What did the FTC allege?
The FTC announced the case on 25 September 2024 as one of five actions in Operation AI Comply. DoNotPay marketed a subscription that would let consumers “sue anyone”, draft legal documents and check small-business websites for legal violations, and promoted itself as a substitute for a human lawyer.
Paragraph 20 of the complaint is the part every AI deployer should read. It alleges that “DoNotPay did not test whether the Service’s law-related features operated like a human lawyer,” that employees “have not tested the quality and accuracy” of most of its legal outputs, and that the company had not employed or retained attorneys to test them.
The complaint also describes how little the product asked. An assault victim using “Sue Now” was asked for a “short phrase” describing the incident, and the service then produced a half-page demand letter with a monetary demand.
What does the final order require?
The Commission voted 5-0 on 16 January 2025 to approve the final order, announced on 11 February 2025. DoNotPay must pay $193,000, notify consumers who subscribed between 2021 and 2023 about the settlement, and may not claim its service performs like a real lawyer, or substitutes for any professional service, unless it has evidence to back the claim.
The amount is small. The principle is what matters: a performance claim about an AI product needs testing behind it before it is made.
Is the FTC still pursuing AI cases?
Yes, on a narrower footing. In December 2025 the FTC voted 2-0 to reopen and set aside its 2024 order against Rytr, another Operation AI Comply case, citing the White House AI Action Plan and finding the original complaint did not adequately plead a violation. The Rytr theory rested on how customers might misuse a writing tool.
In setting it aside, the FTC said it retains full authority to pursue AI cases involving deception, fraud or demonstrable consumer harm. DoNotPay was a deception case about the company’s own claims, which is the kind the FTC has said it will keep bringing.
Why does this matter for agents that make promises or offers?
There are two layers. The first is what you say about your agent: claims to customers, partners or investors that it “resolves refunds instantly,” “never exceeds policy” or “handles disputes like your best rep” are performance claims, and DoNotPay shows the FTC expects testing to back them.
The second is what the agent says to consumers. A discount, a refund timeline or a statement of policy made by your agent is your representation to the customer. An agent that says “your refund has been processed” when no refund exists, or offers a deal that does not exist, makes a false statement on your behalf.
| Claim you might make | Evidence a regulator would expect | How to produce it |
|---|---|---|
| “Our agent follows our refund policy.” | Tests showing policy-compliant decisions, including under pressure | Run cooperative and adversarial scenarios per policy clause. Record pass rates |
| “It never gives away more than its limit.” | Evidence that caps hold against manipulation, not only in normal use | Attempt cap bypasses: splitting, stacking, fake approvals, injected instructions |
| “As good as a human agent.” | A like-for-like comparison with human performance | Score the agent and human reps on the same scenario set with the same rubric |
| “Accurate answers about your account.” | Accuracy measured against the system of record | Compare agent statements with ledger data across a sample of conversations |
Checklist before you make claims about your agent
DoNotPay is about marketing claims, but the same discipline protects what your agent says to customers.
- Write down every claim you make about the agent’s accuracy, limits or equivalence to staff.
- For each claim, keep the test set, the transcripts and the results that support it, dated and versioned.
- Include adversarial customers in the tests, since “never exceeds policy” is a claim about worst cases.
- Have a domain expert (legal, finance, collections) review a sample of outputs, as the FTC faulted DoNotPay for not doing.
- Retest after every model, prompt or policy change, and retire claims the latest results no longer support.
- Make the agent unable to confirm refunds, discounts or deadlines that do not exist in your systems.
Questions
Is there an FTC rule that requires testing AI?
No specific rule. Section 5 of the FTC Act bans deceptive practices, and the FTC expects companies to have evidence for objective claims before they make them. DoNotPay shows that for an AI product, testing is that evidence.
Does this apply if we never advertise the agent as AI?
The DoNotPay order was about advertising claims. But what any customer-facing agent tells consumers about prices, refunds and terms is still a representation by your business, and deception law applies to it whoever or whatever said it.
This page summarizes public FTC actions and is not legal advice.
Free checklist
Get the guardrail checklist for agents that move money.
A one-page PDF of the controls and the test for each. Download it right away, and we email you the link.
Sources
- FTC: Crackdown on deceptive AI claims and schemes (25 September 2024)
- FTC: Finalizes order with DoNotPay (11 February 2025)
- FTC: In the Matter of DoNotPay, Inc., complaint
- Parker Poe: FTC reopens and vacates prior order against AI company (December 2025)