Content that deceives and outputs that discriminate

The marketing manager at a fitness equipment company is building the launch page for a smart rowing machine. The agency left space for four customer testimonials, with name, city and stars. The product has not reached anyone yet. She asks Claude for four realistic testimonials so the page does not look empty.

Claude can write them in seconds. Being able to write them does not make them honest.

Two ethical questions for every use

The ethical implications of using AI at work almost always come down to two questions:

1. Does it deceive whoever receives it?
2. Does it treat some group of people worse?

And one rule that applies to both: responsibility stays with a person. Human accountability is not delegated to the tool.

Content that poses as real

A testimonial with a name and city tells the reader: “a real person bought this and thinks this.” If that person does not exist, it is deceptive content, even if the text is modest and believable. Making it “more modest,” using only common names or adding a “results may vary” note does not change the fact that the customer is invented.

The same goes for fake reviews, profiles of “experts” who do not exist or letters signed by someone who did not write them.

The ethical answer is not to leave the page empty. It is to use something that does not pose as what it is not:

• wait for real opinions from the first customers;
• describe the product and its features;
• show results from your own tests, presented as such.

Outputs that treat a group worse

A logistics company uses a Project where managers paste their notes and get a draft performance review, which they then sign. Halfway through the cycle, someone in HR compares 60 summaries: with similar ratings, the ones for part-time staff use more negative words (“limited availability,” “less committed”), even though the notes only mention their hours. Several managers signed after a quick read because it “sounded professional.” The reviews set bonuses next month.

That is discriminatory treatment: a bias that is already harming real people. The responsible course is:

1. Pause the use until the pattern is understood, even if the deadline is tight.
2. Review with the right people (HR, in this case) the reviews already done, not only the remaining ones.
3. Make it clear that whoever signs answers for the final text. Signing because it “sounds professional” is part of the problem.

What is not enough

Proposal	Why it does not solve it
Adding “be fair and unbiased” to the instructions	It is an intention, not a control; it does not fix what was already signed
Asking Claude to rate each summary for bias	It is the same tool evaluating itself
Removing the hours from the notes and carrying on	It hides the symptom, does not review what was done and does not set who answers
Carrying on because the deadline is close	Real people’s bonuses depend on those texts

The typical traps

• “Modest” testimonials or ones with a footnote, as if that made them honest.
• Fixing a bias with a sentence in the instructions.
• Using Claude as the judge of its own bias.
• Correcting going forward without reviewing what already went out.
• Treating the manager’s signature as a formality.

On the exam

If the scenario asks for invented people, opinions or signatures that look real, the answer is not to publish them and to use content that is real or does not pose as real. If it shows a pattern that harms a group, look for the option that pauses, reviews with the responsible department and keeps responsibility with people. Rule out the statements of good intentions, Claude’s self-assessment and “carry on because of the deadline.”