oddlynew
Sign inStart for free
LibrarySource: Mentioning AI, disclosure evidence and the EU legal floorv7 · Updated Report

Source: Mentioning AI, disclosure evidence and the EU legal floor

Status: source note, research pass on 29 September 2026 by a research agent. Studies and legal texts were read on that date unless a source gives a later date, and section 2 carries its own "Last checked" date; legal texts were read from the Official Journal and Commission publications, the Article 50 guidelines from a mirrored copy, rechecked against the Commission's own copy on 29 October 2026. Evidence classes: M meta-analysis, F field, X controlled experiments, S survey, G design guidance, L binding law, LS official soft law, I inference. Legal content is orientation, not legal advice; dates in this area move, so verify before relying on them and take the open questions to counsel.

0. Bottom line

  1. Since 2 Aug 2026 the law requires AI disclosure in four places that matter for copy. First, any chat a person can mistake for a human (Art. 50(1)). Second, machine-readable marking by whoever provides a generator of synthetic text, images, audio or video (Art. 50(2)); a generator on the market before 2 Aug 2026 has until 2 Dec 2026. Third, AI image, audio or video that resembles real persons, objects, places or events and would falsely appear authentic (deep fakes, Art. 50(4), first subparagraph); an AI product or food photo that misleads about the real product can count. Fourth, AI text published to inform the public on matters of public interest, unless it went through human review or editorial control and a person or company holds editorial responsibility for publishing it (Art. 50(4), second subparagraph). None of these requires AI in a headline or a product name.

  2. Disclosing AI use costs the person who discloses trust (Schilke and Reimann, pooled d = 0.81), and no framing they tested removed it, including "a human reviewed and revised it" and "proofreading only". The penalty on the content itself is different: it disappears when a label says AI only drafted or edited (Altay and Gilardi), and AI editing hurts less than AI authorship (Kirk and Givi). In one study, the task-fit penalty on people who use AI was offset for a task where AI was described as useful (Reif, Larrick, Soll), and giving users control raises use (Dietvorst 2018). Being exposed by a third party is worse than disclosing yourself. So disclose where the law requires it or where discovery is likely, and say it as a plain statement of roles, never as a selling point.

  3. AI is proof only when the reader already wants the mechanism, for example an agent-literate audience for whom their own agents are the subject. It is a liability in the headline for any high-risk, personal or emotional context.

  4. Control and optionality beat avoidance. Where attitudes toward AI vary within an audience, the winning frame is letting the user decide how much AI help to use, not hiding AI. That matches the strongest experimental lever in the literature, which is giving the user control.

1. Evidence digest

StudyDesign, sample, domainFinding that mattersClass, strength
Dietvorst, Simmons, Massey 2015, JEP: General 144(1)5 incentivized forecasting studies (MBA performance, airline passenger ranks)After seeing an algorithm err, people abandon it for a worse human, even when it clearly outperformed the human.X, strong for forecasting
Dietvorst, Simmons, Massey 2018, Management Science 64(3)3 incentivized forecasting studiesUse of an imperfect algorithm rises sharply when people may modify its output even slightly. Satisfaction and perceived accuracy go up.X, strong; directly supports giving users control
Logg, Minson, Moore 2019, OBHDP 1516 experiments, lay estimation and forecastingLaypeople weight algorithmic advice above human advice. The effect fades against their own judgment and among experienced professionals.X, moderate
Castelo, Bos, Lehmann 2019, JMR 56(5)4 lab studies (1,400+) and 2 field studies (56,000+)Less trust in algorithms for tasks that seem subjective. Framing a task as more objective raises use.X+F, strong
Longoni, Cian 2022, J. Marketing7+ studiesAI recommenders are preferred for utilitarian goals and resisted for hedonic ones. Resistance returns when personal matching is needed. It disappears when a human and AI decide together.X, strong; recipes are partly hedonic
Qin et al. 2025, Psychological Bulletin 151(5)Meta-analysis, 163 studies, 442 effects, N=82,078Appreciation only when AI is seen as more capable and personalization is unnecessary (d=0.27). Otherwise aversion (d=-0.50). Aversion is stronger in high-education, high-internet countries.M, strong
Schilke, Reimann 2025, OBHDP 18813 preregistered experiments; students plus CloudResearch panel; grading, job applications, ads, reviews, emails, investment fundsPeople who disclose AI use are trusted less (d from 0.36 to 1.42 against no disclosure, pooled d = 0.81), through lower legitimacy. The effect holds across framings, whether disclosure is voluntary or mandatory, and when AI use was already known. Exposure by a third party hurts more. Pro-tech evaluators show less penalty, but still some.X, strong
Reif, Larrick, Soll 2025, PNAS 122(19)4 preregistered experiments, N=4,439, workplace and hiringAI users are judged lazier and less competent, and managers who do not use AI hire them less. In Study 4 the task-fit penalty was entirely offset for a task where AI was described as useful, though AI users were still seen as lazier.X, strong
Altay, Gilardi 2024, PNAS Nexus 3(10)2 preregistered experiments, US and UK, N=4,976, news headlinesAn "AI-generated" label lowers perceived accuracy and willingness to share by 2.66 percentage points in Study 1, mostly through accuracy. That is a third of the effect of a "false" label, because readers assume full automation. When the label says AI only drafted or edited, the penalty disappears.X, strong
Kirk, Givi 2025, J. Business Research 1867 preregistered experiments, marketing messagesAI authorship of emotional messages cuts positive word of mouth and loyalty, through lower authenticity and moral disgust. The effect is smaller for factual messages and when AI only edits.X, strong
Cicek, Gursoy, Lu 2024, JHMM 34(1)6 experiments, 1,000+ US adults, 8 product and service categoriesAdding "artificial intelligence" to a description lowers purchase intent through emotional trust. The drop is larger for high-risk products.X, moderate (online scenarios)
Luo et al. 2019, Marketing Science 38(6)Field experiment, 6,200+ customers, outbound sales callsUndisclosed chatbots perform like skilled staff. Disclosure before the conversation cuts purchases by more than 79.7%, which later disclosure mitigates.F, strong, but see "Do not rely on"
Crolic et al. 2022, J. Marketing 86(1)Telecom chat data plus 4 experimentsHuman-like chatbots (names, avatars) lower satisfaction, firm evaluation and purchase intent for angry customers, because they inflate expectations.F+X, strong
Jakesch et al. 2019, CHI3 experiments, Airbnb host profiles"Replicant effect": all-AI or all-human sets are trusted equally. In a mixed set, profiles labeled or suspected as AI are trusted less.X, moderate; relevant to any site mixing AI and human pieces
Langer et al. 2022, CHI2 studies (N=397 and N=622), algorithmic decision systemsThe word used ("AI", "algorithm", "computer program") alone shifts perceived complexity and trust.X, moderate
Epstein et al. 20235 countries, about 5,100 people"AI generated" reads as a description of process, not of deception. "Manipulated" and "deepfake" read as misleading.S, moderate
Califano, Zhang, Spence 2024, Int. J. Gastronomy & Food Science 37Online experiment, recipesNo trust gap for standard dishes. Lower trust in AI for innovative recipes.X, moderate
Kraus et al. 2025, Dietetics 4:50Students (n=948), randomized blog evaluationAI and student blogs rated alike. 42% say they would use a blog less if they knew it was AI-generated. Participants want a sign that recipes were tested or reviewed by humans.X+S, weak (convenience sample)
EMAC 2025 proceedings (food ad)n=220Disclosing a GenAI ad image lowered perceived brand effort, then taste expectation, then purchase intent.X, weak
Califano, Spence 2024, Food Quality & Preference 1162 online experimentsUnlabeled, AI food images were often preferred. Saying an image is real raised its appeal; saying it is AI-generated weakened that effect.X, moderate
Microsoft HAX guidelines (Amershi et al. 2019, CHI)18 guidelines, validated with 49 practitionersG1 make clear what it can do; G2 how well; G9 easy correction; G11 why it acted; G17 global controls; G18 announce changes.G, moderate
Google PAIR GuidebookMental Models and Explainability + Trust chapters"Focus on the benefit, not the technology." Avoid "AI magic" messaging. Make the algorithmic nature clear to prevent expectations of human-like ability.G, moderate

Gap: no peer-reviewed German-language study on the word "KI" in B2B or social-sector copy was found. Claims about German skepticism toward the word rest on indirect evidence such as Qin et al.'s country moderators (M, indirect).

2. Legal floor, checked against the Official Journal

Last checked: 27 October 2026, against the AI Act Service Desk timeline (https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act), the Council's press release of 21 October 2026 on the data protection omnibus, and § 5 and § 5a UWG as amended by BGBl. 2026 I Nr. 43 (https://www.recht.bund.de/bgbl/1/2026/43/VO.html, read 29 October 2026). Held: Art. 50 applies from 2 Aug 2026, the Art. 50(2) grace for systems already on the market runs to 2 Dec 2026, Annex III applies from 2 Dec 2027, and the cited UWG paragraphs are unchanged in substance. The monthly Loop: Recheck the legal blocks keeps this line current.

AI Act, Regulation (EU) 2024/1689 (OJ L, 12.7.2024).

  • Art. 113: applies from 2 Aug 2026. Chapters I and II (including Art. 4) have applied since 2 Feb 2025.

  • Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026 (OJ L, 24.7.2026), in force 27 July 2026. It moves the Annex III high-risk obligations (Chapter III Sections 1 to 3) to 2 Dec 2027 and Annex I systems to 2 Aug 2028. Generators placed on the market before 2 Aug 2026 get until 2 Dec 2026 for Art. 50(2). Art. 4 AI literacy is softened to "take measures to support the development of AI literacy". Art. 50(7) is replaced, so the Commission now assesses codes as adequate. Art. 50(1), (4) and (5) are unchanged and apply from 2 Aug 2026.

  • Art. 50(1): providers must design systems that interact directly with people so the person is told it is an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person.

  • Art. 50(2): providers of systems that generate synthetic text, images, audio or video must mark outputs in a machine-readable, detectable way. This does not apply where the system only assists standard editing or does not substantially alter the deployer's input or its meaning.

  • Art. 50(4), first subparagraph: deployers of AI that generates or manipulates image, audio or video content constituting a deep fake must disclose that the content is artificially generated or manipulated (https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50, read 29 October 2026). A deep fake resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful (Art. 3(60)). In evidently artistic, creative, satirical or fictional work, disclosing it in a way that does not hamper the work is enough. The Commission's Article 50 guidelines list an AI product image in advertising that misleads about the real product's appearance as a deep fake.

  • Art. 50(4), second subparagraph: deployers of AI that generates or manipulates text "published with the purpose of informing the public on matters of public interest" must disclose it. The duty lapses if the text had human review or editorial control and a person holds editorial responsibility.

  • Art. 50(5): the disclosure must be "clear and distinguishable", given at the latest at first interaction or exposure, and accessible.

  • Penalty: Art. 99(4)(g), up to EUR 15m or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups (Art. 99(6)) and small mid-caps (Art. 99(6a), added by Regulation (EU) 2026/1744) the lower of the two applies.

  • Annex III 4(a): systems for recruitment or selection, including systems that "evaluate candidates", are high-risk.

  • Art. 6(3): exemptions exist for narrow procedural or preparatory tasks, but a system that profiles natural persons is always high-risk.

  • Art. 26(11): deployers of Annex III systems must tell affected people that such a system is used. Because it sits in Chapter III Section 3, it applies from 2 Dec 2027.

Commission Guidelines on Article 50, C(2026) 5054, approved 20.7.2026 (Commission text, https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems, read 29 October 2026; non-binding).

  • Chat disclosure should be a first-turn statement or a plain-language banner, ideally near the input field. Wording should be simpler for users with low digital literacy.

  • Not sufficient on their own: disclosure only in terms and conditions; generic words like "assistant"; "human-like representations that may mislead users"; site-wide lines such as "Services on this website use AI"; technical descriptions like "uses LLMs".

  • Periodic reminders are likely needed for vulnerable users, explicitly including "those with lower literacy". The system must also disclose whenever it is asked about its nature.

  • The obviousness exception is read restrictively and cannot be relied on where the public or vulnerable people are in the audience.

  • Published text: "public interest" covers topics such as public health and consumer safety. The in-scope examples include a "lifestyle-website article comparing the effects of various diets on a particular disease". The out-of-scope examples include product descriptions without health, safety or sustainability claims, and fantasy novels. Private correspondence is not "published".

  • The human-review exception needs substance: fact-checking at minimum, by someone with relevant knowledge. Spell-checks and cursory approval do not count. Substantive AI edits after sign-off void the exception. The person holding editorial responsibility should be publicly findable, for example in the Impressum.

  • A company that builds a system in-house and puts it into service under its own name is a provider.

Code of Practice on marking and labelling AI-generated content

  • Final version published 10 June 2026. The Commission and the AI Board confirmed it as adequate.

  • Deployers label with the EU "AI" icon or an equivalent, placed near the headline or at the top of published text. A second layer saying "generated" or "modified" is encouraged.

  • Providers should use at least two machine-readable marking layers where needed.

  • The icons are optional. The labelling duty is not.

GDPR (OJ L 119, 4.5.2016).

  • Art. 12(1): information must be "concise, transparent, intelligible and easily accessible", "using clear and plain language".

  • Art. 13(1)(e): the notice names "the recipients or categories of recipients of the personal data, if any". Naming the AI vendors that act as processors is the clearer choice and answers "Which AI vendor sees it?" directly.

  • Arts. 13(2)(f) and 14(2)(g): where Art. 22 automated decision-making exists, people must be told, with "meaningful information about the logic involved" and its consequences.

  • Art. 22: a right not to be subject to solely automated decisions with legal or similarly significant effects.

  • CJEU C-634/21 SCHUFA (7 Dec 2023): an automated score is itself an Art. 22 decision when a third party "draws strongly" on it.

  • CJEU C-203/22 Dun & Bradstreet Austria (27 Feb 2025): "logic involved" means an intelligible explanation of the procedure and principles actually applied. A formula is not enough.

  • The GDPR part of the Digital Omnibus is not law. The Council agreed its negotiating mandate on 21 October 2026, which keeps the Art. 22 right not to be subject to solely automated decisions; the Parliament has not adopted its position, so negotiations between the two have not started (https://www.consilium.europa.eu/en/press/press-releases/2026/10/21/digital-rulebook-council-agrees-position-to-simplify-data-protection-privacy-and-cyber-reporting-rules/, read 27 October 2026; the Art. 22 wording is in the mandate, Council document ST 14756/26, https://data.consilium.europa.eu/doc/document/ST-14756-2026-INIT/en/pdf, read 29 October 2026).

UWG (German unfair competition law).

  • §5(1) and (2) Nr. 1: false or deceptive statements about the method of production, expected results or fitness for purpose are unfair. This covers overstating AI ("KI erstellt perfekte Lebensläufe") and also understating it: "tested in our kitchen" or "written by our editors" when an agent wrote it.

  • §5a(1) and (2): withholding material information is unfair, including giving it in an unclear or late way.

How to apply

  • A chat surface you offer under your own name makes you the Art. 50(1) provider, even when a customer deploys it. With a vulnerable or low-literacy audience, do not rely on the obviousness exception.

  • A generator that substantially writes documents from a user's notes may be an Art. 50(2) provider with a machine-readable marking duty. The 2 Dec 2026 grace applies only if it was on the market before 2 Aug 2026.

  • Articles on public-interest topics (health, diet and disease, food safety, consumer safety, AI policy) need substantive human review with a named editor, or an "AI" label at the headline. Plain recipes and product descriptions without such claims are likely outside Art. 50(4).

  • Private correspondence such as a CV or a cover letter is not published text, so Art. 50(4) does not apply and no visible label is legally required.

  • Recruitment and candidate evaluation can be Annex III high-risk (profiling), with the Art. 26(11) notice from 2 Dec 2027. If a third party draws strongly on a score, Art. 22 applies. Keeping a named human assessment as the decision point is the practical mitigation.

  • A product with no built-in model that connects customers' own agents is not the generator provider; the customers using their agents are the deployers. Keep that true in the architecture if you claim it.

  • Search engine guidance adds a practical layer: give readers context on how content was made, use bylines where readers would ask "who wrote this?", and avoid mass-produced pages without added value.

3. Mentioning AI: principles

A. Required disclosure (the law sets the floor; copy sets the tone)

  1. Every chat surface says it is AI at the first turn, in plain words, near the input field. Where the audience is vulnerable, a persistent badge stays visible and reminders recur. The chat answers truthfully when asked. L+LS, strong.

  2. Chat bots get no human name, human face or "assistant" persona. That is both a legal insufficiency and an expectation trap. LS strong; F+X strong (Crolic); G (PAIR).

  3. Say what the AI did and what the human did: "KI-Entwurf aus Ihren Notizen, von Ihnen geprüft." A bare "AI-generated" makes readers assume full automation. X strong (Altay; Kirk and Givi, editing versus authoring).

  4. Published public-interest text gets substantive human review with a named owner, or a clear AI label at the headline, for example the EU "AI" icon. A cosmetic check does not count. L+LS, strong.

  5. Disclose before someone else reveals it. Being exposed costs more trust than disclosing. X strong (Schilke and Reimann, Study 13).

B. AI as proof (only where the reader wants the mechanism)

  1. For an audience that is organizing its own use of AI, AI is the subject, not the adjective:

    • Name the reader's agents and what they do ("your agents", "the agent you choose"). Do not present the product as the intelligence.

    • Show the mechanism and a verifiable demonstration (one agent starts a task, another continues it), never "AI-powered" or "intelligent".

    • X moderate (Cicek, Langer); M moderate (aversion higher in high-education countries); I.

  2. Credit AI by function when the benefit is clearly a task where AI helps: searching sources, drafting from notes, removing duplicate entry. In one study, Reif et al. found the task-fit penalty for AI users "entirely offset" for a task where AI was described as useful. X, moderate.

  3. Pair any capability line with its limit and a way to correct it (HAX G1, G2, G9, G11). Put each limitation next to the claim it limits. G moderate.

C. AI as liability (keep it out of headlines)

  1. Never lead with AI in high-risk, personal or emotional contexts: hiring and job decisions, CVs, social welfare, emotional brand messages. M strong (personalization needed, d=-0.50); X strong (Cicek high-risk; Kirk and Givi emotional).

  2. Where attitudes toward AI vary within the audience, frame AI as optional help the user controls: "Sie entscheiden, wie viel Hilfe es gibt." It works for users who want less help and for those who want more. X strong (Dietvorst 2018 control; Longoni and Cian hybrid).

  3. Avoid terms that imply the machine decides or authors: "KI-Matching", "Algorithmus", "automatisch generiert", "Unsere KI", "KI-gestützt", "AI-native", "intelligent". X moderate (Langer); I.

  4. Content-type example, recipes: no AI badge in marketing. Disclose provenance site-wide and per article, as a statement of roles. A mixed site with unexplained AI pieces invites the replicant effect. One online experiment found no trust gap for standard dishes and a gap for innovative, AI-invented ones (Califano, Zhang, Spence 2024). X moderate (Jakesch; Califano 2024); S weak (Kraus).

  5. Never let AI change facts after human sign-off, and do not publish nutrition values computed by an LLM without checking. Users trust flawed LLM nutrition figures (Lysova, Trattner, Starke 2026, UMAP '26, US user study N=299). LS strong (post-sign-off rule); X weak.

4. Disclosure checklist per surface type

Marketing page

  • [ ] No AI in the headline or product name unless the reader wants the mechanism (rule 6). The benefit comes first (PAIR).

  • [ ] Any AI claim is falsifiable and has its limit nearby. Nothing overstates capability (UWG §5(2) Nr. 1).

  • [ ] Where AI is the reader's own subject, it appears as what the reader's agents do. A claim such as "no built-in model" is checked against the architecture before it ships.

  • [ ] A site-wide "How we use AI" or methods page, linked from the footer, spells out roles.

  • [ ] No AI imagery: mascots, glowing brains, humanoids.

  • [ ] The marketing page and the methods page do not replace in-context chat disclosure. A generic "Services use AI" is insufficient (Guidelines para 38).

In-product chat

  • [ ] First-turn disclosure plus a persistent badge near the input, in simple language. Other languages if the chat serves speakers of them (I).

  • [ ] Recurring reminders for vulnerable users. Truthful answer to "Bist du ein Mensch?". A handoff path to a responsible human.

  • [ ] No human name, face or "Assistent" persona.

  • [ ] Accessible: screen reader, contrast, plain language.

  • [ ] GDPR Art. 13 notice at collection: purpose, recipients including the model vendor, any transfer abroad, retention. Art. 13(2)(f) logic information if any Art. 22 decision exists.

  • [ ] HAX: say what it can and cannot do, how often it errs, how to correct it.

Generated documents (CVs, cover letters, reports, agent outputs)

  • [ ] If you are the generator's provider: machine-readable marking in metadata, with a second layer where feasible. Deadline 2 Dec 2026 for systems on the market before 2 Aug 2026, otherwise now (counsel).

  • [ ] No visible "AI" stamp on a private document written for a person, such as a CV. There is no legal duty, and a stamp would trigger the social penalty (Reif). The person's facts must be true, and a responsible human reviews.

  • [ ] In the reviewer's interface, mark which parts are drafts and which are confirmed, with one-tap editing (Dietvorst 2018; HAX G9).

  • [ ] Reports that feed decisions about a person: a named human assessment is the decision. No scores go out that a third party would "draw strongly" on (SCHUFA).

  • [ ] Where several agents or people contribute, show which one contributed what, where attribution is verified. It is proof, not a claim of correctness.

Published articles

  • [ ] Classify each piece: plain content (for example a plain recipe), or public-interest content (health, diet and disease, food safety, AI policy).

  • [ ] Public-interest pieces: a substantive fact-check by a competent person, a named editorial owner in the Impressum, and no AI edits after sign-off. Otherwise a clear AI label at the headline, for example the EU "AI" icon.

  • [ ] Every article: one provenance line stating the roles, for example "Rezept mit KI entworfen, von [Name] nachgekocht und geprüft". Include only true steps.

  • [ ] Byline where readers would ask "who wrote this?" (Google). No pages mass-produced without added value.

  • [ ] AI-generated food images are labeled or avoided. A realistic one that misleads about the real dish can be a deep fake that must be disclosed (Art. 50(4)). The "real photo" boost matters (Califano and Spence).

5. Trust and privacy copy

Order of answers: who can read it, then which AI vendor sees it, then what leaves with me. Only then price. I, inference: GDPR Art. 12 and 13 back the content; no study is cited for the order.

  1. Who can read this? Name the exact audience: the members of a private space; for a public space, who can read and through which channels (signed-in users, connected clients with read permission, anonymous search). Say who can write and administer. Say exactly this, and nothing like "secure", "confidential" or "government-grade".

  2. Which AI vendor sees it? When a connected agent reads content, the content goes to the client and model provider the user chose, under their retention, training and location terms. Connecting an agent is a data-access decision, and the copy should say so.

  3. What leaves with me? Name only export and deletion behavior that has been verified live. No claims of complete export, portability or permanence before verification.

  4. Regulated data. Keep patient, client, confidential research and similar data out until the responsible institution approves the data flow, members, agent and model providers, contracts, retention, deletion and incident process. Access boundaries do not replace a processing agreement with each processor (Art. 28(3) GDPR), ethics approval or a security review.

Private is a membership boundary, not a security guarantee. In German, a generic form of the three questions is "Wer sieht das? Welche KI-Anbieter verarbeiten es, wo? Was passiert beim Löschen?", answered in plain words, before any feature talk, and only with verified facts. This doubles as GDPR Art. 13(1)(e) and (f) content.

Never: "secure", "private" standing alone, "DSGVO-konform" without a named basis, or "your data never leaves" when agents send content to a vendor.

6. Conflicts: mandatory disclosure versus not leading with AI

  • The conflict is smaller than it looks. The law regulates where disclosure appears (in context, at first interaction, near published text, on realistic AI images, audio or video), not where copy leads. Headlines and value propositions can stay AI-free, while chat and article surfaces disclose, and so does a realistic AI image that could mislead about the real product.

  • How to satisfy both:

    1. Separate the layers. Marketing leads with the outcome. The methods page and the in-context disclosure carry the AI facts.

    2. Make disclosure a statement of roles, not a label. "KI-Entwurf aus Ihren Notizen; Sie prüfen und entscheiden." Naming the AI and the human role together, shown next to the content at first exposure, can meet the Art. 50(5) "clear and distinguishable" standard, and it cancels the full-automation assumption (Altay; Kirk and Givi).

    3. Put control inside the disclosure. "Sie bestimmen, wie viel Hilfe es gibt." Control is the best-evidenced antidote (Dietvorst 2018).

    4. Use machine-readable marking for documents (Art. 50(2)), rather than visible stamps that stigmatize the people the documents are about.

    5. Use the review exception honestly on articles. A real fact-check plus a named owner removes the Art. 50(4) label, and a line such as "nachgekocht und geprüft" is also the trust cue readers ask for (Kraus).

    6. Affected people versus operators. People who talk to a chat must be told in the chat. Operators choose how much AI to use. Neither needs AI in a headline.

  • Residual tension to accept: the trust cost to the person who discloses persists even with good framing (Schilke and Reimann). Accept a trust cost on disclosed surfaces rather than risk exposure, a UWG §5a omission claim, or an Art. 99 fine.

7. Do not rely on

  • Late disclosure (Luo et al.): chat disclosure is required at the first interaction.

  • The obviousness exception for any public or vulnerable-audience chat.

  • Terms-and-conditions disclosure, a site-wide "we use AI" line, or "powered by LLMs". All are explicitly insufficient.

  • Cosmetic human review (a spell-check, rubber-stamp approval, or AI edits after sign-off) to escape Art. 50(4).

  • Old dates: "high-risk from 2 Aug 2026" is superseded; it is now 2 Dec 2027. The Art. 50(2) grace period to 2 Dec 2026 covers only systems on the market before 2 Aug 2026 and does not cover Art. 50(1) or (4). Outputs and deep fakes generated before 2 Aug 2026 need no retroactive mark or label, and neither does public-interest AI text published before that date. Public-interest AI text generated before 2 Aug 2026 and published on or after it needs the Art. 50(4) label unless the review exception applies (Guidelines para 154).

  • The GDPR Digital Omnibus changing Art. 22 or the information duties: not adopted as of 27 October 2026 (Council mandate only).

  • "Algorithm appreciation" as a licence to headline AI. It fades with expertise and personal stakes, and those describe professional reviewers, hiring and decisions about people.

  • Human-like bots to warm up the experience: expectation trap, legal insufficiency, and backlash from angry users.

  • Absence of evidence on recipes as proof of no penalty. Evidence is thin and US-centric, and the "innovative dish" penalty applies.

  • "Private", "secure" or "DSGVO-konform" as blanket assurances.

Sources (read 29 September 2026 unless a date is given)

Open legal questions

  1. Is a document generator that drafts from a user's notes an Art. 50(2) provider, or does the standard-editing exemption cover it? Whether it was on the market before 2 Aug 2026 decides whether the 2 Dec 2026 grace applies.

  2. When is candidate matching Annex III 4(a) profiling, and so high-risk from 2 Dec 2027?

  3. When does an agent content pipeline make a site a provider under Art. 50(2), rather than a deployer of the model vendor's system?

  4. Which public documents count as public-interest text that needs a named editorial owner?

Oddlynew Copy and Messaging PrinciplesSign in

Referenced by

and more