The short version
- Using AI to write a testimonial is not itself unlawful. 16 CFR § 465.2(a) prohibits a testimonial that materially misrepresents that the testimonialist exists, that they had experience with the product or service, or what that experience was — none of which turns on what drafted the sentences.
- AI-polishing a real customer's words is generally permitted. The limit is 16 CFR § 255.1(b): an endorsement may be conveyed without the endorser's exact words, but not "presented out of context or reworded so as to distort in any way the endorser's opinion or experience".
- The rule reaches whoever passes it on. Section 465.2(b) makes it a violation to disseminate a testimonial the business "knew or should have known" misrepresented those same three facts, so an agency's copy becomes your exposure the moment you publish it.
- The per-violation penalty in 15 U.S.C. § 45(m)(1)(A) is $10,000 as written, raised each January for inflation in the table at 16 CFR § 1.98. It is a maximum a court sets on culpability and ability to pay, not an automatic charge per review.
Nothing in the federal rule on consumer reviews and testimonials mentions artificial intelligence. That is not an oversight waiting on a second rule. The rule polices three facts about a person, and a language model is invisible to all three.
The rule never asks who typed the words
16 CFR part 465 was published at 89 FR 68077 in August 2024 and took effect that October. Section 465.2(a) makes it an unfair or deceptive act to write, create or sell a review or testimonial that materially misrepresents, expressly or by implication: that the testimonialist exists; that they used or otherwise had experience with the product, service or business; or their experience with it.
Read the three clauses as one question. Did a real person have the experience these words describe? If yes, the tool that shaped the sentences is beside the point. If no, the misrepresentation is complete however it was produced.
The column decides it. The row decides nothing.
Who produced the sentences?
Whose experience is described?
Nobody's — it was invented
A named customer's, and they had it
A model generated them
Fake, plainly
No testimonialist and no experience. Clauses (1) and (2) are both engaged.
Drafting, not fabricating
Only clause (3) stays live, and only if the model adds specifics nobody gave you.
The customer wrote them
Still fake
A real person's own words about an experience they never had. Handwriting cures nothing.
Nothing to argue about
The ordinary case, and the reason the rule leaves most testimonials alone.
Rewriting a real review is allowed until the experience changes
The third clause protects not the customer's authorship but their experience, and the Endorsement Guides at 16 CFR part 255 say the same thing from the other direction. Section 255.1(a) requires an endorsement to reflect the honest opinions, findings or experience of the endorser. Section 255.1(b) then allows it to be conveyed without the endorser's exact words, but not "presented out of context or reworded so as to distort in any way the endorser's opinion or experience".
So the question is never how much of the text a model produced. It is whether any statement in it is a fact about the customer's experience that the customer did not supply. Grammar, length and tone are not facts about the experience. A number is. A timeframe is. A named result is. A sentiment nobody expressed is.
How far a rewrite can travel
Tidying the words
Condensing and reordering
Filling in the specifics
Writing the experience
Publishing it on your own site does not put it outside the rule
A common misreading turns on the word "review". Section 465.1 defines a consumer review narrowly — an evaluation submitted by a consumer and "published to a website or platform dedicated in whole or in part to receiving and displaying such evaluations". A quote block on your homepage is not that, so people conclude the rule stops at the platform boundary.
It does not, because the same section defines a consumer testimonial separately and far more widely: any "advertising or promotional message" consumers are likely to believe reflects the opinions, beliefs or experiences of a consumer who has had experience with a product, service or business. That is the quote block. It is also the case study, the video clip and the star rating printed on your pricing page.
| What you are publishing | How part 465 classifies it | What the classification changes |
|---|---|---|
| A star rating left on a marketplace | Consumer review | Also reachable by § 465.4, on incentives conditioned on a sentiment |
| A quote block on your own site | Consumer testimonial | No platform and no submission form. The rule applies anyway |
| A written case study naming a client | Consumer testimonial | Every stated result forms part of "the testimonialist's experience" |
| A composite client built from several | A testimonial by someone who does not exist | § 465.2(a)(1), and § 255.2(c) wants actual consumers or a disclosure that they are not |
Case studies are where generative drafting does the most damage, because the genre invites invention. Ask a model to write up a client engagement and it supplies the percentage, the payback period and the headcount, because that is what case studies contain. If nobody measured them, no hedge in the surrounding paragraph makes them a description of anyone's experience. The same discipline governs a creator repeating a claim on your behalf, which is what the influencer agreement and influencer usage rights are for.
Model release template
Full text, free to read and copy — permission to use a real customer's name, words and likeness, which is also the record proving the testimonialist exists.
Passing on somebody else's fabrication is its own violation
Section 465.2(b) is the half that surprises people. It makes it a violation to purchase a consumer review, or to disseminate or cause the dissemination of a consumer testimonial, which the business "knew or should have known" materially misrepresented those same three facts. The standard is constructive. You need not have generated anything to be inside it.
- An agency's copy is your copy. Testimonials arriving in a campaign package are disseminated by you the moment they go live.
- Anything inherited in an acquisition. The site came with a testimonials page. It did not come with the evidence behind it.
- Your own archive. A five-year-old quote whose author nobody can now identify is one you cannot defend, however genuine it once was.
What "per violation" actually means
A rule matters more than the general prohibition on deception because of 15 U.S.C. § 45(m)(1)(A), which lets the Commission sue for a civil penalty against anyone who violates a trade regulation rule "with actual knowledge or knowledge fairly implied on the basis of objective circumstances" that the act is unfair or deceptive and prohibited by that rule.
The figure written into the statute is $10,000, and nobody is ever sued for $10,000: the Federal Civil Penalties Inflation Adjustment Act requires the maximum to be raised every January. The operative table is 16 CFR § 1.98, and the adjustment effective January 2025 put the ceiling above $53,000. Read the table rather than any article quoting it, this one included.
Two things get lost when that number is repeated. It is a ceiling, not a tariff — § 45(m)(1)(C) directs the court to weigh culpability, any history of prior conduct, ability to pay, the effect on continuing in business, and such other matters as justice may require. And penalties are not the whole of it: 15 U.S.C. § 57b lets the Commission seek consumer redress, including rescission, refunds and damages, generally within three years.
This machinery exists because of AMG Capital Management v. FTC, decided in April 2021, where the Supreme Court held that section 13(b) "does not authorize the Commission to seek, or a court to award, equitable monetary relief such as restitution or disgorgement". After AMG, deception that breaches no rule became expensive to stop and cheap to have committed. Part 465 is what put the money back on the table for fake reviews.
The rule and the guides are not the same instrument
Part 465 — the rule
- Civil penalties per violation
- Consumer redress under § 57b
- Written as flat prohibitions
Both
- Turn on existence and experience
- Reach testimonials on your own site
- Indifferent to who drafted the words
Part 255 — the guides
- Administrative interpretations, not rules
- Bite through section 5 instead
- Cover material-connection disclosure
The tool vendor is not the defendant to watch
In 2024 the Commission sued Rytr, an AI writing service, on a means-and-instrumentalities theory: its testimonial generator allegedly produced "detailed reviews that contain specific, often material details" unrelated to what the subscriber typed, with no limit on volume and "no or de minimis legitimate uses". A consent order that December barred it from selling any review or testimonial generation service for twenty years.
On 22 December 2025 the Commission reopened and set that order aside, finding the pleaded facts did not establish a section 5 violation and that the order burdened a nascent industry without benefiting consumers. Its reasoning is the sentence worth keeping: "Treating as categorically illegal a generative AI tool merely because of the possibility that someone might use it for fraud is inconsistent with our precedents and common sense."
That is not a retreat from the rule. In the same month the Commission sent warning letters to ten businesses over possible part 465 violations, the template example being paying employees to obtain five-star reviews from friends and family. The generator was let go; the businesses publishing experiences nobody had were not. It is the distinction the chatbot disclosure rules draw — the software is never the speaker.
Fixing what is already on the site
Most businesses reading this have generated nothing, and still cannot evidence the testimonials on their own homepage. That is the position the rule is hardest on, and it is fixable in an afternoon.
- 1
List every published testimonial
Homepage, pricing page, case studies, sales decks, app store listings, ad creative, the PDF a salesperson still emails. Anything a reader takes as another customer's experience is in scope.
- 2
Tie each one to an identifiable person
A name, an account, an email thread. A quote you cannot trace to a real customer cannot be defended under the first clause, and comes down first.
- 3
Check every specific against what they said
Numbers, timeframes, named outcomes, competitor comparisons. Anything that entered at the drafting stage rather than from the customer goes, or gets confirmed.
- 4
Get the final wording approved in writing
Send the exact version that will be published and keep the reply. Bundle the name and likeness permission in so one record covers both.
- 5
Set the rules for reviews you host
If customers can post to your site, your website terms of use should say reviewers must have used the product and that incentives are never conditioned on sentiment.
The rule is easier to live with than its reputation suggests, because it never asks you to prove a negative about your tooling. It asks you to name the customer and show what they said. Businesses that could already do that lost nothing in October 2024. Those that could not had the problem long before generative models existed.
Sources
- 16 CFR § 465.1 — definitions
- 16 CFR § 465.2 — fake or false reviews and testimonials
- 16 CFR part 465 — contents and source note
- 16 CFR § 465.4 — buying positive or negative reviews
- 16 CFR § 255.0 — status of the Endorsement Guides
- 16 CFR § 255.1 — general considerations for endorsements
- 16 CFR § 255.2 — consumer endorsements
- 15 U.S.C. § 45 — civil penalties for rule violations, § 45(m)
- 16 CFR § 1.98 — adjusted civil penalty maximums
- 15 U.S.C. § 57b — consumer redress for rule violations
- AMG Capital Management v. FTC (2021)
- 15 U.S.C. § 1125 — Lanham Act false advertising
- Goodwin — FTC finalises the consumer reviews rule, effective 21 October 2024
- Covington Inside Privacy — FTC sets aside the Rytr final order
- DLA Piper — FTC warning letters on fake consumer reviews and AI
General information, not legal advice. This guide explains how these documents and rules generally work. Law varies by jurisdiction and changes, and none of it is applied to your circumstances here. For anything consequential, consult a licensed attorney where you are.
Frequently asked
Is it illegal to use AI to write customer testimonials?
Not on its own. The federal rule at 16 CFR part 465 tests three things: whether the testimonialist exists, whether they had experience with the product, and whether the words describe the experience they actually had. A model drafting sentences for a real customer's real experience engages none of them. A model producing the experience itself engages all three, and that version carries civil penalties.
Can I post AI-generated reviews on my website?
Not if no customer stands behind them. Section 465.1 defines a consumer testimonial as any advertising or promotional message consumers are likely to believe reflects a customer's opinions or experiences, which covers a quote block on your own homepage as squarely as a marketplace listing. Hosting it yourself changes nothing. Section 255.2(c) adds that advertising portraying actual consumers must use actual consumers, or clearly disclose that it does not.
What is the FTC penalty for fake reviews per violation?
The figure in 15 U.S.C. § 45(m)(1)(A) is $10,000, raised every January under the Federal Civil Penalties Inflation Adjustment Act; the current maximum sits in the table at 16 CFR § 1.98, which stood above $53,000 after the adjustment effective January 2025. It is a ceiling rather than a charge. Section 45(m)(1)(C) directs the court to set the amount on culpability, prior conduct, ability to pay and business impact.
Can I use AI to rewrite a real customer review?
Yes, within a limit that has nothing to do with the tool. The Endorsement Guides permit an endorsement to be conveyed without the endorser's exact words, but not reworded so as to distort their opinion or experience in any way. Fixing grammar, cutting length and reordering sentences stay inside that. Adding a figure, a timeframe, a named result or a sentiment the customer never expressed does not.
Is an AI-generated case study a fake testimonial?
If the client is composite or invented, yes — it misrepresents that the testimonialist exists, which is the first clause of the prohibition. If the client is real and the results were measured, it is a testimonial like any other and the drafting can be automated safely. The risk sits in the middle case: a real client with invented specifics. Every number and outcome has to come from them.