JD Miller
Menu

Library / AI in practice

AI in practice

Duty of care in the AI transition

Displacement research, mental-health baselines, and who decides what a model may say, read together as one operating obligation.

By JD MillerPublished · 15 min read

Listen to this piece read aloud in a recorded voice, narrated by Daniel.

I run IT, security, infrastructure, and AI adoption for a small investment firm. Most weeks that means deciding which workflows get a model in the loop, who is allowed to authorize it, and what we do when the output is wrong.

But the more of it I do, the harder it is to keep three questions apart. What does this technology do to jobs? How do people actually use it during a working day? And who decided what the model is allowed to say back?

This essay holds those three together under one heading: duty of care. Every number below links to the page it came from, with the year in the sentence. Where the research measures association rather than cause, I say so. Where I could not confirm a figure on its primary source, I left it out.

One thing I will not do is claim that AI causes suicide. The chain I am willing to defend is narrower: job loss and job insecurity are associated with worse mental-health outcomes in a large body of research; AI is a plausible new driver of displacement; therefore the people who design, deploy, and govern these systems carry a duty of care that most of us have not written down yet.

Source

The mental-health baseline we are starting from

These are the numbers a duty of care has to be measured against.

In the United States, the age-adjusted suicide rate was 14.1 per 100,000 in 2023, and the National Center for Health Statistics reports that it "did not significantly change" from 14.2 in 2018. That is death-certificate data, which is about as hard as public-health numbers get, but it says nothing about why.

Depression is far more common than suicide. Between August 2021 and August 2023, NCHS found that 13.1 percent of Americans aged 12 and older had depression in the previous two weeks. That figure comes from a self-report screening instrument, the PHQ-9, so it measures symptoms rather than clinical diagnoses.

Globally, the World Health Organization estimates that more than 720,000 people die by suicide each year; its regional breakdowns refer to 2021. Many countries lack complete death registration, so this is a modelled estimate rather than a count. WHO's depression fact sheet puts the 2023 global prevalence at about 4 percent, roughly 322 million people.

None of these numbers mention AI. That is the point. They describe the population that any large labor-market shock lands on.

Source

What the research says unemployment does to people

This is the load-bearing link in the chain, so I want to be precise about what each study measured.

Paul and Moser's 2009 meta-analysis in the Journal of Vocational Behavior pooled 237 cross-sectional and 87 longitudinal studies. They found 34 percent of unemployed people showed psychological problems compared with 16 percent of the employed, with an overall effect size of d = 0.51. Most underlying studies were cross-sectional, which limits what can be said about direction.

Milner, Page, and LaMontagne's 2013 systematic review in PLOS ONE identified 16 eligible studies, pooled six cohort studies, and found the greatest suicide risk within five years of becoming unemployed, a relative risk of 2.50 with a 95 percent confidence interval of 1.83 to 3.17 against the employed population. The elevated risk attenuated over time but persisted. Pooling heterogeneous designs across countries is the main weakness.

Nordt and colleagues' 2015 analysis in The Lancet Psychiatry modelled suicide and unemployment across 63 countries from 2000 to 2011 and found unemployment associated with a 20 to 30 percent increased relative risk of suicide in every world region they studied. This is a country-level model; it cannot tell you what happened to any individual, or separate job loss from whatever else moves with it in a recession.

Case and Deaton's 2015 paper in PNAS documented rising midlife mortality among white non-Hispanic Americans from 1999 to 2013, "largely accounted for by increasing death rates from drug and alcohol poisonings, suicide, and chronic liver diseases and cirrhosis." It is restricted to one demographic group whose experience differed from others. It matters here because it is the best-documented case of economic dislocation and despair moving together over a long period.

Read together, the association between losing work and worse mental health is consistent across decades, methods, and countries. It is not proof of cause in any single case, and the effect sizes are moderate. That is why it belongs in a duty-of-care argument rather than a headline.

Source

What displacement looks like, and who it reaches

Displacement splits along several factors, and each one changes who bears the cost.

Start with the baseline that exists regardless of AI. The Bureau of Labor Statistics' Worker Displacement release covering 2023 to 2025 counted 3.3 million workers displaced from jobs they had held at least three years. By January 2026, 66.1 percent were reemployed. Of those back in full-time wage work, about 49 percent were earning as much or more than at the lost job, down from 62 percent in the 2024 survey. The 3.3 million count covers only workers with three or more years of tenure; another 4.1 million short-tenured workers were displaced over the same period. The earnings figure is the one I keep returning to. Getting a job back is not the same as getting your wage back.

Task exposure versus occupation exposure. Eloundou, Manning, Mishkin, and Rock's 2023 paper estimated that around 80 percent of the US workforce could have at least 10 percent of their tasks affected by large language models, and about 19 percent could see at least half of their tasks affected. This measures which tasks a model could plausibly do, not which jobs will disappear. A job that is 40 percent exposed may get busier rather than smaller.

Wage level. The same paper found higher-income jobs more exposed. Brookings' 2024 analysis of the same exposure data found more than 30 percent of workers could see at least half of their occupation's tasks disrupted, concentrated in higher-paying fields such as computer work, management, engineering, finance, and law, with 36 percent of women in high-exposure occupations compared with 25 percent of men.

Tenure and age. This is where the observed data, as opposed to exposure estimates, is most striking. Brynjolfsson, Chandar, and Chen's Canaries in the Coal Mine analysis of ADP payroll records, updated in August 2026, finds employment of 22 to 25 year olds in the most AI-exposed occupations about 19 percent below where it would be relative to less-exposed peers, while experienced workers in the same occupations held flat or grew. The authors call the result descriptive rather than causal, and note the gaps are larger in the ADP sample than in national benchmarks. The Federal Reserve Bank of New York's tracker showed recent college graduates' unemployment "elevated at about 5.6 percent" in its second-quarter 2026 update.

Firm size and sector. The Census Bureau's real-time business survey, analyzed in an NBER working paper in 2024, found the share of US firms using AI to produce goods or services rose from 3.7 percent to 5.4 percent between September 2023 and February 2024, higher among large firms and in the Information sector.

Augmentation versus substitution. Brynjolfsson, Li, and Raymond's Generative AI at Work, first circulated in 2023, studied 5,179 customer-support agents and found a 14 percent average productivity gain, 34 percent for novice and low-skilled workers, and minimal impact on experienced staff. That is augmentation. Whether a firm uses the gain to serve more customers or to employ fewer agents is a management decision, not a property of the model.

Geography. I could not confirm a primary-source figure on which regions are most exposed, so it sits in the open questions below.

Synthesis

The chain I am willing to defend

The displacement literature says AI exposure skews toward early-career workers, higher-wage cognitive work, and larger firms first. The BLS data says even ordinary displacement leaves about half of reemployed workers earning less. The mental-health literature says unemployment is associated with roughly double the suicide risk in the first years and with about twice the rate of psychological problems, across many countries and methods.

I do not need AI to cause anything directly for this to matter. If AI shifts the composition of who gets displaced, and displacement carries a known mental-health burden, then the burden shifts with it, toward people who may have less experience navigating a job loss and fewer years of savings behind them.

What is different this time is where the design choices sit. The augmentation-versus-substitution decision is made by the firm deploying the tool. The exposure profile is shaped by which tasks vendors optimize for. Neither the young analyst nor the support agent gets a vote. That asymmetry is the basis for a duty of care.

Source

How people interface with these systems day to day

The governance question only matters if people are using the systems. They are.

Pew Research Center found in early 2025 that 34 percent of US adults had used ChatGPT, and that 28 percent of employed adults used it for work, up 20 points from 2023. In a separate 2025 report, 51 percent of the public said they were more concerned than excited about AI, against 15 percent of AI experts.

Gallup reported that the share of US employees using AI at work at least a few times a year rose from 40 percent to 45 percent between the second and third quarters of 2025, with frequent use at 23 percent and daily use at 10 percent, across a panel of 23,068 employed adults.

Microsoft's 2025 Work Trend Index found 53 percent of leaders saying productivity must increase while 80 percent of the global workforce reported lacking the time or energy to do their work. Vendor surveys deserve a discount, but the shape matches what I see: pressure from above, exhaustion below.

The most sensitive interface is people bringing their mental state to a chatbot. The evidence here is young and mixed. A 2025 randomized trial in NEJM AI of a purpose-built therapy chatbot with 210 participants found significant reductions in depression and anxiety symptoms over four weeks, with participants recruited through online ads and follow-up limited to eight weeks. A 2025 study at ACM FAccT found general-purpose models expressed stigma toward mental-health conditions and responded inappropriately to some critical conditions, including delusional thinking. And a 2025 evaluation in Psychiatric Services, with 13 clinicians rating 9,000 chatbot responses, found the models aligned with expert judgment on very-low and very-high risk questions but were inconsistent at intermediate risk levels.

Intermediate risk is where most real conversations live. That finding shapes the next section.

Source

Who decides what safe output is

When a model refuses, redirects, or answers, someone made that rule. It is worth knowing who.

Vendors publish some of it. OpenAI's Model Spec sets a chain of command in which "instructions with higher authority override those with lower authority," running from platform to developer to user to guideline, and directs the model toward empathetic responses with crisis resources rather than flat refusal on self-harm topics. Anthropic's Usage Policy, effective September 15, 2025, requires for high-risk uses that "a qualified professional in that field must review the content or decision prior to dissemination or finalization." Its constitution, published January 22, 2026, sets a priority order of safety, ethics, its own guidelines, then helpfulness.

Regulators set the outer boundary. The EU AI Act entered into force on August 1, 2024, defines four risk levels, and requires that when using chatbots "humans should be made aware that they are interacting with a machine." Prohibitions applied from February 2, 2025, general-purpose model obligations from August 2, 2025, and rules for high-risk systems in areas such as employment and education now apply from December 2, 2027, with AI built into regulated products following on August 2, 2028. In the US, NIST's AI Risk Management Framework from 2023 organizes the work into Govern, Map, Measure, and Manage, and is explicitly voluntary.

Researchers test the edges. The XSTest suite from 2024 uses 250 safe prompts that superficially resemble unsafe ones, alongside 200 unsafe contrasts, to measure exaggerated safety behavior: models refusing help that should have been given.

Synthesis

Safe output is a governance question, not a vendor scorecard

I am not interested in ranking vendors on this. What strikes me reading the specs, the law, and the studies side by side is the structure.

Who decides: vendors write the spec, regulators draw the outer line, the deploying firm sets its own policy, and the person at the keyboard decides last, if at all. How visible the rules are: some specs are public, most enterprise policies are not, and users rarely see either. What an appeal looks like: for most people, nothing. A refusal is final and unexplained.

And what happens when a refusal blocks legitimate help: the Psychiatric Services finding on intermediate-risk inconsistency and the XSTest work on over-refusal describe the same problem from both sides. A model can fail by saying too much and by saying too little, and the cost of the second failure lands on the person who needed the answer.

That is a governance problem. It has owners, evidence, and recovery paths, or it should.

Proposal

What I would put in place at a firm like mine

Each of these is testable inside a small firm. If a year from now none of them changed a decision, I was wrong about their value.

Keep a displacement register. For every AI workflow we approve, record whether it augments or substitutes, whose tasks it touches, and their tenure. Review it quarterly. The test: does the register ever change what we approve?

Decide redeployment before substitution. If a workflow will remove hours from a role, name where those hours go before it ships. The test: count the workflows that shipped without a named answer.

Log refusals and give them an appeal path. Internal assistants should log every refusal or redirect, and someone should read the log monthly. The test: how many refusals blocked legitimate work, and did the policy change?

Route crisis language to people. Any internal assistant that encounters distress should surface 988 and our employee assistance program and hand off. This is not a mental-health tool; it is a door. The test: run the intermediate-risk prompts from the clinical literature against our configuration and read the results.

Measure the whole task, not the model. I argued this in Beyond the model frontier and it applies here: the productivity gain is real, the question is who absorbs it.

Open question

What the evidence does not settle

Whether AI displacement carries the same mental-health burden as past displacement. Every study above predates the current wave; the association may be stronger, weaker, or different in shape for higher-wage, early-career workers.

Whether task exposure converts into job loss at all. Exposure estimates and payroll data point the same direction for young workers, but the canaries study is explicitly not causal.

Geography. I could not confirm a primary-source figure on regional exposure, and a firm with one office in one city should care about the answer.

Whether chatbots net help or net harm at intermediate risk. One trial, one audit, and one clinical evaluation are not a literature.

Whether a published spec changes behavior in production. A spec describes intent. The refusal log is how you find out what happened.

I do not have those answers. I have a register, a log, and a door. For an operator, that is where a duty of care starts.

Source ledger

All sources accessed September 19, 2026, and each figure re-checked against its page on September 20, 2026.

Continue reading

To hear about new pieces by email, subscribe.

Write to me about this piece: hellojdmiller@gmail.com