AppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

Algorithmic accountability

Also called algorithmic audit

Algorithmic accountability means that when software makes or shapes decisions about people, someone can check it, explain it, be challenged over it and fix it.

Safety and societyCurious

In plain words

Algorithms already help decide who gets a loan, which job applications reach a human, how exam grades are set, and which welfare claims get investigated. When those decisions go wrong, "the computer said so" is not an answer. Algorithmic accountability says a person or organisation must answer for it.

In practice that means four things. Audit the system, including whether it treats groups of people differently. Explain decisions in terms the affected person can understand. Let people appeal to a human. And fix what the checks find.

This matters because automated decisions can harm thousands of people at once, quietly, before anyone notices. Accountability is how those mistakes come to light and get corrected.

Everyday example

A cricket umpire and the third umpire

An on-field umpire gives a batter out. The batter can ask for a review. The third umpire checks the replays, the decision is shown on the big screen, and if it was wrong, it is overturned. Everyone knows who made the call and how to challenge it. Algorithmic accountability asks for the same review system for decisions made by software.

Where the example breaks: A cricket review covers one ball, in seconds, with cameras everyone can see. An algorithm's errors can be spread across thousands of cases, hidden in data and code, and those affected may not even know software was involved.

Go deeper

How it actually works

The idea grew from real failures, and laws are slowly turning it into duties.

  1. Lessons from real cases

    In 2016 ProPublica reported that a US criminal risk score, COMPAS, wrongly labelled Black defendants as high risk almost twice as often as white defendants; the maker disputed the analysis. In 2020 England's exam regulator used an algorithm to set A-level grades and 39% came out lower than teachers' assessments, until the government dropped it. In the Netherlands, a tax authority risk model contributed to tens of thousands of parents being wrongly accused of benefits fraud; the cabinet resigned in January 2021.

  2. Audits

    An algorithmic audit tests a system's outputs, often comparing results across groups for bias. Since 5 July 2023, New York City requires employers using automated hiring tools to have an independent bias audit every year and publish a summary.

  3. Documentation and explanation

    Model cards, impact assessments and explainability tools record what a system is for, how it was tested and why it reached a result. Without records, no one can check anything.

  4. Appeals and human review

    People affected need a way to contest a decision and reach a person who can overturn it, a role often called human in the loop. The reviewer needs real power and information, not just a button to approve.

  5. Laws

    The EU AI Act requires checks and human oversight for high-risk uses such as hiring and credit, with those duties now applying from 2 December 2027. In the US, an Algorithmic Accountability Act requiring impact assessments was reintroduced in June 2025 and has not passed as of November 2026.

Core questions
Who is responsible? Can it be checked? Can it be explained? Can a person challenge it? Will it be fixed?
Common tools
Bias audits, impact assessments, model cards, decision logs, appeal routes
Example law
New York City Local Law 144: yearly bias audits for automated hiring tools, enforced from 5 July 2023

Questions people ask

What is an algorithmic audit?

A structured check of an automated system, usually by testing its outputs. Auditors look for unfair differences between groups, errors, and whether the system does what its makers claim. Some audits are required by law, such as New York City's for hiring tools.

What is the difference between algorithmic accountability and transparency?

Transparency means being able to see how a system works and how it decides. Accountability goes further: someone must answer for the results, people can challenge them, and problems get fixed. Transparency is one tool for accountability.

What went wrong with the 2020 A-level algorithm?

With exams cancelled during COVID-19, England's regulator Ofqual used a model that leaned heavily on each school's past results. 39% of grades in England came out lower than teachers' assessments, and students from poorer areas were hit harder. After protests, teacher grades were used instead.

Is there an Algorithmic Accountability Act?

In the US, a bill of that name was reintroduced in the Senate in June 2025. It would have the Federal Trade Commission require impact assessments of automated decision systems. It is not law as of November 2026.

Was this clear?

1 is confusing, 5 is crystal clear. One tap, no account needed.

Spotted something wrong, or know a better example? Suggest a change

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked November 2026.