fairness

in English: AI fairness

basicspreading

This is the requirement that artificial intelligence decisions must not place people at an unjustified disadvantage due to their gender, age, origin, or membership in other groups.

In plain terms

Imagine a teacher who grades papers anonymously so that personal preferences do not influence the outcome. Fairness asks the same of machines: the decision should depend on what truly matters, not on whether someone is a woman, elderly, Roma, or lives in a small village.

The machine is not biased out of malice. It learns from old data, and if in the past a company mainly hired men for leadership roles, the system treats this as a pattern and perpetuates it. Fairness means measuring this bias and correcting it if necessary.

It is important because AI now helps decide on loans, jobs, insurance premiums, and medical examinations. A biased clerk affects a few people, but a biased program can affect millions at once with the same error.

Fairness is not a single formula. There are multiple ways to define what counts as fair, and these partially contradict each other, so ultimately humans must decide which one to apply.

How it works

First, they identify which groups need protection (for example, gender, age, origin, disability) and in which decisions, such as credit assessment or resume screening.
They examine the training data to see whether it includes enough people from every group and whether it encodes past injustices.
They compare group results using metrics: do they receive favorable decisions at the same rate, and does the system err against them equally often?
If they find a discrepancy, they intervene: they rebalance the data, adjust the training, or change the decision thresholds afterwards.
After deployment, the system is checked regularly because data and users change. Disputed cases must be reviewed by a human.

An everyday example

You apply for an online loan, and the bank's program rejects it within seconds. If the system scored you down because you live in a certain postal code where many people have been late with repayments in the past, it penalizes you for where you live, not based on your own financial situation. This is a typical fairness error.

Who is affected

jobseeker, new graduateResumes are increasingly pre-screened by software, and if the system is biased, a human eye may never see your application. → Work & HR

customer applying for a loan or insuranceThe scoring system behind credit assessments and insurance premiums can disadvantage a group even if no one intentionally set it up that way.

patient, doctorA diagnostic system trained primarily on data from one patient group may be less accurate for others. For example, programs recognizing skin conditions may be less accurate on darker skin, and systems indicating certain heart diseases may err more frequently in women. → Health

developerYou must measure and document whether your model or agent treats different groups differently. → Code

A common misconception

MythIf the system does not receive gender or origin data, it cannot be biased.

In realityFrom other data, such as address, first name, school, or shopping habits, the machine indirectly infers the same. Fairness must therefore be measured by outcomes; simply omitting data is not enough.

Limits and risks

Different mathematical definitions of fairness are provably impossible to satisfy simultaneously when the frequency of the event under study (for example, late loan repayments) differs between groups from the outset. Which one is applied is therefore always a choice of values.
A 'fair' or 'unbiased' label on a product means nothing if it is unclear what was measured, which groups were involved, and on what data.
Reducing bias often slightly reduces average accuracy, and this trade-off is not always stated openly.
Group-level fairness does not guarantee that a correct decision will be made in your individual case, which is why the possibility of legal remedy and human review is important.

Similar, but not the same

algorithmic bias – Bias itself is the error, meaning systematic partiality, while fairness is the goal and the toolkit used to prevent it.
transparency – Transparency means it is known how the system decides, but a fully transparent system can still be unfair.
equal treatment – Equal treatment is a legal requirement and is mandatory even if a decision is made or prepared by AI; in such cases, responsibility lies with the organization using the system. AI fairness is the technical toolkit that helps meet this requirement through metrics and tests, but it does not replace legal compliance.

Related concepts

Broader

responsible AI, AI ethics, trustworthy artificial intelligence

Narrower

algorithmic bias, fairness metrics, bias testing, indirect discrimination

On the edge

transparency, Explainability, accountability, human oversight, training data

A brief history

The topic came into the spotlight in 2016 when ProPublica reported that the COMPAS risk assessment tool used in some US courts incorrectly classified Black defendants as high risk more frequently. In the same year, several research groups demonstrated that common definitions of fairness generally cannot be satisfied simultaneously. The EU's AI Act, adopted in 2024, requires that training data for high-risk systems be examined for potential biases. These obligations are entering into force gradually.

Where things stand

By the end of 2026, larger developers and regulated sectors (banking, HR, healthcare) already routinely conduct bias testing, but there is no unified, universally accepted standard. Measuring the fairness of language models and autonomous agents is still immature.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

US judge questions Google's handling of publisher content in AI OverviewsAug 31, 2026 · Marketing
Germany's AI migration act raises concerns over safety, fairness and legal standardsAug 27, 2026 · Jog

Also known as

AI fairness · AI fairness · algorithmic fairness · algorithmic fairness · AI fairness · algorithmic fairness · fairness in machine learning · ML fairness

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.