Inequality is often hidden within the average. A platform may report that its users receive an answer within minutes while omitting that people with poor connectivity wait for hours. An administration may celebrate that nearly every application is resolved online while failing to see those who leave before they are recorded. A news organisation may count millions of readings while certain communities appear only as objects of suspicion. An average summarises, but it also flattens. When it becomes the sole description of reality, it erases the distribution of benefits, costs, voice, and power to obtain correction. In the construction of shared truth, that statistical disappearance eventually becomes a political disappearance.
The problem is not the calculation of means. They are useful instruments for identifying trends, comparing periods, and administering resources. Trouble begins when an aggregate figure takes the place of people and is interpreted as sufficient proof of fairness. Two systems with the same average response time may be radically different: one spreads moderate delays, while another offers immediacy to most people and traps a minority in a maze. Democratic quality requires questions about who stands on each side of the figure, which cases never entered the measurement, and who can challenge the category under which they were counted.
Truth as common infrastructure depends on that precision. The circulation of true information is not enough; the conditions for producing, finding, and correcting it must also be reasonably distributed. If only groups with time, money, the dominant language, and technical knowledge can document an error, the public record will acquire a stable bias even when each isolated fact appears correct. Equity is not a sentimental consideration added at the end of an analysis. It determines whether collective knowledge represents the world or merely the part of the world that managed to pass through the procedure.
The average does not reveal who pays for efficiency#
Aggregate metrics usually follow the path of someone who completes an action. They measure how long a person who submits a form takes, how much content is found by somebody who remains on a platform, or how many complaints a service resolves. They often omit those who could not begin, abandoned the process, lost documentation, needed translation, or decided that complaining would cost more than the original harm. Their absence artificially improves the result. The system declares speed because it excluded from the clock the people who endured the greatest delay.
Aggregate metrics usually follow the path of someone who completes an action. They measure how long a person who submits a form takes, how much content is found by somebody who remains on a platform, or how many complaints a service resolves.
Work can also be transferred. A tool may reduce ten minutes of professional attention while requiring thousands of people to scan documents, interpret ambiguous messages, or enter the same details again. The organisation records internal savings but does not record external time. When this labour falls on carers, older people, households sharing devices, or workers with rigid schedules, institutional efficiency is financed by unequally distributed private resources. An honest assessment includes that time even when it does not appear in the accounts.
The choice of unit changes the story. An error rate of two per cent looks small until the errors are found to be concentrated in one neighbourhood, language, or administrative situation. Moderation that is correct in ninety-eight per cent of cases may repeatedly silence testimony about violence if the system mistakes its vocabulary for harmful content. Disaggregation does not mean publishing data that identify individuals. It means designing samples, thresholds, and protections that reveal patterns without turning observation into surveillance.
Professional interpretation also needs intervals, sample sizes, and limitations. A striking difference in a very small group may be unstable; a modest but persistent difference across thousands of cases may demand urgent intervention. Transparency does not mean dumping a spreadsheet. It means explaining what was measured, what was omitted, which uncertainty remains, and which decision depends on the indicator. A number acquires public value when it can be understood, challenged, and related to concrete experience.
The infrastructure of truth also distributes voice#
Networks, media, classrooms, and institutions are not neutral containers. Their forms, opening hours, languages, recommendation systems, and rules of evidence decide which testimony becomes evidence. When one person must submit a long complaint in a language they do not command while another has a direct contact, both formally retain the same right but materially exercise different rights. Equality in the rule cannot compensate for unequal access that the procedure itself could reduce.
Reputation deepens this asymmetry. Claims from familiar accounts receive attention and rapid correction, while claims from peripheral communities may need extraordinary evidence to be heard. The platform then uses accumulated engagement as a signal of relevance and reproduces the initial advantage. The cycle looks spontaneous because each step relies on observed behaviour, but observation had already been organised by previous choices. Responsible governance examines these cycles and does not present a historical outcome as the public's natural preference.
Context is unequally distributed as well. An inaccurate headline may reach millions while its correction remains on an obscure page. A label may accompany a post without explaining what evidence produced it or when it will be reviewed. A national statistic may show improvement while concealing local deterioration. Building common truth requires corrections, dates, sources, and limitations to travel with visibility comparable to that of the initial claim. Otherwise, the system rewards first impressions and demands disproportionate work from the harmed party to alter them.
Early participation can detect these burdens before deployment. It is not a request for generic approval from symbolic representatives. It brings in situated knowledge: how a phone is shared within a household, what risk follows disclosure of an address, why a literal translation causes confusion, or when a help channel is inaccessible. The institution retains responsibility for the decision, but decides with a fuller map. It should also explain which contributions changed the design, which were not accepted, and why.
Three cases in which a good average hid a bad distribution#
Case 1: the municipal portal with a ninety-four per cent success rate. A city digitises applications for energy assistance. Its dashboard reports that ninety-four per cent of complete files are processed within five days. The result looks excellent. A neighbourhood association nevertheless notices that many families never reach complete-file status. The portal times out, rejects photographs taken with older phones, and provides instructions in only two languages. These failed attempts never enter the denominator.
An audit reconstructs the funnel from the first visit and disaggregates by device, postcode, chosen language, and need for assistance, with privacy safeguards. Abandonment is concentrated in three neighbourhoods. The council enables progressive saving, reduces file sizes, provides equivalent telephone and in-person support, and allows missing information to be supplied without losing the original date. Average resolution time deteriorates slightly because previously invisible files are now included, but more households receive assistance. The less impressive indicator describes a fairer service.
Case 2: moderation that protected the majority. A social network says it rapidly removes threats and harassment. Its overall accuracy exceeds ninety-seven per cent. Researchers and rights organisations find that the system often removes posts by activists who document racist abuse in order to denounce it, while coded variations used by abusers remain online. The affected group is small in relation to the whole, so its experience barely changes the average.
The platform creates a dedicated appeals sample, involves speakers and independent specialists, preserves quotation context, and compares error by type of harm and language community. It restores posts, notifies affected people, and reviews accumulated penalties. It also measures time to repair, not merely the initial decision. Improvement does not require weaker general protection; it requires a more accurate distinction between documenting violence and perpetrating it. Safety is no longer evaluated only by volume removed, but also by who can speak without carrying repeated errors.
Case 3: the school map of average performance. A region publishes school-level education results to guide resources. One school's average improves, and the administration reduces additional support. The aggregate hides that progress is concentrated among students with stable attendance, while learners who alternate family care, temporary work, or insufficient connectivity fall further behind. The withdrawal of support reinforces the gap the figure concealed.
The education team examines trajectories rather than annual snapshots and combines results with attendance, access to materials, and protected interviews. The analysis is not allowed to become a permanent individual label. The region maintains flexible support, funds tutoring at varied times, and evaluates whether each group can recover learning. Public communication explains uncertainty and prohibits the use of disaggregation to rank students. The aim is not to manufacture an identical result, but to guarantee real conditions for learning and to correct policy when collective improvement abandons part of the population.
The three cases share a pattern. Each institution measured what its procedure facilitated and called absent what its procedure had expelled. Repair began by changing the denominator, listening to those who abandoned the process, and acknowledging external costs. It also required a non-digital or non-automated route with an equivalent outcome. Equity did not appear as an expensive exception. It appeared as a condition for knowing whether the service fulfilled its purpose.
How to audit an inequality that does not yet have a name#
The first step is a distributional hypothesis. In addition to asking whether a system works, the assessment asks for whom it works, who bears time, risk, and exposure, and who receives the power to obtain correction. The inventory covers the whole chain: awareness of the service, access, comprehension, decision, appeal, and repair. A final average cannot replace observation of each transition. Points where people disappear deserve as much attention as the outcomes of those who reach the end.
The selection of groups must avoid two errors. The first is refusing to observe differences because sensitive data may be involved; the second is collecting identities without purpose or protection. Voluntary surveys, user-profile tests, territorial sampling, qualitative research, and analysis by trusted intermediaries can all contribute. Every item of data needs a purpose, retention period, restricted access, and public explanation. A sensitive variable should not be collected when it is unnecessary. When it is essential to uncover discrimination, enhanced safeguards should govern it.
Indicators should combine access, quality, and consequence. Abandonment, repeated attempts, total time, need for assistance, error rates, appeal outcomes, and time to repair all matter. It also matters whether people understand the explanation and trust the channel without exposing themselves to retaliation. No isolated metric offers the whole truth. Numbers locate patterns; case files explain mechanisms; testimony reveals costs the system did not know how to name.
An audit needs sufficient independence. The team responsible for a product knows essential details but has incentives to defend targets and deadlines. Affected people know consequences but should not carry the burden of proving the entire problem. A body with access to documentation, technical ability, and authority to require changes can connect both forms of knowledge. Its findings should include disagreements and limitations, not merely a final score. An audit unable to change budgets, design, or continuity risks becoming a ceremony.
Pause thresholds should be defined before deployment. A sustained difference between groups, a specified rate of overturned decisions, a serious consequence, or an inability to explain an outcome can trigger review. The threshold does not automate judgement; it prevents each inconvenient signal from being dismissed when it appears. The response identifies an owner, a deadline, containment, and repair. If rights are at risk, prudence may require suspension of part of the system while the matter is investigated.
Institutional sources for turning equity into an obligation#
Public frameworks can translate this discipline into testable duties. They provide neither a universal formula nor permission to copy indicators without context. They do converge on several principles: assessment of impacts, protection of rights, explanation of decisions, access to complaint, and attention to those who bear risk. Read together, they show that distribution is not a footnote to efficiency but part of its legitimacy.
- UNESCO, Guidelines for the Governance of Digital Platforms. They place human rights, transparency, accountability, and multistakeholder participation at the centre of governance. Institutional document.
- European Union, Digital Services Act. It establishes duties concerning statements of reasons, complaints, transparency, and systemic risk assessment for platforms and other intermediaries. Official text.
- Council of Europe, Recommendation CM/Rec(2020)1 on the human rights impacts of algorithmic systems. It calls for continuing assessment, effective oversight, and remedies throughout the life cycle. Official recommendation.
- United Nations Office of the High Commissioner for Human Rights, Guiding Principles on Business and Human Rights. The principles articulate due diligence, prevention of impacts, and access to remedy. Institutional publication.
These frameworks reinforce one conclusion: identifying a gap without assigning responsibility is insufficient. An organisation must be able to change an interface, fund an alternative, revise a rule, correct records, and repair consequences. Transparency helps when it enables action. A table of percentages without an owner, date, or route for challenge merely documents distance. Accountability connects knowledge to authority and establishes who responds when an average leaves somebody out.
A democratic practice for looking beneath the figure#
Public communication should resist both triumphalism and exaggeration. It presents the overall result alongside relevant differences, uncertainty, and corrective measures. It protects privacy and avoids describing a community as the problem. A gap does not belong to the disadvantaged group; it arises in the relationship among needs, conditions, and design. This language directs intervention towards barriers that can be changed and prevents a statistic from becoming a stigma.
Changes in methodology should also be published. If the denominator expands or previously invisible abandonment is included, a series may look worse even though understanding has improved. Explaining the break protects institutional learning. A mature organisation does not reward attractive indicators at the cost of ignorance. It rewards the capacity to discover harm, correct it, and preserve evidence so that it does not return under another name.
Cooperation completes the method. Journalists, teachers, associations, administrations, researchers, platforms, and the public hold different pieces of reality. No party should claim absolute neutrality or delegate all responsibility to another. Sharing protocols, protected data, and alert channels reduces the need for each harmed person to begin from nothing. Institutional humility means accepting that a metric useful today may conceal a new cost tomorrow.
An average describes the centre of a distribution; a democracy must also answer for those who disappear at its edges.
Looking beneath the figure does not prevent innovation or demand perfect certainty before action. It requires measurement from access through repair, preservation of alternatives, attention to abandonment, and decisions made with known limits. When this practice is built in from the start, equity improves both justice and knowledge. Shared truth ceases to be the account of those who passed through the system and becomes an infrastructure capable of recognising those left outside, learning from them, and correcting the distribution of voice, time, and power.
