# Can the FTC go after an AI company because its model hallucinates, or because of the accuracy numbers in its marketing?

**Short answer:** The bigger risk is what you claim, not the hallucinations themselves. In a proposed policy statement published in the Federal Register on July 7, 2026, the FTC said it does not believe hallucinations, which stem from the technological and resource limits of AI systems, in and of themselves raise issues under section 5 of the FTC Act or any other law it enforces, but that a company may deceive consumers if it misrepresents how likely hallucinations are. Accuracy and performance claims are different: the FTC expects you to have evidence for them when you make them. In Workado (formerly Content at Scale AI), the company advertised that its AI content detector was 98.3% accurate. According to the FTC's complaint, it hadn't built or tested the model itself; it relied on results published by the model's developers, students in Norway who trained it on academic abstracts, while Workado marketed it for marketing and other plain-language text. The developers' best result on a mix of non-academic text was 74.5%, and they caught AI-generated non-academic text just 53.2% of the time. The final order, issued August 21, 2025, bars accuracy claims unless they're backed by competent and reliable evidence and requires Workado to preserve its testing protocols, test data, the steps taken to keep test and training data apart, and its statistical analyses, including confusion matrices. If you publish an accuracy number, keep the eval that produced it.

## Checklist
1. List every performance claim you make: accuracy rates, "no hallucinations", "replaces a human", benchmark ranks. Check the website, ads, pitch decks, sales scripts and answers your support team gives. Holland & Knight notes many recent FTC AI cases involve claims made to businesses, not just consumers.
2. For each claim, keep the evaluation behind it: the frozen test set (or its hash), the model and version tested, the protocol, the date and the per-item results.
3. Test on data that looks like your customers' real use. Workado's model was trained on academic abstracts but sold for marketing copy, and trained only on ChatGPT text but sold for Claude, Bard and GPT-4 output.
4. Don't borrow someone else's number. Workado relied on the model developers' published test results instead of testing its own product for the conditions it advertised.
5. Report the number that matches the claim. Overall accuracy can hide a poor detection rate for the class customers care about; keep the confusion matrix, not just the headline.
6. Re-test when anything changes. The Workado order requires evidence at the time a claim is first made and each time it's made again, so a new model, vendor or prompt means a new eval before the old number stays up.
7. Be careful with hallucination claims. Saying your product rarely or never makes things up is itself a representation about the likelihood of hallucinations, so it needs the same evidence.
8. Put limits where people will see them. In the July 2026 statement the FTC repeats that a prominent misrepresentation is unlikely to be fixed by a less prominent disclosure, and says a disclaimer buried in terms of service wouldn't be adequate.
9. Keep the evidence that cuts against you as well. The Workado order requires keeping tests and complaints that contradict, qualify or call into question a claim, for 5 years after the claim was last made.
10. If resellers or partners use your marketing materials, check those claims too; the FTC has charged vendors that supplied deceptive materials downstream. Talk to a lawyer before publishing a performance number.

## Does the FTC treat AI hallucinations as deceptive?
- The July 7, 2026 notice is a proposed policy statement on how section 5's ban on deceptive practices applies to AI companies; comments were due by July 31, 2026. Its main subject is companies that steer outputs toward objectives users don't expect or haven't been told about, including steering done to comply with a state law, which it says is likely to be deceptive.
- In a footnote, the FTC separates that from incorrect output, frequently called hallucinations, that stems not from a design decision but from the technological and resource limitations AI systems necessarily reflect. It says a company may unlawfully deceive consumers if it misrepresents the likelihood of such hallucinations, but that it does not believe hallucinations in and of themselves raise issues under section 5 or any other law it enforces.
- The statement also repeats the basics: deception requires a representation likely to mislead a consumer acting reasonably, and it must be material; implied claims count; and courts have held that companies must substantiate the claims they make. Check whether the statement has been finalized or changed since July 2026.

## Why did the FTC act against Workado's 98% accurate AI detector claim?
- From at least November 2022 until about June 21, 2024, Workado's website said its AI Content Detector was 98% accurate, then 98.3%, at telling human text from text generated by ChatGPT, GPT-4, Claude and Bard, and that it was trained on blog posts, Wikipedia, essays and more. It charged $49 a month for unlimited use from September 2023 to August 2024.
- The FTC's complaint says the detector was an open-source model, RoBERTa-academic-detector, built by students in Norway for an undergraduate thesis. It was trained on equal numbers of human-written and ChatGPT-generated research abstracts, not fine-tuned on non-academic content and not trained on output from other AI programs. Workado relied only on the developers' published tests on academic text. The developers' best accuracy on a mix of human and AI-generated non-academic text was 74.5%, and the model correctly identified AI-generated non-academic text 53.2% of the time, which the FTC called barely better than a coin toss.
- Workado settled without admitting or denying the allegations. The order is in effect for 20 years and also required it to email customers an FTC-written notice saying the FTC says it didn't have proof for its 98% claim, and to file compliance reports one year after the order and then annually for three years.

## What evidence does the FTC expect behind an AI accuracy claim?
- The order defines competent and reliable evidence as tests, analyses, research or studies conducted and evaluated in an objective manner by qualified persons, using methods generally accepted in the profession to yield accurate and reliable results.
- For any evidence it relies on, Workado must preserve: testing protocols and amendments, with the reasoning for its methods and metrics; the test data; the data's sources, distribution and evaluation classes; how test data was processed compared with training data; steps taken to avoid overlap between training and test sets; why the test data fits the expected use cases; and all statistical analyses, including preliminary ones such as confusion matrices.
- That list is a fair checklist for any AI team publishing a number, whether or not the FTC ever asks.

## What other FTC cases involve AI performance claims?
- The FTC's July 2026 statement cites its recent AI cases: Workado (AI-content detection accuracy, August 2025); TheFBAMachine (guaranteed earnings from AI-powered storefronts, July 2025); DoNotPay (claims that its chatbot could replace a human lawyer, January 2025); IntelliVision (facial recognition accuracy, January 2025); Evolv Technologies (AI weapons screening, November 2024); and a complaint against Air AI over its conversational AI's ability to replace human customer service representatives (August 2025).
- In December 2025 the FTC also reopened and set aside an earlier order against the AI writing tool Rytr, saying aggressive enforcement unsupported by facts or law isn't in the public interest, while adding that companies that deceive consumers about their AI's capabilities should be held accountable.

## Ask Rightshold (free, no key)
Add the MCP server https://www.rightshold.com/mcp to Claude, Cursor, ChatGPT, Grok or VS Code (one-click setup: https://www.rightshold.com/start?ref=seo), then ask: "Record a signed receipt for the eval run behind our published accuracy claim: a hash of the frozen test set, a hash of the per-item results, the model version, the metric and the score, so we can show later what the number was based on and when."

```json
{
  "tool": "record_ai_run",
  "arguments": {
    "model": "<model under test, dated version>",
    "provider": "<provider or eval harness>",
    "prompt_sha256": "<sha256 of the frozen test set file>",
    "output_sha256": "<sha256 of the per-item results file>",
    "params": {
      "claim": "accuracy on customer-style text",
      "metric": "accuracy",
      "score": 0.912,
      "positive_class_recall": 0.874,
      "confusion_matrix_sha256": "<sha256 of the confusion matrix>",
      "test_domain": "<what the test data represents>"
    },
    "ran_at": "2026-11-10T12:00:00Z"
  }
}
```
- Needs an API key. Only the hashes and fields you send are stored, so the test set and results stay with you. The receipt is Ed25519-signed, gets an issued_at time from Rightshold's clock and goes on Rightshold's append-only ledger, whose blocks are also anchored to Bitcoin through OpenTimestamps.
- verify_ai_run is free and needs no key: a customer, auditor or regulator can send the receipt and signature plus your results file to check it matches what was recorded.
- Record a new receipt each time you re-run the eval for a new model or release, and keep failed or worse runs too, so the record shows how the number changed over time.
- In recorded mode the receipt marks the run as reported by you. It shows which hashes and score you recorded and when, not that the eval was sound or the claim is substantiated.

_Rightshold doesn't run evals, judge whether a claim is substantiated or advise on FTC compliance. A receipt proves which hashes and fields were recorded and when Rightshold recorded them. The FTC statement described here was a proposal as of July 2026, and the cases are summarized from FTC documents; allegations in complaints aren't findings. Not legal advice; talk to a lawyer before publishing performance claims._

## Sources
- [Federal Register: FTC Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems (proposed, July 7, 2026)](https://www.federalregister.gov/documents/2026/07/07/2026-13628/policy-statement-concerning-the-suppression-of-accuracy-in-artificial-intelligence-systems)
- [FTC: Complaint, In the Matter of Workado, LLC, f/k/a Content at Scale AI, Docket No. C-4822](https://search.ftc.gov/system/files/ftc_gov/pdf/ContentatScaleAI-Complaint.pdf)
- [FTC: Decision and Order, In the Matter of Workado, LLC, Docket No. C-4822 (issued August 21, 2025)](https://www.ftc.gov/system/files/ftc_gov/pdf/ContentatScaleAI-DecisionandOrder.pdf)
- [CyberScoop: Workado settles with FTC over allegations it inflated its AI detectors' capabilities](https://cyberscoop.com/ftc-workado-settlement-ai-detector-false-and-misleading-claims/)
- [Holland & Knight: "Operation AI Comply" 2 Years Later: Continued Enforcement Against Misleading Claims (August 18, 2026)](https://www.hklaw.com/en/insights/publications/2026/08/operation-ai-comply-2-years-later-continued-enforcement)

## Related
- [How to prove a reported eval score wasn't cherry-picked from many runs](https://www.rightshold.com/answers/how-to-prove-a-reported-eval-score-wasnt-cherry-picked-from-many-runs)
- [How to make LLM eval results reproducible and tamper-evident](https://www.rightshold.com/answers/how-to-make-llm-eval-results-reproducible-and-tamper-evident)
- [How to keep an eval set out of model training data and detect contamination](https://www.rightshold.com/answers/how-to-keep-an-eval-set-out-of-model-training-data-and-detect-contamination)
- [Verified AI Runs](https://www.rightshold.com/verified-ai-runs)

_Not legal advice. Updated 2026-11-10. JSON: https://www.rightshold.com/answers/can-the-ftc-go-after-an-ai-company-for-hallucinations-or-for-its-accuracy-claims.json_