Why the label tells you little
“AI-powered” has become a marketing phrase. Sometimes it describes a sophisticated system trained on large amounts of data. Sometimes it describes a few ordinary rules in a spreadsheet, or a simple statistical formula that existed long before the label became fashionable. The label alone does not say which one you are looking at, or whether the product works well.
That does not mean the claims are false. It means they deserve the same practical scrutiny as any other claim about a product. The questions below do not require technical training.
Question 1: What exactly does it do?
Look for a specific task. “Sorts incoming support emails into five categories” is checkable. “Transforms your business with intelligent insights” is not. Specific claims can be tested, and vague ones are hard to hold anyone to.
Also ask where the AI sits in the process. A tool that drafts suggestions for a human to approve is different from one that makes final decisions on its own.
Question 2: Is machine learning actually involved?
Not everything described as AI learns from data. A fixed set of if-then rules can be useful, but it is not machine learning in the usual sense. If a company says its system “learns,” a fair follow-up is: from what data, and does it keep learning after release?
Question 3: What was it trained and tested on?
A model is shaped by its training data, and the match between that data and your situation matters. A tool trained on one country’s documents, one kind of customer, or one hospital’s patients may perform differently elsewhere. Good evidence usually says who and what was tested, and whether the test examples were kept separate from the training examples. Results measured on the same data used for training can look much better than real-world performance, an effect known as model overfitting.
Question 4: How is success measured?
Be careful with a single headline number.
- Accuracy can mislead. Suppose, hypothetically, 1 percent of transactions are fraudulent. A system that labels every transaction as legitimate is 99 percent accurate and catches nothing.
- Different mistakes cost different amounts. A false alarm and a missed problem are rarely equally harmful.
- Context matters. A figure from a laboratory test or a curated demo may not carry over to messy, real conditions.
- Compared with what? A fair comparison is against the existing method or a simple baseline, not against nothing.
An honest vendor can usually explain the measure, the test set and the comparison without hand-waving.
Question 5: Who checked it?
Claims tested by the company itself are weaker evidence than results from independent evaluation or published, peer-reviewed work. Even then, check that the evidence is about this product and this use, not about a related technology. In areas like medical AI, a tool may have been tested only on limited groups or in narrow settings, so approval or evidence for one use does not extend to another.
Question 6: What happens when it is wrong?
Every system makes errors. Ask how mistakes are found, who is responsible, and whether a person can review or overturn an automated decision. This matters most when outcomes affect jobs, money, health or safety. It is also where problems like algorithmic bias can show up, so it is reasonable to ask whether results were checked across different groups.
Question 7: What about the data and privacy?
Ask what information the tool collects, where it goes, how long it is kept, and whether it is used to train future versions. The answer is often in the privacy policy, and a missing or evasive answer is informative in itself.
Warning signs
- Claims of perfect or near-perfect accuracy without any description of testing.
- Words like “human-level,” “revolutionary” or “unbiased” with no explanation.
- Only demos or testimonials as evidence.
- Refusal to say what data was used or how success was measured.
- Promises that the system will replace entire jobs or make guaranteed predictions about people, markets or health.
- Pressure to decide quickly or to avoid independent checks.
Reading news stories the same way
The same questions work for headlines. “AI beats doctors” or “AI predicts crime” should prompt you to look for the study design: how many cases, what the comparison was, and whether the finding was repeated elsewhere. Reports of generative AI tools that sound impressive are also worth checking against their known weak spots, such as confidently stated errors.
A simple habit
You do not need to settle every question to be a careful buyer or reader. Picking two or three that matter for your situation, and noticing whether the answers are specific, is usually enough to separate a well-founded claim from an empty one. Believable claims tend to be narrow, measured and open about their limits.