Transparency is the AI duty most likely to apply to you today, whatever your size and wherever you are. It is also the cheapest to meet. The rules come in four flavours.
1. "You are talking to a machine"
- EU AI Act, Article 50 (since 2 August 2026): people must be told they are interacting with an AI system unless it is obvious. Emotion recognition and biometric categorisation require notice too.
- Utah's AI Policy Act (since May 2024, narrowed in 2025): disclose generative AI when a consumer asks, and proactively in "high-risk" interactions involving health, financial or legal advice or sensitive data. Businesses that disclose clearly at the outset get a safe harbour.
- Texas TRAIGA (since 1 January 2026): government agencies and healthcare providers must tell people when they are interacting with AI.
2. "This content was made by AI"
- EU AI Act, Article 50: providers must mark synthetic audio, image, video and text in a machine-readable way; deployers must label deepfakes and AI-written text on matters of public interest. Systems already on the market before 2 August 2026 have until 2 December 2026 for the marking duty (Digital Omnibus).
- California SB 942 (operative 2 August 2026, after AB 853): generative providers with more than a million monthly users must offer a free detection tool, embed latent disclosures, and offer a visible one.
3. "AI is part of this decision about you"
- Illinois HB 3773 (since 1 January 2026): employers must notify applicants and employees when AI is used in recruitment, hiring, promotion, discipline or discharge decisions.
- NYC Local Law 144: notice to candidates at least 10 business days before using an automated employment decision tool, plus a public summary of the bias audit.
- Ontario Employment Standards Act (since 1 January 2026): employers with 25 or more employees must state on public job postings whether AI is used to screen, assess or select applicants.
- Quebec Law 25: notice when a decision is based exclusively on automated processing.
- Colorado SB 26-189 (from 1 January 2027): notice at the point of interaction, and a plain-language explanation within 30 days of an adverse outcome.
4. "This is what the model was trained on"
- California AB 2013: training-data summaries for generative AI.
- EU AI Act: general-purpose model providers must publish a summary of training content.
Write one disclosure set
You do not need a different notice per law. Draft three reusable texts and place them where each law asks:
- Interaction notice: "You are chatting with an automated assistant. A person can take over on request."
- Decision notice: "We use software, including AI, to help review applications. A person makes the final decision. You can ask how it was used and request a review."
- Content label: "This image/text was generated with AI."
Keep them short, in plain language, and visible before the interaction starts, not buried in terms.
The failure mode: overclaiming
Transparency also means not saying what is untrue. The FTC's case against DoNotPay (2024) turned on marketing a "robot lawyer" without evidence it could do the job; the company paid USD 193,000 and had to warn past customers. Marketing copy about your AI is a compliance document. Say what it does, and what it does not.