Under a YouTube walkthrough of Meta's Muse agent, the second-most-liked comment (29 likes) lands on a doubt a lot of demo-watchers share. The commenter watched the full video, admired the features, and finished with "I can't think of a single use for this myself," because most of what these tools do looks "pretty easily solvable without AI." The doubt has a name, AI washing. Below: what the term covers, the three questions that test any AI feature, and how to tell an honest label from a decorative one.
What AI washing is
AI washing is the practice of putting an AI label on a product whose work is done by something simpler: templates, if-this-then-that rules, or a lookup table. It also covers the softer version, where AI is involved but contributes less than the label suggests.
The term comes from finance. On March 18, 2024, the SEC charged two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., with making false and misleading statements about their use of AI; both settled, with $400,000 in combined civil penalties (SEC, 2024). In a September 2024 SEC video (SEC, 2024), then-chair Gary Gensler restated the underlying rule: advisers "should not mislead the public by saying they're using AI when they're not."
The pattern shows up outside finance too. In June 2025, Gartner estimated that of the thousands of vendors calling their products agentic AI, only about 130 had the autonomy to support the label, and the firm used "agent washing" for the rest: assistants, chatbots, and script automations rebranded. Consumer app features sit one level down. The label describes the ambition, and the code inside sometimes does something blander.
Two reasons the label spread so fast: renting a model got cheap, since an app can call a hosted model per use with no training of its own, and the label draws attention, users, and investor interest. Neither reason guarantees that a given feature needed a model. For anyone buying, the label is a claim to check, like any other claim on a pricing page.
The three-question test
The test is a screening pass, not proof of dishonesty, and it runs on any AI feature, in an app you use or one you're building:
- Name the job. One sentence, concrete: what does the feature produce? "It suggests a reply." "It sorts my expenses into categories." "It writes alt text."
- Ask whether written rules could do it. If you could write the steps on paper ("if the message contains a tracking number, suggest 'where is my package?'"), a rule will do the job with more predictable results and lower cost. If the job needs judgment on messy input (a question phrased five different ways, a paragraph in any order, a photo taken at any angle), rules get brittle and expensive to maintain, and a model is worth testing.
- Decide what a wrong answer costs. A model predicts, and prediction means occasional wrong guesses. A wrong guess in a low-stakes place, a suggested tag you delete, costs a click. In a high-stakes place, a sent email or a paid invoice, wrong guesses cost more, and the feature needs checking machinery no matter what its label says.
The fastest check is to turn the AI off. Some features have a toggle; for the rest, picture the feature rebuilt as plain rules. If the output quality wouldn't move, the label was decoration.
Where the label earns its keep
A washing test should calibrate judgment, not breed blanket suspicion. Plenty of features pass question 2 with a "no": drafting and rewriting text, summarizing a long document, translating, reading a messy spreadsheet, tagging a photo taken in bad light. Writing and maintaining rules for those jobs gets brittle fast, and models handle them well enough to change what an app can do.
We run both sides of that line in this site's pipeline, and the choice happens per job. When we built the comment-sorting step, we kept it simple: a keyword-matching script, no model call, so no API cost. Nightly it groups the day's harvested comments into themes; tonight's run sorted 4,600 comments into 13 themes in 0.64 seconds (one run, on the machine that publishes this site). The drafting and review steps use models, because writing and judging writing need judgment on messy input, which is the case in question 2 where rules get brittle. Same pipeline, two answers to the same question: does this job need a model?
What to do with the answer
What to do next depends on where the feature is:
- The feature is in an app you pay for. Use the test as a pricing check. A feature built on rules doesn't justify an AI premium, and a model-backed feature whose wrong answers cost little is usually fine to leave on. Look for a toggle; comparing the feature with it on and off answers faster than any review.
- You're building something. Start with rules where they pass the test and models where they don't. The rules version is cheaper to run and easier to debug; the model version should justify the difference in accuracy, cost, and error handling.
- A whole demo reads as features you'd never use. If you can't name a use after the full walkthrough, you may have learned something useful: the product doesn't fit your week. The Muse commenter reached that ending after giving the demo its chance, and the test turns a vague doubt into a decision with reasons attached.
When the label is honest and the job needs a model, use the feature and skip the cynicism.
Frequently asked questions
What does AI washing mean?
It's a claim problem: a product or feature carries an AI label while simpler machinery (templates, rules, a lookup table) does the work, or the AI inside contributes less than the label suggests. The SEC's March 2024 enforcement actions brought the term into wide use in finance, and the same pattern shows up in everyday app features.
Is AI washing illegal?
In finance it can be: the SEC's March 18, 2024 actions charged two investment advisers $400,000 combined for false statements about their use of AI, and SEC guidance warns that AI washing "may violate the securities laws." For consumer apps, the same claim is mostly a marketing and trust problem rather than a securities case.
What's the difference between AI washing and agent washing?
Agent washing is Gartner's narrower term for products rebranded as "agentic AI" without the autonomy to match. AI washing is the broader pattern: any AI label on work that simpler machinery does. A washed agent is one variety of AI washing with a specific claim attached.
How do you test whether an AI feature needs AI?
Run the three questions: name the job, ask whether written rules could produce the same output, and price a wrong answer. A reply suggestion driven by "if the message contains a tracking number" passes for a rule; a feature that summarizes a document in any format needs a model. Treat the result as a screening pass, not proof, and let the stakes from question 3 decide how much checking the feature needs.
Want to go further?
Try the test on one feature you already pay for: name its job, picture it as written rules, and decide what a wrong answer costs. Two posts from this Lab cover the neighboring skills: What Is Agentic AI? The Next Step Beyond Chatbots explains agent washing and the questions that sort the "agentic" label, and How to Verify AI Output: A Simple Framework covers the checking habit for the features that pass the test.