Algorithms are now making decisions that shape the texture of ordinary life: who gets a loan, who gets an interview, who gets approved for housing, who gets flagged by a health system. These decisions are made at scale, quickly and without explanation.
And the checking — the auditing of whether these systems decide correctly, fairly and lawfully — is still in its infancy. The asymmetry is the quiet scandal of the algorithmic age.
The scale of algorithmic decisions
The volume of decisions delegated to algorithms is larger than most people realize.
Credit scores, developed by statistical models, shape access to finance for nearly everyone. Resume screeners filter job applicants in many large employers. Risk assessments inform bail, parole and insurance. Medical algorithms help triage patients. The list grows every year, and each entry is a decision that used to require a human.
The scale is not the problem in itself; scale can be efficient. The problem is the asymmetry: the decisions are delegated, but the oversight is not.
Why auditing is hard
There is a reason the auditing lags, and it is not just neglect; it is difficulty.
Modern algorithms are often too complex for a person to fully explain. They are trained on historical data, which means they inherit the biases in that data. They are updated continuously, so an audit describes a moving target. And they are usually proprietary — the details of their design are treated as trade secrets, withheld even from the regulators.
The difficulty is real, but it is also an excuse. The fields that made safety work — aviation, medicine — also faced hard problems and built systems to address them.
The bias in the data
The deepest problem with algorithmic decisions is that they inherit the past.
A model trained on historical lending data will reproduce historical lending patterns, including the patterns that discriminated. A resume screener trained on past hires will encode whatever shaped those hires. The algorithm does not invent the bias; it inherits it — and then applies it faster, more consistently and at a scale no human could match.
This is why the claim that algorithms are more objective than humans is only half true. They are more consistent; they are not more neutral. They make the past’s biases more efficient.
The feedback loop
Worse, the biased decisions feed back into the data that trains the next version.
If a screening algorithm admits fewer candidates of a certain group, those candidates appear less often in future hiring data, and the next model learns an even stronger pattern. The bias compounds. Without intervention, the system does not correct itself; it entrenches itself.
This is the loop that makes algorithmic bias a moving problem rather than a fixed one — and it is why occasional audits are not enough; the monitoring must be continuous.
What a real audit would look like
It is worth being specific about what meaningful oversight would require.
A real audit would test the system on diverse populations, measure outcomes against stated objectives, examine the data for bias, and publish the results. It would include the authority to demand changes when the tests fail. It would be independent — not paid by the company that built the system, and not bound by its confidentiality. It would be continuous, because the system changes continuously.
None of these requirements are exotic. They are the standard expectations applied to other consequential technologies. They are simply not yet applied here.
The accountability gap
Underneath the technical difficulty is an accountability gap that is harder to close.
When an algorithm makes a wrong decision — a loan denied, an applicant rejected, a patient mis-triaged — who is responsible? The developer? The deploying company? The data? The answer is currently: no one in particular. The decision is made by a system, and the system has no legal personality and no conscience. The accountability evaporates into the architecture.
This is the defining governance question of the algorithmic age, and it remains unresolved in most jurisdictions.
The honest conclusion
Algorithms are not going to stop making decisions; they are too useful and too profitable.
The question is not whether to use them. It is whether the societies that use them will build the oversight to match their power — the independent audits, the continuous monitoring, the clear accountability, the right to explanation for the people affected.
The technology that makes the decisions is a decade ahead of the systems that check them. Closing that gap is not a technical project; it is a legal, institutional and political one. And the people whose lives are shaped by the algorithms — which is to say, everyone — have a stake in its outcome.
The errors that are not audited will not stay hidden forever; they will surface, in denied loans, in missed diagnoses, in the small betrayals of the automated decision. The question is whether they will surface because someone was watching — or because someone was harmed.