Marta’s first classification was not made by an algorithm. It was made by a form. At the municipal employment centre, she was told that a new platform would arrange appointments and recommend training paths. She opened the link on her phone while waiting for her son’s class to finish. The screen asked for “most recent job” and accepted only contracts, employers, and dates. For two years, Marta had cleaned homes by the hour, cared for her mother after surgery, and sold prepared food on weekends. There was nowhere to say so. She selected “not active” because it was the only answer that let her submit the application.
Fifteen days later, she received a recommendation for an introductory course and a distant appointment. The adviser saw an unusually thin story: inactivity, high availability, little recent experience. None of it described Marta’s life, yet the platform had treated the absence of a field as the absence of work, responsibility, or skill. The harm did not begin when a model found a distorted pattern. The person had already been flattened before the model received a single row of data.
A form decides who can be understood#
Conversations about algorithmic bias often begin at the end: who received an offer, who was excluded, or which group received a lower score. That scrutiny matters, but it comes too late when the intake form has already forced some lives into narrow boxes. Every field determines which experiences may count as evidence. Every closed menu creates a version of reality that is easy to process. Every impossible question transfers the cost of institutional simplification onto the person answering it.
Conversations about algorithmic bias often begin at the end: who received an offer, who was excluded, or which group received a lower score.
Informal work is not a statistical footnote. It can mean irregular earnings, unpaid help in a family business, delivery work, paid care without a contract, or employment interrupted by illness. Unpaid care is not empty time either: it demands organisation, knowledge, travel, negotiation, and attention. When a system recognises only what leaves a conventional administrative trace, it rewards people who have already been able to document their path and disadvantages people who kept a household running outside that lane.
The effect then travels through the system. An inactivity label may lower appointment priority, narrow the vacancies shown, trigger a “low engagement” alert, or prompt training that does not fit. The organisation later reads those outcomes as neutral observations about the individual. A loop has been built: the interface fails to collect a fact, the decision uses the gap, and the result appears to prove that there was nothing worth collecting.
Ask what the person cannot say#
An audit should not begin only with outcome tables. It also needs a test of language and possibility. What can someone answer if they work by the hour for several households? Where does a person record leaving paid work to care for a child or dependent relative? Can they explain that they have no payslip but do have references, receipts, skills, or returning customers? If the answer is “call this number” or “mention it at the interview,” the design has pushed essential information into a slower, less dependable channel.
The test should be run with people who know these circumstances, not just with the team that wrote the list of options. Watch where a person hesitates: whether “not active” sounds like an accusation, whether they fear that disclosing an irregular situation will close doors, whether they give up because they cannot tell what proof will be accepted. Those points of friction are design findings. They are not fixed by a friendly note at the bottom of the screen. They require changes to the fields, guidance, and rules that use the answers.
A careful solution does not require turning a life into an endless case file. It can provide broad categories, a short description box, and a route for alternative evidence. The key is to distinguish “not recorded,” “not asked,” “the person chose not to disclose,” and “not applicable.” Those are different states. Collapsing them into a silent zero is a policy choice, not an unavoidable technical limitation.
A correction needs a date and an owner#
When Marta found the problem, she clicked “request correction.” The confirmation said that her request would be logged, but it did not say who would read it, which record could change, or whether her appointment would be reconsidered before the assigned course began. A correction channel without a deadline can become a suggestion box: it exists, but it does not protect someone who needs a decision changed in time.
A useful process starts by returning a readable account of what the system understood. The person should see the fields used, identify the one that misrepresents their situation, and add an explanation in ordinary language. They should know what evidence is admissible even if no contract exists: a statement, references, a care schedule, receipts, or a conversation with a case worker. These materials will not carry identical weight, but none should be excluded simply because it does not resemble a payslip.
Then there must be a dated response. For instance: immediate confirmation of receipt, human review within five working days, and a decision before an exclusion is applied or a relevant vacancy closes. If the issue cannot be resolved in time, the prudent step is not to worsen the person’s position using the disputed data. That is not the same as accepting every claim without checking it. It means a questionable classification should not cause irreversible harm while it is being examined.
Ownership matters as well. “The system” cannot answer a question or explain a rule. The organisation must name a team or role able to amend the record, alter the recommended path, and document the reason. When a request is denied, the person needs to know what evidence was missing, what other evidence could help, and how to request a second review. A generic reply protects a workflow; a specific reply allows the workflow to learn.
Treat missing data as a warning signal#
The work does not end by adding a care option. The centre must measure what follows. It can record how many people use the correction route, which fields are challenged, how long each case takes, and whether recommendations change after information is completed. It should also compare abandonment and opportunity allocation for people on the standard path and people reporting non-standard work histories. If the second path takes longer, produces fewer appointments, or ends with fewer offers, inequality has merely moved location.
The data need careful handling. A category describing informal work can expose vulnerability if it is shared without limits. An audit should therefore collect only what is needed, restrict access, and separate supportive use from punitive use. Disclosing care responsibilities to explain a work gap must not become a reason to screen someone out on the assumption that they are unavailable. The stated purpose of a field must impose real limits on everyone who consults it.
Changes should be logged and revisited. If the form adds an option, someone should check three months later whether people can find it, understand it, and obtain an effective route from it. Decisions about vocabulary, evidence, and deadlines are not interface details; they are access policies. Leaving them unexamined allows a well-meaning improvement to become another hidden button.
The next practical step for an employment service is both simple and demanding: take a real application, ask five people with different work histories to complete it, and list every fact they could not express. For each absence, the team must decide whether to create a field, accept alternative evidence, or provide human review, then publish a correction deadline. Fairness does not begin when an algorithm calculates more accurately. It begins when nobody has to call herself “not active” in order to proceed.
