Law and Policy: Rules on AI-Generated Text and Its Labels

No law anywhere bans writing with AI. What the rules do is narrower and more specific: they require labels in certain settings, forbid certain deceptions, set who owns what, and govern how companies may describe their tools. Several of the best-known "AI labeling" laws do not reach plain text at all. This lesson goes through the rules that matter for machine-written words, as of October 2026, and says what each one requires and who it covers. It takes no side on whether any of them is good policy. For the wider picture of AI law, see AI law and regulation in course 15.

The EU AI Act, Article 50

Regulation (EU) 2024/1689, the AI Act, entered into force August 1, 2024. Chapter IV, Article 50, holds its transparency duties. Its general date of application, set in Article 113, is August 2, 2026, and Article 50 has no earlier or later date of its own. So these duties are now in effect. Four paragraphs matter for text.

Paragraph	Who	What it requires	Exceptions in the text
50(1)	Providers of AI systems meant to interact directly with people	Design the system so people are told they are interacting with AI	Where that is obvious to a reasonably well-informed, observant person; some law enforcement uses
50(2)	Providers of systems, including general-purpose ones, that generate synthetic audio, image, video or text	Mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated, as far as technically feasible	Assistive functions for standard editing; systems that do not substantially alter the input or its meaning; some law enforcement uses
50(4), first part	Deployers of systems that make deepfake images, audio or video	Disclose that the content was artificially generated or manipulated	Lighter duty for evidently artistic, satirical or fictional works
50(4), second part	Deployers of systems that generate or manipulate text published to inform the public on matters of public interest	Disclose that the text was artificially generated or manipulated	Text that went through human review or editorial control, where a natural or legal person holds editorial responsibility; some law enforcement uses

Paragraph 5 says the information must be given "in a clear and distinguishable manner" no later than the first interaction or exposure.

Two points get missed. First, 50(2) is the only place where text and the word "detectable" meet. It is a duty on the companies that build generators, and it is qualified by technical feasibility and the limits of different content types. Lesson 4, Watermarks and provenance, explains why marking text is harder than marking images. Second, the public-interest text duty in 50(4) falls away when a human with editorial responsibility reviewed the text. A newspaper that has an editor sign off on an AI-drafted article is outside that sentence. An automated site that posts unreviewed AI news is inside it.

The 2026 changes

Regulation (EU) 2026/1744, the "Digital Omnibus on AI," was signed July 8, 2026, published in the Official Journal July 24, and entered into force on the third day after publication. For Article 50 it did two things. It added a paragraph to Article 111 giving providers whose generative systems were on the market before August 2, 2026 until December 2, 2026 to comply with the marking duty in 50(2). And it rewrote 50(7) so the Commission assesses whether codes of practice on detection, marking and labeling are adequate.

The Commission's Code of Practice on Transparency of AI-generated Content was published in June 2026. It has a section for providers on marking and detection and a section for deployers on labeling deepfakes and AI-generated text. Signing is voluntary, but the Commission says the Article 50 duties themselves are legal obligations, and that it and the AI Board confirmed the code is an adequate way to show compliance. Companies that do not sign must show their own measures are adequate, case by case, to national market surveillance authorities. The Commission reported about 190 signatories by the end of July 2026.

Penalties: Article 99(4) lists the Article 50 transparency duties among those that carry administrative fines of up to 15 million euros or, for a company, up to 3 percent of total worldwide annual turnover, whichever is higher.

The FTC's rule on fake reviews

The Federal Trade Commission's Trade Regulation Rule on the Use of Consumer Reviews and Testimonials was published August 22, 2024 at 89 FR 68034 and took effect October 21, 2024. It is codified at 16 CFR Part 465.

Section 465.2 is the part about invented reviews. It is an unfair or deceptive practice for a business to write, create or sell a consumer review, consumer testimonial or celebrity testimonial that materially misrepresents that the reviewer exists, that the reviewer actually used the product or service, or what the reviewer's experience was. Buying such reviews, or spreading testimonials the business knew or should have known were false, is covered too. Other sections cover buying positive or negative reviews, undisclosed insider reviews, company-controlled "independent" review sites, suppressing reviews, and selling or buying fake social media influence.

The rule never says "AI" in its operative text. It does not need to: a review a chatbot wrote for a person who never bought the product misrepresents that the reviewer used it. The FTC's explanation in the Federal Register was direct: "AI-generated reviews are covered by the final rule." Because it is a trade regulation rule, violations let the FTC seek civil penalties, which the agency said was one of the rule's purposes.

On December 22, 2025, the FTC sent warning letters to ten companies under the rule, according to an August 2026 summary by the law firm Kohrman Jackson and Krantz, which put the maximum civil penalty at $53,088 per violation, subject to inflation adjustment.

FTC actions on AI writing and detection tools

Workado: claims about a detector

In April 2025 the FTC filed a complaint against Workado, which sells an "AI Content Detector." Workado advertised that the tool spotted AI-written text with 98 percent accuracy. The FTC said independent testing found accuracy closer to 53 percent on general-purpose content, and that the model had been trained or fine-tuned to classify only academic content even though Workado said it was built from a wide range of material, including blog posts and Wikipedia entries. The Commission later voted 3-0 to make the consent order final; news reports of the final approval appeared at the end of August 2025. The order bars Workado from making claims about the effectiveness of any AI detection product unless the claims are supported by competent and reliable evidence when made, requires it to keep that evidence, requires it to email eligible customers about the settlement, and requires compliance reports for four years.

Rytr: a writing tool, an order and a reversal

In 2024 the FTC brought a case against Rytr, an AI writing assistant, alleging that a feature for generating reviews and testimonials could be used to produce fake reviews that would be posted elsewhere. Rytr did not post reviews itself. It agreed, without admitting liability, to a consent order barring it from offering any AI service capable of generating consumer reviews or testimonials.

In December 2025 the FTC reopened the case and set the order aside by a 2-0 vote. According to a summary by the law firm Parker Poe, the Commission concluded the complaint did not adequately plead a violation of Section 5 of the FTC Act, that the theory relied on possible misuse by customers rather than on Rytr's own conduct or identified consumer harm, and that the order barred a whole category of AI functionality with lawful uses. It cited the administration's AI executive order and AI Action Plan. The fake review rule itself was untouched. A business that uses any tool to write fake reviews of its own products is still covered by Part 465.

The U.S. Copyright Office

The Copyright Office's position on AI-generated text comes from three documents.

The March 2023 registration guidance. A statement of policy published at 88 FR 16190 and effective March 16, 2023 says applicants have a duty to disclose AI-generated content in a work submitted for registration and to briefly explain what the human author contributed. People who use AI can claim copyright in their own contributions using the Standard Application. AI-generated content that is "more than de minimis" should be explicitly excluded, in the "Limitation of the Claim" section, with a description such as "[description of content] generated by artificial intelligence." Anyone whose registration was already issued without that disclosure should correct the record through a supplementary registration.

Part 2 of the AI report, January 29, 2025. The Office concluded that existing law can answer copyrightability questions without new legislation. Using AI as a tool to assist human creativity does not affect protection. Copyright covers a human author's original expression even when the work also contains AI-generated material, but it does not cover purely AI-generated material or material where the human lacked enough control over the expressive elements. Prompts alone, given how generally available technology works, do not provide sufficient control. A human can own the selection, coordination or arrangement of AI output and creative modifications of it. Each case is decided on its facts. The Office found the case had not been made for new protection for AI-generated content.

Thaler v. Perlmutter. Stephen Thaler applied to register a picture, "A Recent Entrance to Paradise," naming his AI system, the "Creativity Machine," as sole author. The Office refused. On March 18, 2025, the U.S. Court of Appeals for the D.C. Circuit affirmed in an opinion by Circuit Judge Millett (No. 23-5233): the Copyright Act of 1976 "requires all eligible work to be authored in the first instance by a human being." The court did not decide whether the Constitution also requires human authorship, and it did not reach Thaler's argument that he was the author because he made and used the machine, finding that argument waived before the agency. The Supreme Court denied Thaler's petition for review on March 2, 2026 (docket 25-449).

The Thaler work was an image. The rule is the same for words: a novel, report or article generated entirely by a model has no human author for copyright purposes. Text a person wrote, or meaningfully selected, arranged or changed, can be protected for those human parts. And if you register a work, the Office expects you to say which parts are which.

Idaho

Synthetic media in elections: Idaho Code 67-6628A

House Bill 664, the "Freedom From AI-Rigged (FAIR) Elections Act," passed the House 61-8-1 on March 7, 2024 and was signed by the governor on March 25, 2024 (2024 Idaho Session Laws, chapter 172). It is codified at Idaho Code 67-6628A.

The law defines synthetic media as an audio recording or video recording of a person's speech or conduct, created with generative adversarial network techniques or other digital technology, that is realistic but false, that a reasonable person would take for a real event that did not happen, and that gives a very different impression than the unaltered original would. An AI-written flyer, email or letter is not an audio or video recording, so it is outside the statute.

What the law does when it applies:

• A candidate whose speech or action is deceptively represented through synthetic media in an electioneering communication may seek an injunction, and may sue the information content provider for general and special damages. A court may award the prevailing party reasonable attorney's fees and costs.
• The plaintiff must prove use of synthetic media by clear and convincing evidence.
• It is an affirmative defense if the communication carries the disclosure "This (video/audio) has been manipulated." For video, the text must be easily readable and stay on screen for the whole video. For audio, it must be spoken clearly at the beginning and end, and at least every two minutes if the audio runs longer than two minutes.
• The creator is the one liable. The medium carrying it is liable only if the medium strips the disclosure or alters content into synthetic media. That exception does not apply to a federally licensed broadcast station transmitting a communication subject to 47 U.S.C. 315.

Section 67-6628A does not set its own time window. But it uses the term "electioneering communication," which Idaho Code 67-6602 defines for the whole chapter as a communication that unambiguously refers to a candidate, is distributed within 30 days before a primary or 60 days before a general election, and reaches voters for that office. That definition excludes news articles, editorials and letters to the editor in publications not controlled by a candidate, committee or party.

Chatbots: the Conversational AI Safety Act

Idaho's 2026 Senate Bill 1297, the Conversational AI Safety Act, was signed March 31, 2026 (2026 Idaho Session Laws, chapter 249) and takes effect July 1, 2027. It applies to operators of conversational AI services, meaning publicly available programs that primarily simulate human conversation through text, visuals or voice, with exclusions such as tools designed and marketed for developers and researchers. If reasonable people would be misled into believing they are talking to a human, the operator must clearly and conspicuously disclose that the service is artificial intelligence. For minor account holders, the disclosure must be a persistent visible notice, or appear at the start of each session and at least every three hours. The bill also requires a protocol for responding to users who express suicidal thoughts. The attorney general enforces it, with civil penalties of $1,000 per violation up to $500,000 per operator, or actual damages if greater; it creates no private right to sue. Unlike the election law, this one reaches text, because a chat window is text.

California's AI Transparency Act

California's SB 942, the AI Transparency Act, is in the Business and Professions Code starting at section 22757. AB 853 (Statutes of 2025, chapter 674) amended it and set its operative date at August 2, 2026. The official legislative site, leginfo.legislature.ca.gov, refused automated access while this lesson was checked, so the details below come from Justia's reproduction of the code and should be read against the official text.

• Who: a "covered provider," meaning a person that creates, codes or otherwise produces a generative AI system that has over 1,000,000 monthly visitors or users and is publicly accessible in California.
• Detection tool (22757.2): a covered provider must offer a free, publicly accessible AI detection tool, with an API, that lets users check whether image, video or audio content was created or altered by the provider's system and shows any system provenance data, but no personal provenance data.
• Disclosures (22757.3): the provider must offer users the option of a visible, hard-to-remove label on image, video or audio content, and must embed a latent disclosure in content its system creates, conveying where feasible the provider's name, the system name and version, the time and date, and a unique identifier. If a licensee disables the latent disclosure, the provider must revoke the license within 96 hours of discovering it.
• Platforms and devices: large online platforms must detect and display standard provenance data starting January 1, 2027; GenAI hosting platforms may not knowingly offer systems that lack the required disclosures, also from January 1, 2027; capture device makers must offer and default to latent disclosures on devices first produced for sale in California on or after January 1, 2028.
• Penalty (22757.4): $5,000 per violation, each day a discrete violation, in actions brought by the attorney general, a city attorney or a county counsel.

The law's definition of a generative AI system mentions text. The duties do not. The detection and disclosure sections cover image, video and audio content, alone or combined. A chatbot's written answers carry no California latent disclosure requirement under this act.

Which rules reach plain text

Rule	Reaches AI-written text?	Who it covers	Status, October 2026
EU AI Act 50(2)	Yes: machine-readable marking	Providers of generative systems	Applies from Aug 2, 2026; Dec 2, 2026 for systems already on the market
EU AI Act 50(4)	Yes: published public-interest text, unless human-reviewed with editorial responsibility	Deployers	Applies from Aug 2, 2026
16 CFR Part 465	Yes: fake reviews and testimonials, however written	Businesses	In effect since Oct 21, 2024
Copyright Office guidance	Yes: disclosure and exclusion in registrations	Registrants	Effective Mar 16, 2023
Idaho Code 67-6628A	No: audio and video only	Creators of electioneering communications	In effect since Mar 25, 2024
Idaho SB 1297 (2026)	Yes: chatbots that could pass as human	Operators of conversational AI services	Takes effect Jul 1, 2027
California AI Transparency Act	No: image, video and audio duties	Providers with over 1,000,000 monthly users; later platforms and device makers	Operative Aug 2, 2026