# AI Washing: How to Spot Products That Have AI Only in the Name

Author: Marin Niehues, Niehues Consulting · Published: 2024-04-22 · AI Strategy & Business Case
Source: https://www.niehuesconsulting.de/en/blog/ai-washing-spot-products-claiming-ai/

You can spot AI washing when a vendor promises AI but names neither the model nor the data nor any measured results. A product with genuine AI can be described: there is a task the model solves, data it works with, a metric by which its quality is measured, and a test on your own data. If one of these is missing, you are looking at a promise that you should have proven before you buy.

Since ChatGPT appeared in November 2022, product pages, presentations and proposals have used the term AI frequently in their titles. This article shows how decision-makers can test vendor claims, which warning signs to look for and how to build a pilot that proves the promise with your own data. This article provides an overview and does not constitute legal advice.

**An AI claim holds once the vendor names the task, the data basis and the measured result, and the product passes on your data. The product name proves nothing.**

## What is AI washing?

AI washing is marketing that attributes artificial intelligence to products or services although they use none or only an insignificant part. The term is modelled on greenwashing. The pattern is older than the current wave: in 1770 Wolfgang von Kempelen presented the "Mechanical Turk", a supposed chess automaton in which a hidden person made the moves. Its modern counterpart is a product whose results are produced by staff by hand in the background, or whose "intelligent" function consists of fixed rules.

The problem reaches beyond an annoyance in procurement. In March 2024 the US Securities and Exchange Commission announced that two investment advisers, Delphia and Global Predictions, had agreed in a settlement to civil penalties of 225,000 and 175,000 US dollars over false and misleading statements about their use of AI. The US Federal Trade Commission had already pointed out to businesses in February 2023 that claims about AI must be provable. Regulators treat inflated AI promises like any other advertising claim: they have to be true. For buyers a practical conclusion follows. Anyone who buys a product for its AI function should have that function proven before signing.

## Why is AI washing so widespread?

AI washing is widespread because the term has no sharp boundary, buyers can hardly assess the technology from outside, and pressure for results inside the company creates demand for "something with AI". Everything from a simple rule set through statistical methods to a large language model is called AI, and a label costs the vendor less than developing a model. How that pressure arises is described in the article on [the pressure to adopt AI](/en/blog/ai-fomo-companies-investing-wrongly/). How expectations develop over time is shown in the article on the [hype cycle](/en/blog/ai-hype-cycle-progress-versus-announcement/).

## How do you recognise products that have AI only in the name?

You recognise them by four patterns: the vendor stays vague about the technology, names no measured results, refuses a test with your data, or shows a demonstration that does not transfer to your case. Statements that are untypical for learning systems add to this.

**Warning signs in the vendor conversation**
Warning sign | What may lie behind it | Follow-up question
"Our AI does this automatically" with no detail | Fixed rule set or manual rework | Which task does the model solve, and what happens when it is uncertain?
No statement on model and data | Bought-in standard component without own development | Which model runs, how was it adapted, where is the data held?
Accuracy "above 95 per cent" without a measurement basis | Measured on easy examples or on training data | On which data and with which metric was it measured?
Demonstration with prepared examples | Examples were selected | Can we test for an hour with our own data?
No route for error cases | Human checking in the background | What percentage of cases do people check, and who are they?
Price rises with the label AI | Mark-up for a label | What does the same function cost without the AI component?

The link to measured quality deserves a closer look. A model can only be judged when you know what its quality is measured by. For a classification these are, for instance, the share of correct assignments and the share of missed cases. For a language model it is the rating of a sample of answers by experts from the business area. A single percentage without a measurement basis says little, because it varies with how easy or hard the test cases were. The article on [checking AI success stories](/en/blog/ai-case-studies-check/) shows the same approach for case studies.

## Does a product have to contain machine learning to be useful?

No, a useful product does not have to contain machine learning, and a rule set can be the better solution when the task is stable and the rules are known. What counts is the honesty of the label and the benefit for your process. A rule set is traceable, cheap to run and behaves predictably. A learning model has advantages where rules cannot be written down, for example with free text, images or many exceptions.

For your decision the question is therefore which technology suits the task. A vendor who says plainly "This is a rule set with text recognition in front of it" is a better partner than one who talks vaguely about AI. Anyone who argues like that can be tested, and that is the basis of a robust business case.

## How do you test an AI claim with a pilot?

You test an AI claim with a pilot on your own data, with a metric set in advance and with criteria that decide between continuing and stopping. The pilot is an effective protection, because it moves the burden of proof to the vendor.

1. Define the task: which specific process should improve, and what is the value today?
2. Provide test data: a sample from real operations, including the difficult cases, with known correct results that the vendor does not see beforehand.
3. Set the metric and threshold: at what result is the pilot passed, for example hit rate, time saved or share of cases without rework?
4. Run it: the vendor processes the test data, and experts from the business area rate the results.
5. Decide: if the product passes, introduction follows. If it fails, the pilot ends without creating an obligation.

The effort is manageable, because after preparation the measurement is quick. The article on [ChatGPT in the company](/en/blog/chatgpt-in-business-sober-review/) shows how the pattern carries over to language models.

**Ten questions for vendors making AI claims**
Question | Good answer | Warning sign
Which task does the AI solve? | Concrete process with input and output | General description
Which technology is behind it? | Model type and origin are named | Retreat to trade secrets
What was the model trained or adapted with? | Type and volume of data | No statement
How is quality measured? | Metric, test data, result | One percentage without a basis
How does the system behave on errors? | Documented route with human review | It makes no errors
Which data leaves our company? | Clear information on storage location and processing | Unclear answer
Can we test with our own data? | Yes, before the contract | Demonstration only
What happens without the AI component? | Scope of function is described | The product fails
Are there references with a comparable task? | Contact person and measured values | Logos only
What does exit look like? | Data export, termination route | No answer

## Spotting AI washing with Niehues Consulting

[Niehues Consulting](/en/) is a consultancy for AI management and organisation design from a delivery perspective, based in Düsseldorf. In the service area of AI management, the work includes business analysis and requirements, the business case and the selection and sequencing of applications, as well as testing vendors against the requirements of operations. Founder Marin Niehues brings AI into operations, in projects ranging from VC-funded scale-ups to DAX 40 corporations and in more than 20 countries on six continents.

The approach follows the Niehues Delivery Model. Three review questions matter when procuring AI products: which business outcome should change measurably? Does the prototype prove the benefit in the real process with real data? Is the business case confirmed with real values? The scope and terms of an engagement are discussed in the free initial consultation, which you can arrange through the [contact form](/en/#kontakt).

**Conclusion**
AI washing ends where the task, the data basis and the measured result are on the table. A test with your own data, a metric set in advance and ten questions are enough to separate promises from performance. An honest rule set is better than an unproven AI.

**Sources**
- US Securities and Exchange Commission: press release 2024-36 on proceedings against Delphia (USA) Inc. and Global Predictions Inc., March 2024
- US Federal Trade Commission: Keep your AI claims in check, business guidance, February 2023

## Frequently asked questions

### What is AI washing?

AI washing is marketing that attributes artificial intelligence to products that use none or only a little. The term is modelled on greenwashing. Buyers spot it when task, data basis and measured results are missing.

### How can I tell whether a product really contains AI?

Ask about the task the model solves, the data it works with and how its quality is measured. A product with genuine AI can answer these questions concretely. A test on your data before the contract provides the proof.

### Is a rules engine not AI?

A fixed rule set is generally not counted as machine learning. It can still be the right solution when task and rules are stable. It becomes a problem only when it is presented as a learning system.

### What questions should I ask an AI vendor?

The ten questions in this article cover the main points: task, technology, training data, measurement, error behaviour, data flow, a test on your own data, scope without AI, references and exit. Write the answers down so that they can be compared.

### How much does a test with your own data cost?

The effort consists mainly of preparing a sample with known results and the time of the experts who rate the output. Serious vendors provide a test before the contract. The scope and terms of an engagement are discussed in the free initial consultation.

### Can AI washing have legal consequences?

Yes, inflated advertising claims can have legal consequences, as the US Securities and Exchange Commission's proceedings of March 2024 show. For buyers, proof in the contract and in the pilot comes first. The legal assessment of an individual case is a matter for specialist lawyers.

### Is a product with human review in the background dishonest?

No, human review in the background is often right and even desirable. It becomes dishonest when the vendor conceals it and presents the result as fully automatic. Ask about the share of cases reviewed and about data protection implications.

### How do I tell genuine generative AI from an interface on top of someone else's model?

Ask which model runs and what the vendor itself contributes, such as data integration, adaptation and quality assurance. A product built on a third-party model can be valuable if the ingredients are named openly. Price and own contribution should fit together.