The myth that most distorts debate about workplace algorithms says: “when a machine decides, human bias disappears.” It sounds reassuring because it presents the system as a neutral set of scales, but it confuses calculation with justice. Software can apply a rule consistently while inheriting unfair data, measuring the wrong proxy or imposing priorities that nobody debated. Understanding that difference helps young workers, families and teams ask precise questions without falling into either technological fear or blind obedience.
Imagine Diego, twenty-two, applying for a warehouse job. He records a short video and receives an automated rejection minutes later. The company says the process is objective because every applicant encounters the same tool. Yet an equal procedure does not guarantee equal opportunity. The model may have learned from old hiring decisions, penalised a pause related to disability or treated a poor connection as lack of interest. The important question is not whether an angry person caused the rejection, but which criteria produced it and whether they can be justified.
The myth begins with an incomplete truth#
Computers can avoid some inconsistencies. They do not arrive tired for an interview, change mood between the first and last candidate, or forget instructions while checking thousands of records. That regularity has value. The mistake is to infer that a consistent rule is necessarily a fair rule. A requirement for night availability may be applied identically to everyone but disproportionately exclude people who provide care, study or depend on public transport, even where night work is not essential to the role.
Computers can avoid some inconsistencies. They do not arrive tired for an interview, change mood between the first and last candidate, or forget instructions while checking thousands of records.
There is no completely raw data either. Somebody chooses what to collect, how to classify it, how long to retain it and which absences count as poor performance. Even a seemingly simple variable such as parcels per hour depends on the assigned zone, weight, incidents, machinery and help given to colleagues. When a system turns that number into an individual score, it shrinks a collective reality. Mathematical precision in the calculation cannot repair poverty in the original question.
A model often learns from the past. If historic promotions favoured people available outside normal hours, a system may discover that availability predicts advancement without understanding that it reflects one particular culture. If a word on a résumé correlates with rejected candidates, software may use it as a signal despite no causal connection to competence. Removing names or sex does not guarantee neutrality: postcode, career breaks, school, or linguistic patterns may operate as proxies.
Objectivity is not the same as absence of intention#
Algorithmic discrimination does not require a programmer who intends harm. It can arise from poorly chosen objectives, incomplete samples or unequal impacts. Debating only whether somebody acted in bad faith is therefore a distraction. A responsible assessment examines outcomes: who is excluded, how often, in which circumstances and with what route to correction. It also examines necessity. If a variable is not needed to perform the job, predictive power alone does not make its use legitimate.
The word “intelligence” creates another misunderstanding. These systems do not understand a life story as a listening person might. They detect patterns and produce classifications, predictions or recommendations. A risk score is not a property of the individual. It results from a model, data, threshold and particular moment, and may change when any of those elements changes. Treating it as a discovered truth turns a probabilistic tool into a moral label.
Who formulates the objective matters as well. Optimising retention can sound benign, but a model seeking people like those who stayed for years in a hostile workplace may reject candidates who would thrive under better conditions. Optimising speed may encourage skipped breaks. Reducing absence may penalise health rights. Before judging the algorithm, write in ordinary language which behaviour it rewards, what harm it could cause and who receives the benefit.
A simple test for dismantling false neutrality#
A family can teach a young worker five questions. First, which actual decision changes: an interview, shift, bonus or continued employment? Second, which data feed the system and where did they come from? Third, what does the score represent and what can it not represent? Fourth, which errors have been tested across different groups? Fifth, which person can review and change the outcome? If the organisation answers only “that is our algorithm,” it has not yet explained the decision.
At work, those questions can become a one-page record. Note the declared purpose, affected population, main variables, point of use, threshold, consequence and accountable human. Then test edge cases: disability, pregnancy, medical leave, changed name, accent, old device, unstable connection, interrupted employment or caring responsibilities. The exercise is not designed to prove in advance that the system fails. It discovers where failure may occur before one person bears the damage.
Comparison must include an alternative. Saying a tool is ninety per cent accurate does not reveal whether it improves the old process, whether the remaining errors cluster in one group or whether mistakes can be reversed. Wrongly rejecting an applicant and showing another vacancy is not equivalent to dismissing a worker because of a defective alert. The more severe and irreversible the consequence, the more evidence, oversight and ability to suspend the process are required.
What rules say and what they cannot decide for us#
The European Artificial Intelligence Act classifies certain systems used for recruitment, selection, working conditions, promotion, termination, task allocation and performance evaluation as high risk. The official text is available through EUR-Lex, Regulation (EU) 2024/1689. The classification does not prove that every use is unfair. It recognises that these tools can profoundly affect livelihoods and requires risk management, documentation, data quality, human oversight and monitoring during operation.
In Spain, legal worker representatives have a right to information about the parameters, rules and instructions underpinning algorithms or artificial-intelligence systems that influence working conditions and access to or continuation in employment, including profiling. The amendment appears in the official Spanish state gazette. This right does not make every system understandable automatically, but it prevents operation from being treated as an exclusively private matter between employer and vendor.
Article 22 of the General Data Protection Regulation addresses specified decisions based solely on automated processing that produce legal or similarly significant effects. In applicable cases its safeguards include human intervention, the opportunity to express a point of view and the right to contest a decision. Read the official GDPR text and avoid universal promises: workplace tools do not all fit the provision in the same way, and concrete facts matter.
Law establishes a floor; it cannot decide alone what kind of work people want. Measurement can be lawful while damaging autonomy, cooperation or mental health. Joint research by the International Labour Organization and the European Commission's Joint Research Centre describes operational benefits alongside risks of intensification, surveillance and declining job quality in real sectors. The report is available from the ILO's official research repository.
What meaningful human review looks like#
Putting a person at the end is not enough. If the reviewer sees only a score, must resolve hundreds of cases per hour or fears a penalty for disagreeing, their involvement is decorative. Meaningful review displays relevant data, explains the criterion, allows contextual information to be heard and grants authority to change the outcome. It also records what was checked and why the decision was confirmed or corrected, allowing repeated errors to improve the system.
The complaint channel must be accessible and safe. A young worker can retain the advertisement, contract, internal policy, emails, screenshots and a timeline. Instead of claiming “artificial intelligence discriminated against me,” it is more useful to write: on this date my shift changed; my manager attributed it to a score; that score contains two authorised absences; I request correction, an explanation and review by somebody authorised to change it. Precision provides more protection than an accusation that cannot be tested.
The organisation should not treat each complaint as an isolated exception. When several people correct the same variable or one group appeals more frequently, there may be a design problem. Useful indicators include the share of decisions changed, response time, errors by type and group, complaints, harms and subsequent modifications. Auditing is not producing a certificate once. It means observing the system as data, employment and workplace practices change.
A family conversation without technophobia or naivety#
When a son or daughter says an app assigns impossible shifts or rejected them, a family can listen before reaching a conclusion. It may be an automated decision, a company rule implemented in software or a convenient explanation from a manager. Separate established facts, messages received and suspicions; then identify the desired result: information, corrected data, restored hours, reviewed discipline or support. A person need not understand source code to demand an understandable process.
The opposite myth should also be rejected: not every algorithm is necessarily harmful. A well-bounded system may detect scheduling conflicts, distribute workloads transparently or warn about safety risks. The question is not machine versus human. It is which combination of tool, rules, participation, evidence and control best protects people. Humans also make mistakes and discriminate; automation can reveal patterns, but only when impact is measured and there is willingness to correct it.
A minimum agreement before deployment should specify purpose, permitted data, prohibited uses, testing before and after launch, worker participation, training, accountable human, challenge channel, time limits, data deletion and suspension criteria. It must anticipate change: updating the model, expanding the workforce or applying it to another decision requires a new assessment. Generic consent when joining a company does not legitimise any future surveillance or replace employment and data-protection duties.
The sentence that improves the debate#
Instead of asking “is the algorithm objective?”, ask “objective for which purpose, measured with which data, compared with which alternative and reviewable by whom?” That sentence forces specificity. It permits benefits to be recognised without erasing harms and risks to be identified without turning technology into a villainous character. Above all, it restores decisions to their proper place: an organisation chooses purposes, thresholds, consequences and safeguards even when it buys the software elsewhere.
Diego need not prove that a machine held prejudices. He can ask what was assessed, how it influenced the outcome, whether a person reviewed the rejection and how to correct data or provide context. The company need not promise perfection; it must show that it understands limitations, compares outcomes, hears challenges and stops harmful uses. The neutrality myth impoverishes debate because it closes questions. Leaving it behind is not distrusting every innovation; it is requiring that power over working life remain explainable, shared and accountable.
