The Truth Supply Chain

This is an illustration, not a real case, and not any organisation I have worked with.

A disability program brings in an automated intake. A year later, two numbers sit side by side. Accommodation requests are down. Time to resolution is down. Both are reported as success, and both are accurate.

There is a third number nobody reports: how many people needed something and decided not to ask. The records cannot show it, because those people never entered them. Only asking the workers themselves can. And in this illustration, nobody asked.

That third number lives somewhere the reporting was never pointed — in somebody’s decision at a reception counter to say it’s fine, never mind. It may have risen while the other two fell.

I spent ten years in disability-management practice before I wrote any of this. I have written about disclosure, inspectability and refusal separately. They are one problem, and treating them separately is how it keeps getting missed.

One question under all of it

What does a system have to protect when it makes, supports or records a decision about a person?

Accuracy matters. But first the system has to protect its supply of truth, and that supply has three links.

The truth has to be speakable: a person has to be willing to say it.

It has to be kept within limits: the system has to take only what it needs and decide only what it should.

And it has to be checkable: the person has to be able to see what was done with it, and say so when it is wrong.

That path, from what a person is willing to say, to what a system takes and decides, to what they can check afterwards, is what I mean by the truth supply chain.

Break any one link and the system keeps running. It keeps producing files and decisions. It just stops being about the person.

Link one: disclosure comes before data

A disability file draws on more than the person. There are physicians’ statements, payroll records, absence triggers, independent assessments. But most of what the system knows — including what reaches the doctor — starts with what the person chose to say, and when.

The file is what is left after that person decided how much it was safe to say.

The research on that decision is largely Canadian, from Monique Gignac and her colleagues at the Institute for Work & Health, and it is sobering. In a survey of 882 Canadian workers drawn from a national panel, all with chronic conditions that affected their work, a quarter had told neither their supervisor nor their co-workers (Woticky et al., 2025). In another sample of 591, 42 percent had not told their supervisor, and their decisions turned on their own goals and the culture around them more than on how much their health was affecting the job or what support was available (Gignac et al., 2025).

The finding I keep coming back to is the most recent. Gignac, Pienkowski and Bowring (2026) asked 695 workers not only whether they had disclosed to a supervisor but whether they were hiding anything. About 45 percent were actively hiding health information. Roughly one in five had disclosed and were still hiding part of it — and that group reported more absenteeism, worse productivity and more job disruption.

Disclosure comes in parts. People tell some of it, and which part depends on what they believe happens next.

That makes this a problem of what gets generated, before anything is recorded. Every line in a file can be accurate, and the file can still be missing the cases that matter most — because the people they belong to decided, in advance, what the system could be trusted with.

I worked this link through in an earlier piece. The short version: the person’s decision to speak is the input. What the system does afterwards is processing.

Link two: what the system will not ask for, and will not decide

The second link, what a system will ask for and what it will decide, is the one that gets designed by accident.

Every system that touches a person’s account makes choices, written down or not: what it will ask for, what it will decide, and where it hands the decision back to a human being. Most are built to want more on all three. More data is assumed to mean better decisions.

In a disability process that assumption deserves suspicion, because what a system asks for is itself a disclosure condition. A form that wants a diagnosis when it only needs a functional limitation. An intake that records every word when it needs three facts. Each tells the person on the other side something about what happens to what they say, and they adjust what they say accordingly.

The research here is early, but it points this way. Monitoring that is more transparent and less invasive draws better reactions from workers than monitoring that is not (Ravid et al., 2023). Where workers see real organisational support for disclosing a mental-health concern, they are more willing to disclose and more likely to do it (Lyubykh et al., 2025). One recent paper argues for making common supports — flexible hours, quiet space, a choice of format — available to everyone, so disclosure is needed only for what is genuinely individual (Kalmanovich-Cohen & Stanton, 2025). And a study of the people on the employer’s side of accommodation — supervisors, disability managers, union and occupational-health representatives, labour lawyers — starts from the bind they work in: believing they would benefit from health details, and bound by law to keep them private (Gignac et al., 2021).

Asking for less eases that bind. My argument — and it is an argument — is that it is also part of what makes a system safe enough to tell the truth to.

The other half of this link is what a system will decide. I built one tool to a rule about that. It answers disability and return-to-work questions, and it is built to refuse to answer without a citation. Not to prefer one: to refuse. A confident guess about something that affects a person’s livelihood is worse than no answer. I mention it once, as evidence that a refusal can be built in, not as a recommendation.

So I count refusal as part of the product.

Link three: inspectability, not explanation

The third link is what happens after.

An explanation tells a person what was decided. Inspectability gives them a way to check whether it holds up: the rule applied, the information it was applied to, the moment of judgement, what was left out, how to challenge one step, and what changes if they are right.

I set that test out in Inspectability. Its place in this chain is simple. A person who cannot see what was done with what they said has little reason to say more next time. Inspectability is a fairness standard for the decision in front of you. It is also what keeps link one open for the next.

How the chain could break all at once

Suppose a disability process brings in automation. Here is what could move.

It would change who might read what a person says. It would change how long that lasts and where it can travel. It would change whether there is still a human on the other end, and whether the person believes they could correct the record.

Few people would stop disclosing outright. Some might delay. Some might tell only the part that will not raise a flag, or take the day as a personal day instead of a sick day, or phone a manager they trust instead of using the channel the system can see.

The records would get shorter and more consistent. Requests would fall. Every number the automation was bought to improve would improve.

Then suppose somebody audits the system, checking whether its decisions were accurate and fair on the records it had.

They might well be. The missing truth would be in neither the model nor the audit, because both start from the same records. The audit’s blind spot is built in: it can only see what was recorded.

There is an old warning about this kind of number. Donald Campbell put it in 1979: the more a number is used to make social decisions, the more pressure there is to corrupt it — and the more it distorts the very work it was meant to monitor. In a disability program, nobody has to cheat for that to happen. Once lower request counts are the proof that the automation worked, nobody has a reason to ask why they fell. And what fell is where early intervention starts.

The research on being watched at work points the same way, though it does not measure disclosure, and I will not pretend it does. Electronic monitoring has not been shown to improve performance and is linked to more stress (Ravid et al., 2023). In a national Canadian sample, feeling surveilled was linked to distress through job pressure, lost autonomy and a sense of privacy violated (Glavin, Bierman & Schieman, 2024). The closest study on speaking up, of 351 employees and their supervisors at one large Chinese platform company, found that being directed, scheduled and monitored by an algorithm made people less likely to speak up with ideas for change, while algorithmic feedback made them more likely to (Lin et al., 2026). The design decided which way it went. Those are conditions under which I would expect people to say less. They are not proof that they do.

The same chain outside disability

The first link is not special to disability files. Any process that makes telling the truth costly will start to hear less of it.

Take a complaints process. A person is being treated badly at work, sets a boundary, and the other party escalates. By the time anyone looks, it reads as a two-sided conflict, and the process is built to treat it as one. The escalation is what makes it look mutual. And the policy can end up rewarding whoever stays composed, which is rarely the person being targeted.

In a process built that way, the target learns what happens to what they say, and may say less next time. So may the people who watched. A quieter complaint log can then be read as improvement.

No automation is needed for that, and that is the point. Automation does not invent this failure. It can scale it, and make it harder to see.

Where this could be wrong

The evidence does not all point one way.

People sometimes tell a machine more than they would tell a person. When participants in a health screening believed the interviewer was automated, they disclosed more (Lucas et al., 2014). In consumer settings, people share sensitive information with AI agents more readily because they expect a machine not to judge them (Kim et al., 2022). One study found no difference in how intimately people disclosed to a chatbot and to a human, though people trusted the human more (Croes et al., 2024). Another found that making an agent look more human made people less willing to answer sensitive health questions (Zhu & Broadbent, 2025). And in the workplace study closest to this essay’s setting, after using a digital decision aid, more than a third of the 100 workers who took part in its evaluation decided to disclose a mental-health condition at work (Stratton & Glozier, 2024).

Automation can raise disclosure as well as lower it.

What the research does show is where the comfort runs out: at the question of who else will know. In Kim’s studies, the advantage of telling an AI reversed when people expected it to share what they said with others. More directly, Lipsey and Shepperd (2019) found that people were more likely to avoid learning their own health risks when an employer or an insurance company might see the result than when only researchers would. And in a small set of interviews with disabled people who use AI assistants — not yet peer-reviewed — participants weighed every disclosure against two listeners at once, the assistant and the company behind it, and described an assistant’s memory letting something said once drift into places it did not belong, even as it spared them from repeating it (Taheri et al., 2026).

A disability file is the case where the employer or the insurer does see it. It can be read, scored, sent to a carrier, used to decide. The question a worker faces there is what happens to me once this is written down.

That is the actual claim. Disclosure depends on what people believe happens next, and automation changes that belief.

Stated so it can fail, the prediction is directional. Where automation widens who can read a disclosure, keeps it longer, and removes the human channel, people will disclose later and less completely. Where it adds plain notice, keeps a human channel and gives a real right to correct, they will not.

Here is what would show I am wrong:

• In programs that widened the audience and removed the human channel, the timing and completeness of disclosure hold steady, compared with similar units that did not automate.
• Programs that added plain notice, a human channel and a right to correct see no difference in what people are willing to say.
• An anonymous survey of workers finds no more hiding after automation than before.

Each of those can be checked. None of them can be checked from the records alone.

The first question

Everyone is asking whether the model is fair. The question before it is this: do people still tell the system what it needs to know?

That is answerable, and the standard it implies runs across all three links.

Measure disclosure before and after, and not only in the records. Requests, early self-reports and time from first difficulty to first request, for the year before deployment and the year after, compared with a unit that did not change. And an anonymous survey asking workers what they held back, because partial disclosure never shows up in a count. If disclosure fell, that is the first thing to explain.

Say what the tool does, in plain words, before anyone has to ask. What it reads, what it does with it, who sees the output, how to challenge it. People deciding under known conditions decide differently from people who have heard there is “some AI thing now.”

Ask for less. Collect what the decision needs. Make common supports available without a diagnosis.

Keep a human channel, and make sure using it does not mark anyone.

Show what is missing. A system that produces a confident answer from an incomplete file should say what it does not know.

Make correction possible, one step at a time, quickly.

Anyone can do this without understanding a model. What it takes is someone who has decided that the people a system is about are part of the system.

The verdict

We judge systems by what they do with the information they are given.

That is half the test. A system also has to be judged by what its presence makes people unwilling to say — because that is the information it will never be given, and no audit of its records will ever find it.

Automate the paperwork. Do not automate away the reason somebody was willing to open the file.

References

Campbell DT. Assessing the impact of planned social change. Evaluation and Program Planning. 1979;2(1):67-90. doi:10.1016/0149-7189(79)90048-x

Croes EAJ, Antheunis ML, van der Lee C, et al. Digital confessions: the willingness to disclose intimate information to a chatbot and its impact on emotional well-being. Interacting with Computers. 2024;36(5):279-292. doi:10.1093/iwc/iwae016

Gignac MAM, Bowring J, Jetha A, et al. Disclosure, privacy and workplace accommodation of episodic disabilities: organizational perspectives on disability communication-support processes to sustain employment. Journal of Occupational Rehabilitation. 2021;31(1):153-165. doi:10.1007/s10926-020-09901-2

Gignac MAM, Bowring J, Saunders R, et al. Why are disclosure decisions so difficult? Understanding factors that encourage and discourage workers with a chronic disabling condition from disclosing health information at work. Journal of Occupational Rehabilitation. Published online August 29, 2025. doi:10.1007/s10926-025-10326-y

Gignac MAM, Pienkowski M, Bowring J. Working with a chronic health condition: is choosing not to disclose the same as hiding information? Journal of Occupational Rehabilitation. Published online April 24, 2026. doi:10.1007/s10926-026-10387-7

Glavin P, Bierman A, Schieman S. Private eyes, they see your every move: workplace surveillance and worker well-being. Social Currents. 2024;11(4):327-345. doi:10.1177/23294965241228874

Kalmanovich-Cohen H, Stanton SJ. Moving beyond disclosure: rethinking universal support for neurodivergent employees. Frontiers in Psychology. 2025;16:1547877. doi:10.3389/fpsyg.2025.1547877

Kim TW, Jiang L, Duhachek A, et al. Do you mind if I ask you a personal question? How AI service agents alter consumer self-disclosure. Journal of Service Research. 2022;25(4):649-666. doi:10.1177/10946705221120232

Lin W, Zhang M, Zhang W, et al. Will employees still speak up under algorithmic management? The differential effects of distinct algorithmic functions—evidence from the Meituan platform in China. Systems. 2026;14(5):569. doi:10.3390/systems14050569

Lipsey NP, Shepperd JA. The role of powerful audiences in health information avoidance. Social Science & Medicine. 2019;220:430-439. doi:10.1016/j.socscimed.2018.11.037

Lucas GM, Gratch J, King A, et al. It’s only a computer: virtual humans increase willingness to disclose. Computers in Human Behavior. 2014;37:94-100. doi:10.1016/j.chb.2014.04.043

Lyubykh Z, Turner N, Weinhardt JM, et al. Facilitating mental health disclosure and better work outcomes: the role of organizational support for disclosing mental health concerns. Human Resource Management. 2025;64(5):1243-1262. doi:10.1002/hrm.22310

Ravid DM, White JC, Tomczak DL, et al. A meta-analysis of the effects of electronic performance monitoring on work outcomes. Personnel Psychology. 2023;76(1):5-40. doi:10.1111/peps.12514

Stratton E, Glozier N. Adaptation and implementation of an employee mental health disclosure decision aid tool in a real-world sample. Translational Behavioral Medicine. 2024;14(3):172-178. doi:10.1093/tbm/ibad072

Taheri A, Tazike M, Carrington P, et al. When disability disclosure travels: memory, privacy, and contextual integrity in conversational AI. arXiv. Preprint posted online September 19, 2026; not peer-reviewed. doi:10.48550/arXiv.2609.22720

Woticky G, Jetha A, Tompa E, et al. Disclosure decisions of workers living with a chronic health condition causing disability at work: are decisions to disclose to co-workers and supervisors different? Journal of Occupational Rehabilitation. 2025;35(3):675-687. doi:10.1007/s10926-024-10235-6

Zhu T, Broadbent E. Effects of a computer’s human likeness on disclosure in psychosocial assessments: a randomised trial. Computers in Human Behavior. 2025;169:108683. doi:10.1016/j.chb.2025.108683