An automated decision can look neutral because it arrives as a score, traffic light or message generated by a platform. Yet people have already made crucial choices: which data count, which objective is optimised, how much error is acceptable and who bears the consequences. When a system filters applicants, allocates shifts, measures productivity or recommends dismissal, the challenge is not merely to improve an algorithm. It is to preserve a genuine opportunity to hear context, correct injustice and accept responsibility. For a young person entering employment, understanding this human side is both protection and a way to participate confidently in decisions that will shape everyday life.
The decision never belongs only to the machine#
Imagine Maya applying for her first supermarket job. She meets the stated requirements, but a tool rejects her application within seconds. Nobody explains whether a word in her CV, an employment gap, her postcode or a personality test affected the result. Saying “the system decided” erases the people who chose to purchase it, configure it and act on its output. An organisation remains responsible when technology comes from an external supplier. It must understand the purpose, test whether the tool discriminates and identify who can correct it. An interface cannot become an institutional hiding place: a screen does not sign contracts, manage a team or answer to someone who has been harmed.
Imagine Maya applying for her first supermarket job. She meets the stated requirements, but a tool rejects her application within seconds.
Automation does not always mean deciding without people. Some tools organise information, flag anomalies or provide recommendations. The danger appears when a recommendation becomes an order in practice. If a manager always accepts a score because only twenty seconds are available, human review is decorative. Meaningful intervention requires time, authority, understandable information and permission to depart from the output without punishment. Ask: who reviews it, what independent evidence can they see, can they speak with the affected person, and is the reason for confirming or changing the proposal recorded? Without those conditions, a “validate” button merely disguises an automated decision and creates false reassurance.
Where a system may influence working life#
Recruitment is the most visible example, but it is not the only one. Software may decide who sees an advert, extract traits from a CV, score a recorded interview or predict early resignation. Inside a workplace, it may allocate shifts, routes and tasks; set targets; compare completion times; flag supposedly unusual conduct; calculate incentives; or recommend training, promotion and discipline. On delivery and transport platforms, a rating can reduce offers or disable an account. Each use affects equality, income, health and reputation differently. A general authorisation to “use AI” is therefore meaningless. Every purpose requires a separate assessment, suitable evidence, boundaries and a route for a person to challenge the outcome.
The first practical question is whether a data point has a legitimate connection to the work. Serving a customer well may be relevant; sleeping badly for one night, living far away or caring for a relative may not be. Historical data are not neutral truth either. If women were promoted less often, or carers routinely received worse shifts, a model trained on those records may learn inequality as though it represented business success. Removing sex from the dataset does not solve the problem when other variables operate as proxies. A serious test compares outcomes across relevant groups, investigates differences and modifies or withdraws the system when the resulting harm cannot be justified.
Everyday data quality matters too. A badge reader fails, GPS loses its signal, a camera misses protective equipment or a prejudiced customer gives an unfair rating. When such events feed directly into evaluation, small errors accumulate until they appear to be a pattern. The worker needs a chance to supply context: equipment failed, the route changed for safety, time was spent helping a colleague or the task was measured incorrectly. A sound process keeps contextual information and allows correction before an adverse measure. It does not force an employee to prove that a black box is wrong; it requires the employer to demonstrate that its information is suitable and its conclusion reasonable.
What preserving the human side really means#
Humanity is not achieved by attaching a friendly sentence to a rejection. It means treating the individual as someone with a voice, rather than a row in a table. Before an important decision, the person should know what is assessed, why it is assessed and what consequences may follow. Afterwards, they need a useful explanation: the principal factors, relevant data and a path to correction. They do not need all the source code to understand that their shifts fell because an acceptance rate included orders that never reached their telephone. An explanation stating only that “the profile did not meet the threshold” provides no opportunity to learn, respond or identify an error.
Human review must happen before harm becomes irreversible. If an account is disabled and an appeal takes two months, eventually winning does not pay the missed rent. Decisions affecting income, continued employment or health require short deadlines, an accessible route and, where possible, suspension while review takes place. The reviewer should be independent enough to question the original objective and able to restore shifts, correct pay or remove a warning. They must listen without turning an appeal into an impossible examination. The question is not merely “did the model malfunction?” It is also “is this individual decision fair when all available information and the seriousness of its consequences are considered?”
Some decisions should not be delegated. A machine may help locate inconsistencies, but dismissal, discipline, refusal of a disability adjustment or judgement of a person’s future potential demands particular care. The European Artificial Intelligence Act classifies various systems used in employment and worker management as high risk. This does not mean every use is forbidden, but it brings requirements concerning risk management, data quality, documentation, human oversight and other safeguards. The complete official text is available at EUR-Lex: Regulation (EU) 2024/1689. The practical duty is to anticipate foreseeable harm instead of waiting for hundreds of individuals to complain.
Rights that turn a complaint into an answer#
Data protection provides essential rules when personal data are used to profile people or make decisions. The General Data Protection Regulation establishes principles including lawfulness, transparency, data minimisation, accuracy and purpose limitation. It also regulates certain decisions based solely on automated processing that produce legal or similarly significant effects. The official text is at EUR-Lex: Regulation (EU) 2016/679. Not every score triggers precisely the same legal right, but no employer should treat that technical distinction as permission to conceal how a process works. Clear information, access to personal data, correction of inaccuracies and proper impact assessment remain basic safeguards.
Spain also provides a collective dimension. Workers’ legal representatives have a right to information about the parameters, rules and instructions on which algorithms or artificial intelligence systems are based when they influence decisions capable of affecting working conditions, access to employment or continued employment. The consolidated Workers’ Statute is available from the Spanish Official State Gazette. Collective information can expose a pattern that an individual claim cannot see: for example, a criterion that systematically cuts one group’s hours or penalises breaks that are essential for safe working. It also allows workers to compare promises made before deployment with real effects afterwards.
When challenging an outcome, document events rather than trying to guess the algorithm. Keep adverts, terms, messages, screenshots, schedules, scores and observable changes. Record dates and consequences: “On 4 May my indicator fell after I declined a shift incompatible with rest requirements; since then I have received six fewer hours.” Ask in writing what data were used, which important factors shaped the outcome, who made the decision and how to request human review. Correct inaccurate data and make a concrete request: restore access, recalculate a score, review a warning or explain a pay difference. A works council, union representative, data protection channel or labour inspectorate may turn an isolated case into a structural review.
A family conversation about a first job#
Parents can help without monitoring or replacing the young worker. Before a contract is signed, they can read together which applications will be compulsory, which permissions they request and whether location is collected outside working hours. Simple questions reveal a great deal: how is an incorrect clock-in fixed, who allocates shifts, can one poor customer rating disable an account, is there a telephone number or only a form, and can someone speak to a person? The aim is not to create fear, but to make asking for clarity feel professional. Digitised work remains an employment relationship; accepting an application is not the same as surrendering privacy, rest or an explanation.
During the first weeks, a private and proportionate record of schedules, assigned tasks and important communications can help. There is no need to record everything or retain customer information. Note only discrepancies that would later be hard to reconstruct. If an application directs someone to work unsafely, the priority is to stop and follow prevention procedures, not obey in order to protect a metric. Persistent anxiety, fear of disconnecting, destructive competition or unpredictable income should not automatically be blamed on poor personal adjustment. They may reveal a problem in job design that requires collective discussion, health and safety attention or professional support rather than another productivity tip.
A useful opening sentence is: “I do not want to solve it for you; I want to help you organise what happened.” Then separate four layers: the verifiable event, the system’s interpretation, the consequence and the request. For example, GPS recorded a stop; the system labelled it inactivity; an incentive was lost; the worker requests review of the route and incident. This structure avoids an abstract argument about whether AI is good or bad. It also protects the young person’s autonomy: they choose whom to contact and which documents to share, unless there is an immediate safety risk or a serious violation that requires urgent assistance.
How an employer should buy and govern technology#
A responsible organisation begins by defining the problem without naming a product. If it wants to reduce turnover, it should investigate pay, schedules, management and learning before purchasing a resignation predictor. It then specifies which decision the tool will support, which data are excluded, which level of error is tolerable and which groups may be harmed. A pilot should include workers and their representatives, measure false positives and false negatives, and compare the proposed system with a less intrusive alternative. The supplier must explain limitations, version changes, security, subcontractors and audit options. “The model is proprietary” cannot prevent the employer from understanding the product well enough to meet its responsibilities.
Control continues after launch. Results should be monitored by location, shift, contract type and other relevant groups; complaints investigated; human overrides recorded; dangerous shortcuts created by metrics examined; and a retirement date established. A system may perform well in a pilot and deteriorate when the workforce, demand or environment changes. It may also achieve the intended output and remain unacceptable: increasing speed by causing injuries is not success. Governance dashboards should include quality, safety, equality, stability and worker experience. If production is the only measure, production is the only reality the optimisation can see, and every unmeasured cost is quietly transferred to people.
Participation improves both fairness and accuracy. The person performing a task knows exceptions that are invisible to a designer: an inaccessible address, a conversation that prevents a return, or a pause required to clean a machine safely. Using this knowledge means more than seeking symbolic approval at the end. Workers should help define objectives, test data and design challenges, with a genuine option to stop deployment. Managers and reviewers also need training: they must recognise automation bias, request independent evidence and explain why they confirm or overturn a recommendation. A final human signature has value only when it reflects independent judgement rather than habitual obedience to a score.
A simple test before accepting the output#
Apply five questions to any important automated employment decision. First, what exact decision is being made and who is accountable? Second, which data were used, and are they accurate, relevant and current? Third, what harm would an error cause, and can the measure be paused during review? Fourth, does the reviewer understand the context and possess authority to change the result? Fifth, is collective impact checked as well as the individual case? If an answer is missing, the process is not mature. The remedy may be better information, removal of a variable, a longer review period, a second reviewer or withdrawal of the system from that decision altogether.
The human side is not nostalgia for a workplace without technology. Entirely human decisions can also be arbitrary, opaque and discriminatory. A carefully bounded tool may reveal inconsistencies, distribute burdens better or prompt checks that people previously forgot. The difference lies in governance: legitimate purpose, suitable data, participation, explanation, effective review and repair. For someone beginning working life, learning to ask precise questions and preserve evidence is as important as mastering an application. For an employer, listening to a challenge is not an obstacle to innovation. It is essential information about whether the system is reliable, fair and worthy of continuing to influence people’s livelihoods.
