Is AI washing only about products containing no AI?

No. Exaggerating the capabilities or benefits of a real AI component can also be misleading.

Does human review mean a product is fake?

No. Human involvement can be appropriate when it is described accurately.

Can small businesses assess claims?

Yes. Define the promised result and test representative work, including difficult cases.

AI Glossary / AI and machine learning / Beginner

AI washing

Definition

AI washing is the misleading presentation of a product, service or organization as using AI, or as benefiting from AI, beyond what the available evidence supports. It can involve overstating capabilities, disguising ordinary features or claiming a system uses AI when it does not.

What it means

An AI label can attract attention, investment and a higher price. The label becomes misleading when the impression it creates is materially different from the product's actual operation or demonstrated performance.

Not every simple algorithm is evidence of AI washing. AI has a broad history, including methods that are less fashionable than large language models. A product can also use a legitimate AI component without every feature being AI-driven. The question is whether the specific claims accurately describe what buyers receive.

For example, “suggests possible categories for human review” and “autonomously manages every classification decision” imply different capabilities. Honest communication identifies the feature, its limits and the role of people. It does not rely on the audience treating the word AI as a guarantee of quality.

How it works

AI washing often begins with a gap between technical evidence and commercial wording. A prototype result may become a general promise. A feature planned for a later release may be described as already available. Human review may disappear from a diagram even though the service depends on it.

Detection starts by making a claim testable. Identify the task, inputs, output, level of automation and conditions under which it works. Ask for a demonstration with representative examples and compare the result with the claimed benefit.

Performance evidence should specify the baseline. “Twice as fast” means little without knowing the earlier process, workload and quality standard. Evidence about a narrow pilot should remain described as a narrow pilot. The aim is accurate purchasing and communication, not deciding that only one fashionable model architecture deserves the AI label.

Three practical examples

A supplier advertises fully automated document processing, but employees manually complete most documents behind the scenes without that dependence being disclosed.
A website describes an unreleased prediction feature as part of the current product, creating an inaccurate impression of what customers can use.
A sales presentation attributes a general workflow speed improvement entirely to AI even though the measured change also removed several manual approval steps.

Think of it this way

Think of a food label that highlights an ingredient. The label should help a shopper understand the product, rather than imply that a trace ingredient changes everything about it.

The analogy has limits because there is no single universal percentage of AI that defines an honest product claim. The useful comparison is between the overall impression and the evidence behind it.

Worked example

Checking an automation claim

Suppose an invented vendor says its service “automatically completes 95% of invoices.” A buyer submits a representative set of 200 invoices. The service produces 190 draft records, but only 130 require no corrections or manual field entry.

If “completes” means ready for the next agreed processing step without intervention, the observed rate is 130 ÷ 200, or 65%. The 95% figure instead describes draft production. Both counts can be real while the marketing wording misleads.

The buyer should agree on what completion means, record the errors and ask whether the sample reflects intended use. This single test does not establish every aspect of a supplier's conduct. It does show why a defined outcome matters more than an impressive percentage.

Illustrative case study

A fictional scheduling-software business adds a model that proposes meeting titles. Its marketing draft calls the whole product an autonomous AI scheduling assistant. The product team points out that users still choose attendees, resolve conflicts and confirm every booking.

The company revises the description to explain the actual title-suggestion feature and demonstrates it with ordinary customer inputs. It keeps separate evidence for existing scheduling automation and the new model-based feature.

The case shows that correcting AI washing need not mean removing AI from the product. It means aligning the promise with the function customers can actually use.

Common misconceptions and mistakes

Using a simple model automatically makes a claim deceptive. Complexity is not the test; accuracy and substantiation of the claim matter.
A successful demonstration supports every promised use case. Carefully selected examples can omit important failure conditions.
Buying an AI component validates the whole product's claims. The surrounding application and the way it is presented still need evidence.

Business use

Buyers can reduce confusion by requesting a feature-level description, representative evaluation and clear account of human involvement. Sellers can maintain a small evidence record linking major claims to product versions and test results.

For website content, prefer concrete wording about what the tool does. A supported statement about summarizing a particular document type is more informative than broad promises about intelligence or transformation. Regulatory treatment depends on the jurisdiction and facts; the underlying communication discipline is to avoid unsupported impressions.

Check your understanding

A service generates a draft for 95% of documents, but only 65% meet the agreed completion standard without intervention. Which number supports an automatic-completion claim?

Answer: The 65% figure matches that standard. Draft generation and completed work are different outcomes.

Related terms

Automation
Benchmarks
Human-in-the-loop
AI hype
benchmark
proof of concept
model evaluation

Learn next

Read benchmark and proof of concept. Take one AI claim from a proposal and rewrite it as a measurable task with a clear baseline and acceptance standard.

Sources

US Securities and Exchange Commission, SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence: https://www.sec.gov/newsroom/press-releases/2024-36. This is a historical primary-source example of enforcement concerning AI representations, not a statement of law for every jurisdiction. Scenarios and calculations above are fictional.

Previous: AI slop Next: AI watermarking