Artificial intelligence is increasingly used to make decisions about hiring and lending.

Today many banks and companies use artificial intelligence to decide who will receive a loan or who will be invited to a job interview. This technology has clear advantages, but it also creates serious risks.

The main advantage is speed and consistency. A computer can read ten thousand applications in a few minutes, while a human manager needs several weeks. The machine is also never tired and never in a bad mood, so the applicant who is number five hundred receives the same attention as the first one, which is not true for a tired human being.

Another advantage is that some people receive an opportunity for the first time. Many people have no credit history because they never took a loan, so a normal bank refuses them automatically. A system that looks at other information, for example regular payments for electricity or rent, can see that this person is reliable.

On the other hand, these systems learn from old decisions. If a company employed mostly men for twenty years, the computer learns that successful candidates are men, and it starts to reject women. The machine does not understand that this is discrimination; it only repeats the pattern from the past.

Another problem is that nobody can explain the decision. When a bank refuses a loan, the customer wants to know the reason, but often even the engineers cannot explain exactly why the system said no, so the person cannot correct the mistake or complain.

In conclusion, automatic systems are fast and can help people without a credit history, but they can repeat old discrimination and give decisions that nobody is able to explain.

Automated systems increasingly decide who is shortlisted for a job or approved for credit. The advantages are considerable, but so are the failure modes, and both follow from the same feature: these systems learn from what has been done before.

The first advantage is consistency. Human decisions vary with fatigue, mood and the order in which applications are read, and these variations are invisible and unappealable. An automated system applies the same rule to every case, which means that where it is wrong, it is wrong in a detectable pattern — and a detectable pattern can be measured, challenged and corrected in a way that a thousand individual human judgements cannot.

The second is access. Conventional credit assessment excludes anyone without a borrowing history, which disproportionately affects the young, migrants and people who have simply never needed credit. Systems that assess a wider range of evidence, such as regular rent or utility payments, can identify reliability in people whom traditional scoring would have refused without examination.

The central disadvantage is that training data records past decisions, including discriminatory ones. A model trained on twenty years of hiring at a firm that favoured men will treat maleness as a marker of suitability, and removing the field for sex does not solve it, because other variables act as proxies. Discrimination therefore survives in a form that appears objective and is harder to contest.

A further difficulty is opacity. Complex models often cannot produce a reason for a particular decision, so a rejected applicant has nothing to appeal against and nothing to correct. Errors also replicate at scale: a biased recruiter affects one company, whereas a biased model licensed to hundreds affects an entire sector simultaneously.

In conclusion, these systems offer consistency and can widen access to credit, but they inherit historical bias and produce decisions that are difficult to explain or challenge.

Both the promise and the danger of automated hiring and lending come from a single property: these systems learn by finding patterns in decisions that were made before. That is what makes them consistent, and it is also what makes them capable of reproducing the past with perfect fidelity, including the parts of it nobody would defend.

The advantage most worth stating is not speed but consistency, and it has an underrated consequence. Human decisions vary with fatigue, hunger and the order of the pile — variation that is invisible, unrecorded and effectively unchallengeable. An automated system applies the same logic to every applicant, which means its errors form a measurable pattern. Paradoxically, a biased algorithm is easier to prove biased than a biased panel: you can test it on ten thousand synthetic applications and observe exactly what it does, which no employment tribunal can do to a hiring manager's intuition.

The second advantage is genuine expansion of access. Conventional credit scoring cannot assess anyone lacking a borrowing history, so it refuses the young, recent migrants and the financially cautious not because they are risky but because they are unmeasured. Models drawing on rent, utilities and cash-flow patterns can distinguish reliability within that group, extending credit to people a traditional system declined without ever examining them.

The corresponding disadvantage is that the training data is a record of historical behaviour, not of merit. A recruitment model built on two decades of a firm's decisions learns whatever those decisions encoded, and if the firm promoted men, the model treats the correlates of maleness as predictive. Deleting the sex field achieves little, because proxy variables persist — a sport, a school, a career gap, a postcode — and a sufficiently capable model reconstructs the protected characteristic from them without ever being told to.

Opacity then removes the remedy. A rejected applicant told only that a model declined them has nothing specific to contest and nothing to correct, and the ordinary defence against an unfair decision — ask why, demonstrate it is wrong — is unavailable. Scale compounds this decisively: a prejudiced recruiter harms the candidates of one firm, whereas one widely licensed model applies the same misjudgement across an entire industry, so an individual's error becomes a systemic exclusion.

Weighed together, these are not arguments for or against the technology so much as arguments about governance. The same consistency that makes automated bias dangerous is what makes it auditable, and whether these systems widen opportunity or entrench exclusion depends almost entirely on whether anyone is permitted to inspect them.