AI washing means making false or inflated claims about how a product uses AI. The gap may be in what the system does, the data it uses, or the results it claims.
Here, the term refers to sales claims about AI. It does not mean the AI features of a washing machine.
Two SEC orders from March 18, 2024 show why the exact claim matters. Both concern gaps between AI claims and the tools in use.
What is AI washing, and what does the term mean?
The SEC’s September 4, 2024 explanation warns against saying a firm uses AI when it does not, or uses it in a way it does not. That is a useful starting point. It is not a test that proves how a given tool works.
An AI washing definition should keep the claim apart from a model error. A tool may use AI and still get things wrong. Our AI hallucination cases cover that kind of error.
Human help does not, on its own, make a claim false. A team can state which tasks people do and which a model does. Check whether those words match the work.
Missing proof needs its own label. If a buyer cannot check a claim, mark it unverified. That does not prove the seller lied.
Two SEC cases: read the findings separately
These AI washing examples come from settled administrative orders, not criminal convictions. Both firms settled without admitting or denying the findings, except the SEC’s jurisdiction.
The findings cover the conduct in those orders. They are not tests of either firm’s current products.
Delphia: the claimed data input was missing
The Delphia order, Release 6573, describes claims made from at least August 2019 through August 2023. Delphia said it used AI and machine learning to study retail clients’ spending and social-media data.
It said those data helped shape its advice. The SEC found that no such data was used in its investing process.
The distinction is in paragraphs 3 and 4. Delphia had built algorithms to manage portfolios.
They used a range of sources. It planned to use client data as an input.
The order says that plan did not become the claimed feature. The firm did collect some client data at times.
It did not feed those data into its investing algorithms. Having algorithms did not prove the stated use of those data.
The order also lists efforts to fix claims after an SEC exam in 2021. Some claims changed.
Others stayed false or misleading through August 2023. A fix in one document did not fix the other claims.
Section IV.C imposed a $225,000 civil penalty. The lesson is to check the claimed data path. Do not infer that all of the firm’s software was made up.
Global Predictions: a chatbot did not generate the advice
The Global Predictions order, Release 6574, describes its PortfolioPilot platform. Paragraph 3 says algorithms made the advice on how to allocate investments.
A chatbot passed on that advice. It did not create it.
Paragraph 4 addresses a claim that the tool had “expert AI-driven forecasts,” which the SEC found it did not. It also finds fault with the claim to be the “first regulated AI financial advisor.” Keep those claims apart from the fact that a chatbot was there.
Paragraph 7 covers a claim to beat IMF forecasts by 34%. The order says the claim left out what the figure meant and when the test was run.
It was a relative error rate. It was not a 34% return on an investment.
The test was from December 2021. The claims stayed public for about two years.
The order says Global Predictions could not provide records to back the 34% claim. Do not read that figure as a proven result here.
Section IV.C imposed a $175,000 civil penalty. Together, the two orders imposed $400,000 in civil penalties.
That sum is not money investors lost. It is not a total for AI washing across the industry.
The orders include other findings beyond those shown here. They do not support an offshore-labor story. Nor do they show that no AI existed anywhere in either firm.
Separate folders keep two case records apart.
Three different claims need different evidence
Start with the words the buyer is asked to trust. “Uses AI,” “uses your records,” and “improves results” raise distinct questions. One answer cannot settle all three.
| Claim | Evidence to request | What remains open |
|---|---|---|
| An AI part exists | Named part, version, and role in the tool | Whether it does the stated task |
| It uses named data or does a task | A clear account of inputs, steps, outputs, and human work | Whether the task yields the claimed gain |
| It yields a result | Test method, baseline, sample, dates, and limits | Whether the result holds for your use |
A visible chat box shows an interface. It does not show what makes the answer behind it.
The Global Predictions order shows why this matters. You need not guess how unseen software works.
The same care applies to data claims. A tool may collect records.
That alone is not proof it uses them to make the stated output. Ask where each input goes and which step uses it.
For a result claim, read the test scope before you repeat the score. Our guide to accuracy-study limits shows why you need to know how a test was run and on what data. A test on one task does not prove all the gains claimed for a tool.
Keys reveal only part of the piano’s mechanism.
Questions to ask a supplier
Choose one dated statement from the page or proposal you are reviewing. Keep its exact words.
A broad label such as “smart” loses that detail. Ask the supplier to identify the system and version behind it.
Then ask what proof backs that claim. Map the part, provider, data path, and human role, as in an AI use and data-flow inventory. Do not share private client data in a casual request for proof.
Here is a fictional review record, not an SEC form or a product test we ran:
| Field | Example entry |
|---|---|
| Exact claim | “Uses uploaded reports to draft answers” |
| Scope | Seller’s trial version; made-up reports only |
| Proof received | Demo shown; data path not explained |
| Status | Unverified for the claimed use of reports |
| Owner and next check | Buyer asks the seller to explain and test that input path |
For this fictional record, ask to see which passages from the made-up report support the answer. Change one passage and repeat the same prompt. Keep both outputs and the version used.
That proposed check can test the visible task. It cannot, on its own, prove how every internal step works or how the tool will perform on all reports. Do not treat it as a test completed for this article.
The demo alone does not close the question. A later check may support a claim, show a gap, or leave it open. Write down what you saw and what you still do not know.
For a broader view, use product claims and evidence. Keep this AI check focused on the claim at hand.
One stone receives a close inspection.
Limits of the check
A buyer may lack access to code, logs, or private test records. That limits what they can know.
A slick demo is not a full view of a system. Nor is an unanswered request.
These two 2024 orders also do not describe all current SEC AI regulation. Their findings and terms have a set scope. They are not a legal checklist for every software company or a judgment about today’s products.
The SEC video below adds context from 2024. Gary Gensler was then the agency’s chair. Use the signed orders for the case findings.
Before you repeat an AI claim, note its words, proof, date, and limits. If a key fact is missing, keep it unverified.
Name the next check. That is a review step, not a shield against all false claims.