Experiments involving 1,401 people found that biased AI feedback could make human judgments more biased too.
A customer sends a short message asking when a refund will arrive. An AI tool labels it angry. If a support worker sees that kind of label often enough, they may begin to read brief messages as hostile before considering what the customer actually said. That is an example of how repeated AI errors could influence human judgment; it is not a customer-service finding from the study discussed here. The research used controlled tasks, including judgments about images. Its concern is relevant beyond a single wrong answer: people may learn a pattern from a tool and carry that pattern into their own decisions.
This complicates the common reassurance that a human will check the AI. A reviewer is not automatically independent of the system they use all day. If its judgments gradually shape what the reviewer expects to see, a person approving the result may repeat the same error rather than catch it. Reviewing the original material matters as much as having someone click approve.
Moshe Glickman and Tali Sharot ran a set of tests. In all, 1,401 people took part. The team studied how AI feedback shaped people’s views. Tasks included judging faces and other images. When people interacted with biased AI, their own judgments became more biased. Some did not grasp how much the AI had shaped them. The researchers also found that accurate AI could improve judgments. So the lesson is not that all machine advice is harmful. The direction of the advice matters.
The paper appeared online in late 2024 and in the 2025 volume of Nature Human Behaviour. These were set tasks. They did not measure how much unfair bias people pick up at work. Read the original study for the designs and limits.
Here is an invented example. A tool sorts customer messages by how angry they sound. It keeps treating brief messages as rude. A worker sees those labels all day. Over time, the worker might begin to expect hostility from any short message. Yet a customer could be writing on a phone, using a second language, or making a brief factual request. The message length is observable; the judgment about anger needs more evidence.
The customer-service scenario is an illustration, not a result from the image tasks. It shows the distinction worth examining in a workplace: whether staff check a label against the source, or gradually begin to interpret the source through the label. A review that looks only for dramatic failures may miss a consistent small error. If brief messages are repeatedly treated as hostile, the pattern matters even when no single classification seems serious enough to investigate.
Read the message before reading the tool's judgment, when you can. Write what you can actually point to. The customer asked twice for a refund. The customer used a short sentence. The customer named a missed delivery. Then look at the label. Does it add a useful fact? Or does it turn a guess about mood into a claim about the person? A note such as “package arrived late” identifies a problem the team can address. “Difficult customer” adds a judgment about the person that may persist after the delivery issue is resolved. Prefer the specific observation unless there is a clear reason for the broader label.
This also helps when people disagree. You can compare the words on the page. It is harder to settle a dispute about who seems to be a pain.
One accurate answer tells you little about a pattern. Make a small set of examples where you know what a fair answer should consider. Use short and long messages. Add a serious complaint with polite words. Add a simple request with blunt words. Ask whether the tool treats the same facts in the same way. Change a name or a style of writing. If the label changes too, look closer. Do not assume every difference proves bias; the wording may change the meaning. Keep the comparison as fair as you can.
Real choices about people need more care. Use trained people to check such tools. A quick home test cannot prove a tool is fit to hire staff or offer loans. It may still show why a label is wrong.
Rejecting every recommendation would not produce independent judgment either. Compare the recommendation with the evidence and record a reason for accepting or rejecting it. That gives the next reviewer something more useful than an unexplained approval. Instead, keep a reason for accepting or rejecting the advice. The label says angry, but the message only asks for a date. Or the label says low priority, but the note says the service has stopped working. That reason is useful to the next person too. It leaves a trail that can be checked. An unexplained click on approve leaves almost nothing.
When a tool keeps making the same mistake, report the pattern. Fixing each case helps one customer, but if you never report the pattern, the next person may face it too.
A review process should preserve access to the original material and make disagreements possible. For a support message, that means reading the customer's words and recording the actual issue, rather than inheriting a label about their character. It also means collecting repeated errors so the same mistake is not quietly passed from one worker to the next. The experiments do not tell us how much bias any particular workplace tool creates. They do show why human review needs to be designed carefully. A person who sees only the tool's interpretation has less opportunity to question it than someone who can examine the evidence for themselves.
