A model has no opinions, but it has habits. They come from what it read, and they show up in its answers with regularity. On uses that touch people, that can have very concrete consequences.
Definition
Algorithmic bias is a systematic distortion in the results produced by an automated system, which regularly disadvantages certain situations or groups.
The important word is systematic. A one-off error is not a bias. A bias repeats, in the same direction, and therefore becomes predictable once spotted.
The origin is rarely intent. A model learns from texts produced by people, which carry their own imbalances. It does not invent them, it reproduces them, sometimes amplifying them because it averages them.
Where it shows up in a small business
| Use | Possible bias |
|---|---|
| Sorting applications | Favouring certain backgrounds or phrasings |
| Sales writing | Repeating stereotypes by target sector |
| Analysing customer reviews | Misreading regional or informal turns of phrase |
| Translation | Assigning a default gender to occupations |
| Prioritising requests | Underrating badly written messages |
The first line is the most sensitive: automated sorting of job applications is regulated in Europe, and a fully automated decision producing a significant effect on a person raises a legal problem on top of the ethical one.
What you can concretely do
Test on deliberately varied cases. Run the same content through while changing an irrelevant element, a first name, a region, a phrasing. If the result changes, you have found a bias.
Keep a person in the loop on anything concerning individuals. See Human in the loop.
Ask for the reasoning. A model asked to explain its ranking makes bias visible far faster than a score alone.
Do not hand the final decision to a model when it affects someone. Sorting can be assisted, the decision must stay human.
The most dangerous bias is the one that suits you. A model systematically setting aside the most complex requests looks efficient, until the day you notice it was setting aside your most profitable customers.
Frequently asked questions
Can bias be removed entirely?
No, and be wary of tools claiming otherwise. It can be measured, reduced and compensated for, but a system trained on human data carries the mark of that data.
Is a newer model less biased?
On the most visible biases, yes, providers work on it actively. On fine-grained biases specific to a trade or a language, improvement is far slower and does not spare you from testing in your own context.
Does this affect RAG too?
Yes, twice over: the model brings its biases and your document base brings its own. A RAG connected to unbalanced documentation will produce unbalanced answers with the confidence of a citation.
How do you set up simple checks?
By building an Evaluation you replay whenever the model or the instruction changes. Our Claude Cowork course covers that verification practice alongside going live.