Say it only if you can prove it

Build-along · ~10 min at your own keyboard

AI writes fast. It does not answer for what it writes. You do. Here is what binds a small business that uses AI. Every rule carries a tag: [P] primary text, [S] secondary source, [M] unverified. The dates and sources are below the lesson. One: the FTC acts on deceptive claims. 'AI-powered', 'guaranteed' and earnings claims all need proof that you hold. [S] Two: fake or AI-written reviews and testimonials are banned. So is paying for reviews that must be positive. [S] Three: if someone got paid or got a free product to endorse you, say so clearly. That reaches an AI-made endorser a customer would read as real. [S] Four: raw AI output is not copyrightable. A logo or jingle made only by AI is not yours to protect. [S] Five: an AI voice that calls customers, or a bulk text, needs prior consent, and each call or text is priced on its own. [M] Six: Tennessee protects voice and likeness, AI copies included. Never clone a singer's, employee's or customer's voice without written consent. [S] Seven: if customer data went into a consumer AI tool, tell the owner the same day. Tennessee runs a notice clock. [S] Eight: hiring law applies to AI screening, whoever sold the tool. AI may rank. A person decides and writes down why. [S/M] And the paste warning that sits under all of it: do not paste customer lists, employee records, client files or contracts into a consumer plan. Training and disclosure risk are the reasons. Business plans differ; the vendor notes below say how. [S] Three uses are clearly fine: drafting, summarizing, brainstorming and analyzing your own non-sensitive data; marketing copy that a person reviews and that real proof backs; a customer chatbot that says it is a bot. [S] The pattern that gets people sanctioned is a claim with no proof behind it. The FTC still brings AI-washing, fake-review and earnings-claim cases; one of them was against an AI sales tool that promised results it could not show. [S] One related order against a writing tool was set aside. That limits what a tool maker answers for. It does not limit what you answer for. [S] The rule to repeat: AI drafts, I decide, and I can prove every claim. Draft copy, not legal advice. Sources tagged [P] primary, [S] secondary, [M] unverified. Reviewed by: nobody yet.

Ladder · Level 1The long version: Prompting ↗

Do this

0/5

Checkpoints

▸You should now see a proof location beside every claim, or the claim cut or marked NEED PROOF.
▸You should now see one plan name written down for the chat that made it.
▸You should now be able to say the rule from memory: AI drafts, I decide, and I can prove every claim.

The three ways this goes wrong

Every claim got a proof location that is really a feeling: 'we're known for it', 'customers love it'. — A proof location is a thing you could hand to someone: a file, a receipt, a signed quote. Go back through and replace each feeling with a thing, or cut the claim.
The plan name is 'ChatGPT' or 'Claude' with no plan. — Open the assistant's settings or billing page and read the plan name off it. Free, Plus, Pro, Team and Enterprise carry different rules, and the vendor note below says which.
Customer or employee data went into a consumer plan and you decided to deal with it later. — Tell the owner today. Then turn off training in the assistant's settings, or move that work to a business plan. The later you tell them, the worse the clock looks.

Dated facts in this lesson: The FTC acts on deceptive claims under section 5 of the FTC Act. 'AI-powered', 'guaranteed' and earnings claims all need proof the business holds before the claim is made. [S] secondary source; the statute page was not re-read (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); The FTC's Consumer Reviews and Testimonials Rule (16 CFR Part 465) bans fake reviews and testimonials, including AI-written ones and reviews the business did not receive. In effect since October 21, 2024. [S] secondary source (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); The FTC Endorsement Guides (16 CFR Part 255) require a clear disclosure when an endorser was paid or got a free product, and read an AI-made endorser a customer would take as real the same way. [S] secondary source (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); The Copyright Office's AI report, Part 2 (January 2025), says output made only by AI is not copyrightable; human authorship is required. The Supreme Court declined to hear Thaler v. Perlmutter on March 2, 2026, so that stands. [S] secondary source (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); FCC ruling 24-17 (February 8, 2024) treats AI-generated voices as 'artificial' voices under the TCPA, so an AI voice that calls customers, or a bulk text, needs prior consent; statutory damages run $500 to $1,500 per call or text. [M] not verified against the ruling text (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); Tennessee's ELVIS Act (T.C.A. 47-25-1101 and following, in effect July 1, 2024) protects a person's voice and likeness, AI copies included, so cloning a singer's, employee's or customer's voice needs written consent. [S] secondary source; bill page not re-fetched (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); Tennessee's breach notification law (T.C.A. 47-18-2107) requires notice to affected people within 45 days of discovery. [S] secondary source; the statute page returned an error when checked (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); Hiring law applies to AI screening whoever sold the tool. Illinois HB 3773 (in effect January 1, 2026) bars discriminatory AI use in employment decisions and requires notice; in Mobley v. Workday (2025) a court let an age-discrimination collective proceed on the theory that a vendor can be the employer's agent. [S/M] secondary sources; neither primary page re-fetched (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); The FTC still brings AI-washing, fake-review and earnings-claim cases: FTC v. Air AI (2025) alleged deceptive earnings and performance claims for an AI sales tool. One order against a writing tool (Rytr) was set aside in 2025, which limits what a tool maker answers for, not what its user answers for. [S] secondary sources; case pages not re-fetched (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); Claude consumer plans: training on your chats is a user setting; opting in extends retention to up to 5 years, otherwise 30 days. Claude Team, Enterprise and the API do not train by default and delete after 30 days, with zero-retention options on Enterprise and API. [S] secondary source; privacy center not re-fetched (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); ChatGPT consumer plans train on chats by default unless the user opts out. ChatGPT Business, Enterprise and the API do not train by default. Whether consumer terms bar commercial use was not verified. [S/M] OpenAI's pages returned an error when checked (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); Gemini Apps (consumer): with activity on, human reviewers may read chats and retention runs up to 3 years. Gemini in Workspace is not used to train. [M] consumer row unverified; Workspace row [S] (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it); Consumer Copilot terms differ from the work product; check its privacy menu. Microsoft 365 Copilot does not use prompts to train foundation models and stays inside your tenant boundary. [S] secondary source (researched Oct 5, 2026, not yet confirmed by a person: check the source before you rely on it).

Share what happened

Paste the piece with a proof location, a cut, or NEED PROOF beside every claim, then one line with the plan name of the chat that made it.

Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].

🔒 Don't paste confidential, personal-health, or payment info.