier and obligation from a plain-English intake, separating the limited- and minimal-risk editorial tooling — summaries, recommendation, synthetic graphics, chatbots — from any Article 5 red flag and from the isolated high-risk HR system. It produces an auditable classification, an Article 50 disclosure register, and an Article 4 AI-literacy and role-evidence trail.

The engine is deterministic and rule-based: no model inference, no hallucination. The same intake answers produce the same finding, with the rule that fired shown, so every classification is reproducible and audit-defensible. GPAI provider compliance remains a partial, roadmap area.

Frequently asked questions

Is AI used in journalism or content generation high-risk under the EU AI Act?

No. High-risk status comes from Annex III via Article 6(2), and ordinary editorial uses — automated summaries, headline testing, content generation, news recommendation — do not appear on any Annex III list. They are not high-risk AI systems. The duties that do apply are Article 50 transparency, the Article 5 manipulation limits, and Article 4 AI literacy. A media company only touches the high-risk regime through a side door, such as deploying an AI CV-screening tool as an employer under Annex III point 4(a).

Do publishers have to label AI-generated articles and synthetic images?

Increasingly, yes. Under Article 50(4), deployers who generate or manipulate image, audio or video constituting a deepfake must disclose it as artificially generated or manipulated. Under Article 50(2), providers of generative systems must mark outputs in a machine-readable format. These content-marking duties apply from 2 December 2026 under the Digital Omnibus, which is adopted. Article 50(4) keeps a carve-out for evidently artistic, satirical or fictional works so disclosure does not hamper the work's display.

What is the artistic and satirical carve-out in Article 50(4)?

Where AI-generated or manipulated content forms part of an evidently artistic, creative, satirical or fictional work, Article 50(4) limits the disclosure obligation so it does not hamper the display or enjoyment of that work. This protects editorial illustration, satire and creative formats. It is not a licence to deceive: a synthetic news clip presented as real footage falls outside the carve-out, and covert deepfakes designed to mislead can cross into prohibited manipulation under Article 5(1)(a).

Is a media company a GPAI provider if it uses a large language model?

Usually not. A publisher fine-tuning or wrapping a third-party foundation model is normally a downstream deployer; the GPAI obligations in Articles 51–55 sit with the model provider. The role only flips under Article 25(1) if the publisher puts its own name or trademark on a high-risk system, substantially modifies it, or changes its intended purpose so it becomes high-risk. Because editorial use is not Annex III, that flip rarely produces a high-risk system — but the role analysis still has to be run and documented.

What about copyright and training data for AI used in publishing?

Copyright is a live, adjacent question. GPAI model providers face Article 53 duties to put in place a copyright-compliance policy and publish a sufficiently detailed summary of training content. As a deployer, a publisher is not the provider, but those duties shape vendor due diligence and contracting. Note that the EU AI Act governs transparency and provider duties; the substantive copyright position for training and outputs is set by EU copyright law, which sits alongside, not inside, the AI Act.

What are the penalties if a publisher gets Article 50 disclosure wrong?

Most Article 50 breaches fall in the second tier: up to EUR 15,000,000 or 3% of worldwide annual turnover under Article 99(4). Using prohibited manipulative or deceptive AI under Article 5 is the top tier — up to EUR 35,000,000 or 7% under Article 99(3). Supplying incorrect or misleading information to authorities is capped at EUR 7,500,000 or 1% under Article 99(5). For SMEs and start-ups, fines are capped at the lower of the percentage or the fixed figure under Article 99(6).

When do the relevant deadlines hit for media and publishing?

Article 5 prohibitions and Article 4 AI literacy have applied since 2 February 2025. GPAI obligations under Articles 51–55 have applied since 2 August 2025. Article 50 content-marking duties apply from 2 December 2026 under the Digital Omnibus, which reached provisional agreement on 6–7 May 2026 and was adopted: the European Parliament passed it in plenary on 16 June 2026 and the Council formally adopted it on 29 June 2026. Only publication in the Official Journal remains — expected before 2 August 2026 — a formality that does not affect the dates. Until then the enacted statute governs, so plan against the stated dates — a stop-the-clock was rejected and not everything is delayed.

Manage your EU AI Act compliance in one place

Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.

Start free trial →