AI washing is marketing that claims, exaggerates, or implies artificial intelligence capability a product does not have. In security, it usually shows up as an "AI" label on a rules engine, a report generator, or a feature that exists only in a demo. Buyers can spot it by asking for evidence.
A definition from an actual enforcement action
The term is not just industry slang. On 18 March 2024 the SEC announced settled charges against two investment advisers, Delphia (USA) Inc. and Global Predictions Inc. According to the SEC, Delphia made claims about using AI and machine learning with client data that it could not back up, and Global Predictions made false claims such as calling itself the "first regulated AI financial advisor". The SEC called this "AI washing" and said it hurts investors.
Those cases were about investment advice. They are useful to security buyers for a narrower reason: they show that "we use AI" is a factual statement that can be false.
Common patterns in security products
- Relabeling. A long standing rules based detection is renamed as AI without any model involved.
- Demo only. A feature works in a scripted environment and has no general availability date.
- Wrapper features. A language model summarizes an existing scan result and is marketed as a detection engine.
- Unfalsifiable claims. "AI powered protection" with no stated task, metric, or limit.
- Borrowed credit. Citing a research paper or a model provider as if it described the vendor's own system.
None of these is automatically dishonest. A summary feature can be genuinely useful. The problem starts when the description claims more than the feature does.
How to check a claim
| Claim | Evidence to ask for |
|---|---|
| "Detects unknown threats with AI" | What input, what output, how evaluated, on whose data |
| "Autonomous remediation" | Which categories, what approvals, what happens on failure |
| "Reduces false positives" | Measured against what baseline, on which projects |
| "Trained on proprietary data" | What data, whether customer data is included, how it is governed |
Ask for the answers in writing. A vendor that is proud of a feature will usually be glad to describe it precisely.
Why it matters beyond marketing
If you buy on an inflated claim, you may pass it along. Your own security questionnaires, board slides, or customer commitments might repeat what the vendor told you. Keep a record of what you were told and what you verified.
Try it yourself
Click through it at your own pace, or open it full screen.
Watch how a noise reduction claim is shown with a before and after count. A claim you can see counted is easier to trust than one you are asked to believe.
Where Safeguard fits
Safeguard lists generally available capabilities and keeps early access items separate. You can check the SBOM and dependency scanning on your own project with the free scan at gold.safeguard.sh/scan.
Quick check
Is using a language model in a product bad?
No. The issue is claiming more than the feature does, not using a model.
Who defines AI washing?
There is no single definition. The SEC used the term in its March 2024 action against two advisers.
Sources
- SEC press release 2024-36, 18 March 2024: https://www.sec.gov/newsroom/press-releases/2024-36
Last reviewed: 19 October 2026.
