Is an Algorithm’s Accuracy Enough to Make It Fair?

An Application That Never Reached the Table

In this imagined scene, Noura applies to a small training programme. She reads the requirements, assembles her work, and submits before the deadline. A polite automated rejection arrives. Asked why, an employee explains that the system ranks applications very accurately. Noura accuses no one of bad intentions; she wants to know what was measured in her application and whether a human can review it. The number reassuring the institution has not yet answered her question.

Imagine, meanwhile, a team exhausted by hundreds of applications. Manual screening was slow and inconsistent, perhaps affected by favouritism. The team sees an algorithm as a chance to improve work, not an excuse for injustice. Both parties have legitimate questions: how can selection improve, and how can we ensure some applicants do not pay for improvement without appearing in the overall result? Dividing humans and machines into opposing camps is insufficient. The tool operates within human arrangements choosing its goal, data, and appeal process.

This article considers fairness in decisions supported by algorithms. It tests no current product and issues no legal judgement on an institution. It explains why “high accuracy” needs further questions: accuracy in predicting what, for which group, with which error considered more harmful, and under whose responsibility? Historical studies remain tied to their test samples and the systems of their time.

Questions to begin with:

• Does average performance conceal large differences in errors between groups?
• Does the system measure the quality we seek, or a readily obtainable indicator?
• What do we do when different fairness criteria conflict?
• Does an affected person have an understandable path to correcting an error?

1. When the Average Is Correct but Insufficient

Begin with an imagined numerical example, not a study result. One hundred applications comprise ninety from one group and ten from another. The system correctly classifies eighty-nine in the first group and misclassifies all ten in the second. Overall accuracy is eighty-nine per cent. Mathematically correct, this statement does not reveal that the smaller group received not one correct decision. The arithmetic contains no trick; the deficiency lies in the question we stopped at.

If the system improves enough to classify the first group’s ninetieth application correctly, average accuracy rises to ninety per cent while the second group remains unchanged. A presenter can truthfully describe improvement while some users experience none. The lesson is not that an average is worthless, but that it answers a general question. Where error distribution matters, we need to look behind it and know group sizes, distinguishing a rate based on few cases from a broader result.

This seems obvious with numbers on paper but disappears when a dashboard displays one large number. A concise interface is not inherently at fault; abbreviation is necessary. The danger is letting it replace the explanatory report. Someone reviewing a system affecting opportunities needs to know what is absent from the screen: use categories, tested environments, and cases lacking sufficient data.

Joy Buolamwini and Timnit Gebru’s 2018 Gender Shades study provides a specific empirical example. Testing three commercial systems classifying gender from facial images, it found differences in error at the intersection of gender and skin tone. Errors for darker-skinned females reached 34.7% in one system, while the highest error rate for lighter-skinned males was 0.8%. This is a comparison within a historical gender-classification test, not a judgement on identity recognition accuracy or every system available today.

The example’s value is making “For whom does the system work?” part of evaluation. It does not permit jumping from differing performance to knowing its complete cause or a designer’s intention. Review can reveal a problem requiring repair without explaining every root. Responsible research distinguishes observing disparity, explaining it, and judging a particular use: related steps, but not one step.

For someone affected by a decision, an average has no magical power to make an individual case fair. Even an excellent overall system can err about one person. Review therefore remains necessary however high accuracy rises. This does not contradict trust in measurement; it specifies what measurement can establish and what requires another procedure to restore someone’s right.

2. Who Chooses What the Machine Learns to Call Success?

The goal precedes the question of accuracy. Return to the imagined training programme. The institution may seek people who will benefit from training without being able to measure that directly before admission. It uses another indicator: prior grades, historical evaluations, or completion of an online form. The indicator may help or fall short. Correctly predicting it establishes predictive ability for that indicator, not an automatic solution to who deserves the opportunity.

Suppose “success” means completing every activity on time. Someone able to learn may struggle because the schedule conflicts with caregiving. Confusing ability with ease of complying with one arrangement answers a different question from the declared one. This illustrative example describes no particular institution. It shows why defining an outcome requires discussion with people who understand the field and those affected by selection.

Not all available information suits every purpose. An institution may know login counts without those being synonymous with understanding. It may know an application’s writing style without that encompassing suitability for training. Ease of measurement creates an understandable administrative temptation: what we can count sometimes looks firmer than what needs interpretation. Accurate counting of the wrong thing gives no knowledge of the thing we want.

Removing a group name or sensitive characteristic is not enough to guarantee disparities disappear. Hardt, Price, and Srebro’s 2016 work on equality of opportunity discusses how other features may proxy an omitted characteristic. It proposes a statistical criterion concerning chances of a positive outcome among those with actual positive outcomes, while identifying limits of purely statistical tests. This is a specific technical definition, not a comprehensive definition of justice.

The phrase may seem small, but prevents an undeserved sweeping promise. If researchers achieve fairness by one criterion, we should not delete that qualification and say the system is fair in every sense. It is like a building passing a lighting test and being declared suitable for every user in every respect. The test matters, but its value depends on what it actually tested.

Responsible work requires stating a goal in language a non-specialist can question. Are we predicting performance or expanding access to learning? Evaluating the past or discovering potential never given room to appear? An institution may need several goals and should state their priorities. The algorithm does not conduct this debate for us; it implements a choice or learns from examples already embodying human choices.

3. Why Might Two Serious Fairness Tests Disagree?

Imagine a committee seeking a system assigning probabilities of success in a task, then using scores for admission. One member asks whether the same score means similar probabilities across groups. Another asks how many capable people from each group were wrongly excluded. A third asks whether admission rates are equal. These are different questions. Each may have ethical or practical grounds, but shared use of “fairness” does not make their tests synonymous.

In a simple numerical example, twenty people receive a probability of one half and ten succeed. This is consistent with the score’s meaning in that small group, but does not identify which ten will succeed, determine who should be admitted, or prove the test suitable. It illustrates calibration only, not proof of a model’s calibration from one observation or few cases.

Alternatively, we may want to ensure we do not miss capable people at differing rates. We examine those who would achieve the desired outcome and ask how many received a positive decision. This resembles a question about an entrance gate more than about the truthfulness of a probability score. Actual ability itself may be hard to determine when rejected applicants never had a chance to try, so what data establish and what we assume must be stated.

In a paper available in 2016 and presented at ITCS in 2017, Kleinberg, Mullainathan, and Raghavan established conflict among three specific conditions for risk scores: calibration within groups, equal average scores across groups within the negative class, and the corresponding equality within the positive class. In their framework, the conditions coexist only in restricted cases, including perfect prediction or equal base rates. This is a mathematical result under its assumptions, not a declaration that every form of justice is impossible.

Readers need not memorise equations to benefit. Ask anyone calling a system fair: by which criterion, and why choose it for this use? If another criterion fails, how do you explain that? The conversation moves from seeking an absolute stamp to evaluating an explicit argument. The argument may be strong or weak, but can be reviewed rather than accepted as a comforting word.

Conflict does not license choosing whichever metric makes us look best. Data quality, goal definition, or service delivery may improve before accepting a harsh trade-off as inevitable. Nor does a theorem about one set of conditions prove every ethical disagreement can be solved by equation. It identifies a constraint in a particular model and returns responsibility to humans for explaining what they wish to protect.

Observed differences in group rates should not be treated as their own natural and final explanation. A rate describes data and may be shaped by earlier opportunities, sample selection, or outcome recording. The number alone gives no causal story about people. The distance between describing data, explaining their origin, and choosing how to act on them must remain clear.

4. A Decision Is a Chain of Responsibility, Not a Result Appearing Alone

A 2022 report from the US National Institute of Standards and Technology, NIST, notes that biases may produce harmful outcomes even without intent. This moves discussion from proving bad intentions to examining observable effects. Lack of intent does not excuse correction, just as disparity does not automatically reveal motive.

Noura’s imagined decision passes through many stages: writing programme requirements, collecting applications, selecting training data, designing a model, setting a threshold, and converting outputs to acceptance or rejection. If something fails at the end, each party pointing backwards is insufficient. Whoever chose this tool for this position is responsible for understanding its limits and providing ways to address effects.

Imagine a model scoring applications to help order review, then becoming a final exclusion tool under time pressure. Its equation may be unchanged, but its function has changed. A tool helping staff start with easier files is not necessarily suitable for closing other people’s files. Actual use must therefore be described, not the use written in an initial presentation from which practice has drifted.

“There is a human in the loop” alone is insufficient. The person may be able to object, or may be expected to sign quickly without time or information. In the latter case, human presence is merely formal. What can the reviewer actually do? Can they see reasons and underlying data? Can they change the result and document why, or only pass it to the applicant?

For users, notification that a decision occurred differs from understanding what can be corrected. A wrongly recorded employment date needs a correction route. An objection to institutional criteria requires knowing who reviews them. This need not expose others’ data or publish private information; it requires an appropriate explanation and a clear accountability path.

These are organisational proposals ethically inferred by the article, not claimed legal obligations in every country. Procedures may differ by field and data sensitivity, but the question remains: can a person reach whoever bears responsibility? If the sole answer is “the system said so”, a human choice has been hidden behind a tool’s name.

An individual complaint may reveal a general flaw or an isolated record error. We should not decide beforehand. Examine the case, ask whether it recurs, and repair what is established. A team should not equate acknowledging an error with condemning the entire project. A correctable system may deserve greater trust than one preserving an appearance of infallibility at users’ expense.

5. In the Light of the Quran and Islamic Wisdom

The Quran does not give a statistical definition of model calibration or equal error rates. It places responsibility for justice on people and resists using interest or affiliation to justify favouritism. God says:

The verse commands justice and testimony, mentioning parents, relatives, wealth, poverty, and warning against following desire. Connecting it with reviewing contemporary tools is ethical reasoning: accepting examination of a system we built or purchased even when the findings are uncomfortable. We should not attribute to the text a technical choice among criteria it was not addressing.

In Surat Al-Ma’idah, the text prohibits hatred from causing departure from fairness:

This recalls that procedural justice is tested when it concerns people with whom we are uncomfortable too. In the article’s reflection, demanding explanation when a decision harms us while accepting silence when it benefits us is insufficient. The ethical standard exceeds both the tool owner’s comfort and their wish to announce success.

A hadith describes the just as:

The hadith concerns the merit of the just and responsibility for what they oversee. We do not use it as a ready-made certificate of a particular system’s fairness, but as a reminder that delegating a task to a tool does not remove the responsibility of whoever oversees it. Translating this responsibility into procedures in each field requires knowledge, expertise, and discussion, not citation alone.

6. A Serious Objection: Is the Machine Not Better than Human Favouritism?

A programme director may say: previous decisions were arbitrary and slow, so why impose a test the humans themselves cannot pass? This matters. Fair comparison does not place a real machine beside an ideal person. Available alternatives should be compared under the same conditions, recognising documented improvements in consistency, speed, or reduced errors.

Improvement does not make harm irrelevant. Total errors may decline while a small group’s problem grows, as the numerical example showed. Depending on use, a tool may facilitate examining decisions or make appeals harder. Improvement and costs therefore need describing together. We neither reject technology for imperfection nor grant immunity because it is better in one respect.

A researcher may also object that popular talk of “algorithmic bias” confuses statistical disparity with ethical injustice. This is sound too. Not every group difference sufficiently proves impermissible discrimination, and not every equality sufficiently proves fairness. Goals, context, earlier opportunities, and data quality matter. Statistical evidence helps formulate sharper questions without containing all answers.

We should also avoid applying old tests to changed products. The 2018 study was presented as a lesson in evaluation method, not a report on today’s services. Papers on criteria were presented as contributions with specific assumptions, not final agreement on one definition of justice. Knowledge gains strength by making these limits visible, not deleting them.

Practical Applications: Questions before Adopting a Decision

If responsible for a screening tool, describe one decision it will affect and distinguish its output from the final decision. Write what it actually predicts and where the label treated as ground truth comes from. Request more than an average: sample sizes, error types, and confidence limits for underrepresented groups. An attractive percentage should not replace understanding.

Then identify who reviews errors and how users reach them. Test the route: submit a hypothetical correction to inaccurate data and verify it reaches the authorised person. Specify what calls for suspending use, narrowing scope, or reassessment. These are proposed accountability steps, not formal certification or a substitute for specialist expertise.

If affected by a decision, begin with specific questions about information used, possible correction, and the review authority. Retain what you submitted and the reply without publishing others’ data. Do not infer discrimination from rejection alone or assume an automated answer makes review impossible. The aim is moving from guessing causes to requesting examinable information.

Test these questions in an imagined meeting before Noura’s programme launches. The technical lead says the tool saves hours. The training lead asks whether rejected files will undergo regular sampled review. A support employee suggests rejection messages explain data correction rather than merely thank applicants. No proposal alone settles fairness, but each makes a hidden point observable. Discussion becomes cooperation among different expertise, not a contest in technical vocabulary.

The team may discover screening saves time for deeper review of borderline cases. Or it may discover no time is saved after correction work is counted. We do not know beforehand, so benefit must be measured across the whole process. Moving a burden from an employee to applicants resubmitting documents is not costless saving; it changes who bears the cost. Naming that change helps determine whether it is justified and how to ease it.

It helps to record cases a tool should not decide alone: documents it cannot read, experience unlike its training examples, or applications in an untested language. We do not assume every system contains these cases; teams should investigate their own. Acknowledging an unknown case is more accurate than issuing a confident-looking score no one has sufficient grounds to interpret.

An institution needs a memory of corrections. If Noura’s decision changes because information was transferred incorrectly, correction should reach the source so it does not recur in her next application or another’s. If revision is a reasoned exception, record it as such rather than confusing it with data repair. Different problems require different remedies. Grouping everything under “error” may repair a case while leaving its cause active.

Even those who do not build systems can read critically. When an advertisement calls a tool fair, look for the test’s name, date, and scope, not just its percentage. When sharing a report of disparity, preserve the distinction between historical findings and current conditions, and between classifying a characteristic and recognising identity. Accurate technological criticism is part of fairness: a sweeping claim should not be answered with another sweeping claim.

In either role, separate three demands: correcting your case, understanding selection criteria, and examining effects across user groups. Each may need a different process. One general accusation can obscure detail; separation clarifies what we know and need to know. No particular outcome is guaranteed, but discussion becomes more verifiable.

Conclusion: A Number Needs Someone Responsible

Accuracy matters in many uses, but does not encompass justice. We need to know the goal, error distribution, meaning of the chosen criterion, and handling of objections. Equations do not make these questions less human: decisions ultimately concern opportunities, time, and lives.

The imagined programme’s system may genuinely be useful. Noura does not ask to restore chaos, but that overall success not close questions about her file. Justice begins when we can say who decided, on what basis, and who can correct established error.

Questions to leave with:

• What information is still missing when I hear a high accuracy rate?
• Do I know the difference between my intended goal and the indicator the system measures?
• Which kind of error deserves priority protection in the use I have in mind?
• Is there genuine review, or merely someone repeating the screen’s result?