Systemic Coping Syndrome

The people who stand to gain most from AI are often the people least placed to act on it, and the organisations they work in are often the least able to see what it is for. The best-known experiments find that the largest gains from generative AI go to the less skilled and the less expert. The evidence on who actually uses it is more mixed, and the mixed part is revealing: in the largest linked survey I know of, almost every worker was aware of ChatGPT and use was widespread among younger and less experienced workers, yet the people using it had already been earning slightly more than their peers at the same tenure, and the barriers that workers named for not using it, even where they expected it to halve their time, mostly pointed at their employers. Telling workers how much time it would save them did not change their behaviour. I think much of that gap is not about individuals at all. It is about the organisations they work in, and I have started calling the condition behind it systemic coping syndrome:

Systemic coping syndrome is an organisational pattern in which coping with persistent dysfunction displaces attempts to correct it, and the coping practices themselves help reproduce the dysfunction.

Put more briefly, the organisation has become better at accommodating its problems than at fixing them.

The argument of this essay is that systemic coping syndrome does three things at once. It hides the opportunity at the level where it is largest, because the people who would benefit are busy absorbing the dysfunction that AI could remove, and nobody is asking them which steps should not exist. It blocks the organisational changes that turn one-off individual gains into compounding ones. And, perversely, when AI does arrive in such an organisation, it tends to be used to cope faster rather than to stop coping, which makes the dysfunction cheaper to live with and therefore less likely ever to be fixed.

I should say up front that I build AI systems for law firms, so I have an interest in organisations adopting AI, and you should read what follows with that in mind. I have tried to compensate by citing the evidence that cuts against me, by naming the rival explanations, and by stating predictions that could prove me wrong. None of the remedies at the end requires buying anything.

From learned helplessness to systemic coping

The term borrows from, and deliberately extends, learned helplessness, so the starting point is what that term means and what its discoverers now say about it.

In 1967 Martin Seligman and Steven Maier reported that dogs exposed to electric shocks they could not control later failed to escape shocks that they could have escaped (Seligman and Maier, 1967, Journal of Experimental Psychology 74(1), 1-9). Their theory was that the animals had learned that nothing they did mattered, and that this learning undermined trying.

Nearly fifty years later the same two researchers published a review whose abstract contains a sentence that every manager should read twice: "the original theory got it backward" (Maier and Seligman, 2016, Psychological Review 123(4), 349-367). On the neuroscience, passivity in response to prolonged aversive events is not learned. It is the default. What is learned is control: the detection that one's actions affect outcomes, which then switches the default passivity off.

I am going to be careful with that finding, because it comes from animal neuroscience and an organisation is not a shuttlebox. I use it as an analogy that changes the question, not as evidence about offices. The question it changes is the right one, though. If passivity is the default and agency is what has to be learned, then an organisation in which people rarely see their actions change anything does not need to teach anyone to be helpless. It only needs to fail to teach them that they are not.

Researchers moved the concept up a level long ago. Martinko and Gardner proposed learned helplessness as an explanation for performance deficits in organisations in 1982 (Martinko and Gardner, 1982, Academy of Management Review 7(2), 195-204). Uzbay and Erdoğan have recently proposed a framework for collective learned helplessness, a society-wide loss of belief that circumstances can change (Uzbay and Erdoğan, 2025, Theory and Society 54(5), 879-890). And an ethnographic action-research study of Holacracy adoption found that formally distributing authority was not enough, because people conditioned by hierarchy did not acknowledge or use the authority they had been given until deliberate practices helped them unlearn the habit (Valkiainen, 2026, Human Relations 79(11), 1368-1400).

The layers look like this:

The last clause of the third row is what makes it systemic. Helplessness describes a state of mind. Systemic coping syndrome describes a feedback loop, and the loop runs through the work itself rather than through anyone's morale. "Syndrome" is a metaphor here, not a clinical claim. Nobody is being diagnosed with anything.

What coping looks like from the inside

The most important feature of systemic coping syndrome is that it does not look like passivity. People inside it are often extremely busy, and they are busy being conscientious.

The cleanest empirical description I know of comes from Anita Tucker and Amy Edmondson's study of nurses in nine hospitals (Tucker and Edmondson, 2003, California Management Review 45(2), 55-72). When supplies were missing or information was wrong, the nurses they observed usually fixed the immediate problem themselves, by fetching the item, chasing the information or working around the gap, and then carried on. Tucker and Edmondson call this first-order problem solving. It is competent, caring and fast. It also means that the underlying process does not change, because nobody upstream learns that it failed. The paper's title states the consequence: hospitals don't learn from failures.

Chris Argyris described the same shape decades earlier as the difference between single-loop learning, which corrects an error within the existing rules, and double-loop learning, which questions the rules (Argyris, 1977, "Double Loop Learning in Organizations", Harvard Business Review 55(5), 115-125). His later work on organisational defensive routines describes habits that protect people from embarrassment and threat while concealing the causes of the problem, and thereby block the second loop (Argyris, 1990, Overcoming Organizational Defenses, Allyn and Bacon).

In a coping organisation you see things like these:

1. Workarounds that are maintained and taught to new joiners, instead of the broken step being removed.
2. Expectations that are managed rather than outcomes that are improved.
3. Approvals sought to share blame rather than to improve a decision.
4. Good managers who shield their teams from the dysfunction rather than challenge its source.

Every one of those is individually rational. Every one of those consumes exactly the time and attention that would be needed to notice that the dysfunction could now be removed. And every one of those makes the dysfunction a little more survivable, which removes a little more of the pressure to fix it. That is the loop.

Why the people who would benefit most see it least

Now put AI into that picture.

Who benefits. In a study of 5,172 customer support agents, access to a generative AI assistant raised issues resolved per hour by 15% on average. For agents in the lowest skill quintile the gain was 36%, while for the most skilled agents there was no significant productivity change (Brynjolfsson, Li and Raymond, 2025, Quarterly Journal of Economics 140(2), 889-942). In a field experiment with 758 Boston Consulting Group consultants, on tasks inside the model's capability frontier, those below the average performance threshold improved by 43% and those above it by 17% (Dell'Acqua et al., 2023, Harvard Business School Working Paper 24-013). In a preregistered experiment with 453 professionals doing writing tasks, ChatGPT cut time by 40%, raised quality by 18%, and reduced inequality between workers (Noy and Zhang, 2023, Science 381(6654), 187-192).

Who uses it. A survey of 18,000 Danish workers in 11 occupations exposed to generative AI, linked to national register data, found ChatGPT use to be widespread, "especially among younger and less-experienced workers". The same paper found that women were 16 percentage points less likely to have used it for work, and that "despite its potential to lift workers with less expertise, users of ChatGPT earned slightly more already before its arrival, even given their lower tenure" (Humlum and Vestergaard, 2025, PNAS 122(1), e2414972121). The authors read the earnings finding as "more able workers are more likely to use ChatGPT", and note that years of schooling and school grades were not correlated with use.

The same paper is even more useful on why people did not use it. Almost everyone was aware of the tool. Among workers who believed it could halve the time a task took them, only about 23% planned to use it within the next two weeks. Asked why they did not intend to use it despite that potential, 42% said they needed training and 36% said their employer actively restricted their use, while existential fears about redundancy or dependence were the least cited reason, at under 9%. And when the researchers randomly informed workers about the time savings, usage did not shift, which they read as confirming that these or other frictions are causally holding adoption back. Employer restrictions were more likely to bind in occupations handling sensitive information, legal professionals among them.

I want to be precise about what that adds up to. The adoption gap does not run along experience: less experienced workers were, if anything, more likely to use the tool. It runs along prior earnings at a given tenure, which the authors interpret as ability, and along gender, which my argument does not explain. The individuals mostly did see the opportunity, and information about its size changed nothing. What stood between them and it were an organisational decision (restrictions) and a gap that organisations are at least partly responsible for closing (training). The evidence does not prove my thesis. It does show the shape that the thesis predicts: the benefit is largest at the bottom of the skill distribution, adoption tilts towards people who were already doing well, and the reasons people give for not using it point at their employers.

Systemic coping syndrome explains that shape through three mechanisms, each of which is a property of the organisation rather than of the person.

Coping consumes the slack that noticing requires. Knowing that a tool exists is not the same as noticing which of your own steps it could remove. That takes spare attention: a moment to wonder whether this weekly task needs to exist at all. The people carrying the most workarounds have the least of that attention left. My hypothesis, which I have not seen measured, is that in a coping organisation the workaround load falls disproportionately on the junior and on the people closest to the broken process, which is exactly where the benefit evidence says the gains are largest.

A history of non-contingency teaches people that proposals go nowhere. If the last three suggestions you made disappeared into a committee, the rational estimate of the value of a fourth is low. By analogy with the neuroscience above, and only by analogy, what is missing may not be a belief that was taught but an experience of control that never happened. The closest real evidence is the Danish information experiment: telling people about the gains did not move them, which is what you would expect if the obstacle is not knowledge. Noy and Zhang's experiment contains a second small data point that is consistent with this. Participants exposed to ChatGPT during the experiment were twice as likely to report using it in their real job two weeks later, and 1.6 times as likely two months later. The experiment was not designed to test my claim, and the same result is equally consistent with the simpler explanation that people need training, so I count it as compatible, not confirming.

Permission is withheld at the organisational level. Blanket restrictions, unclear policies and the absence of approved tools are decisions made by organisations. Individuals in a coping organisation do not challenge them, for the reasons above.

None of that requires anyone to be lazy, unintelligent or afraid of technology. It requires only that the organisation keep rewarding endurance over correction.

Why the loss is organisational, and why it compounds

Here is the part that I think is most often missed, and the reason why "learned helplessness" alone is the wrong frame.

It helps to separate two things. The first is the one-off gain: a person uses AI and a task takes less time. The studies above measure that, and in a coping organisation it is reduced but not eliminated. The second is the capacity to compound: the organisation's ability to keep finding and removing the next dysfunction, using what it gained from removing the last one. Systemic coping syndrome costs some of the first and most of the second, and the second is where AI's organisational value lies, for three reasons.

Reinvestment. When a team removes a workaround it gets back time. If that time is spent finding and removing the next workaround, each improvement funds the search for the next, and the gains build on each other. If the freed time is absorbed by more coping, the gain is a one-off. Two organisations with the same tools and the same initial gain can therefore end up in very different places, and the difference is the reinvestment rate, not the tool. That is a claim about mechanism, not a measurement: I do not know of a study that has measured the reinvestment rate directly, and the size of the divergence depends on it.

Non-rivalry and network effects. A prompt, a template, an agent, a checked workflow or a cleaned dataset that one person builds can be used by everyone else without being used up. That property is non-rivalry, and it means that the value of building something scales with the number of people who will use it. On top of that there are genuine network effects, where each additional user makes the tool more valuable to the others: a structured handover drafted with AI is only useful if the team receiving it can consume it, and a shared knowledge base gets better as more people correct it. The Brynjolfsson, Li and Raymond paper points at a version of this inside a single firm. The assistant was trained to distinguish the successful behaviours of the firm's top performers, and the authors hypothesise that part of the effect comes from the system embodying the best practices of high-skill workers, including tacit ones, and passing them to novices. If that hypothesis is right, knowledge that previously lived in a few heads became something closer to shared infrastructure. An organisation that only uses AI person by person, in private, never builds that infrastructure.

Reorganisation. For the two historical cases economists have studied most closely, electrification and computing, the big gains arrived only when organisations rebuilt around the technology. Paul David's account of electrification argued that factories got little productivity benefit from swapping steam engines for electric motors, and that the gains came later, once the factory itself was redesigned around distributed power (David, 1990, "The Dynamo and the Computer", American Economic Review 80(2), 355-361). Brynjolfsson, Rock and Syverson formalise the same pattern as a productivity J-curve, in which complementary intangible investments, such as new processes, skills and organisational designs, come first and the measured gains come after (Brynjolfsson, Rock and Syverson, 2021, American Economic Journal: Macroeconomics 13(1), 333-372). I am assuming that generative AI follows the same pattern, which is a bet rather than a finding.

Reorganisation is precisely the thing a coping organisation cannot do. It is double-loop learning at the scale of the whole firm, and systemic coping syndrome is, by definition, the state in which the first loop has displaced the second.

So the cost of systemic coping syndrome is not mainly the forgone gain from a tool. It is the forgone capacity to compound. A coping organisation and a learning organisation can start the year with the same software licences and, if the reinvestment mechanism holds, end it on different trajectories, with the gap widening for as long as one of them keeps reinvesting and the other keeps coping.

The perverse case: AI that makes coping cheaper

The more uncomfortable consequence is that AI does not simply bounce off a coping organisation. It gets absorbed into the coping.

If the dominant habit is to work around broken steps, then the first thing people will do with a capable assistant is work around broken steps faster. The document that should not need reformatting gets reformatted by a model. The report that nobody reads gets generated in seconds instead of hours. The approval chain that adds nothing gets an AI-drafted memo for every link. Each of those is a real local saving, and each of them lowers the cost of living with the dysfunction. Since the main pressure to fix a dysfunction is its cost, AI used this way removes the pressure and entrenches the problem.

The condition under which this happens can be stated precisely. It happens when AI lowers the cost of coping more than it lowers the cost of correcting. Coping is mostly labour, which AI is good at reducing. Correcting is mostly authority, coordination and approval: someone has to be allowed to delete the step, and everyone downstream has to agree. AI does much less for that. Where correcting a process is mainly a labour problem, AI should help fix it. Where it is mainly a permission problem, AI will mostly make the workaround cheaper.

That is the claim in this essay that I think would most surprise a well-read sceptic, so I want to state it in a form that could be wrong: in an organisation with systemic coping syndrome, AI adoption will mostly automate existing workarounds rather than remove the steps that they work around, and the number of steps in core processes will not fall, and may rise, as adoption grows. If high-coping organisations turn out to delete steps at the same rate as everyone else once they have AI, I am wrong about this part.

When it is not the syndrome

There are boundaries that keep the concept honest, and they matter because a label like this can easily be turned into a way of blaming employees.

Realistic judgement is not helplessness. Learned helplessness, in the strict sense, is persistence of passivity after control has become available again. If challenging the process is still punished, if the last person who proposed a change was sidelined, then not proposing changes is an accurate reading of the environment. That is not the syndrome. It is the organisation working as designed, and the fix is to change what is rewarded and punished, not to run a workshop on mindset.

Coping is sometimes correct. In a crisis, absorbing the dysfunction and getting the job done is exactly right. More than that, the resilience-engineering tradition argues that people's everyday adaptations are a large part of why complex systems work at all, and that treating every deviation from the written process as a defect gets things backwards (Hollnagel, 2014, Safety-I and Safety-II, Ashgate). I accept that. Some workarounds are the system's real process and should be written down, not removed. Coping becomes the syndrome only when it is the permanent operating mode, and when the adaptations hide failures that someone upstream could have fixed and never hears about.

Incentives can produce the same symptoms. In professions that bill by the hour, removing work can remove revenue, so a firm that does not adopt AI may be responding rationally to its pricing rather than suffering from anything. I take the billable hour seriously as a rival explanation, and the two can be told apart. The incentive explanation predicts that the same firm will remove workarounds eagerly where efficiency is pure gain to the firm: fixed-fee matters, and internal, non-billable work such as onboarding, conflict checks or internal reporting. Systemic coping syndrome predicts that the workarounds persist there too. If they do not, the billable hour is doing the explaining, not my concept.

A simple diagnostic follows from those boundaries: ask whether the last few well-founded proposals for change were acted on, ignored or punished. Acted on means a learning organisation. Punished means a rational fear problem. Ignored, followed by the same workaround being taught to the next new joiner, is systemic coping syndrome.

What is genuinely new here, and what is not

Most of the parts are not new, and the closest prior work deserves to be named first. Nelson Repenning and John Sterman documented a "capability trap" in which people too busy firefighting cannot invest in improvement, managers attribute the resulting poor performance to the workers, tighten control, and the organisation settles into "conflict, mistrust, and control structures that prevent useful change of any type" (Repenning and Sterman, 2002, Administrative Science Quarterly 47(2), 265-295). Peter Senge's "shifting the burden" archetype describes a symptomatic fix that relieves the pressure to apply the fundamental one, until the capacity for the fundamental fix withers (Senge, 1990, The Fifth Discipline, Doubleday). The warning against using technology to speed up steps that should not exist is older still: Michael Hammer's "Reengineering Work: Don't Automate, Obliterate" made it the core of business process re-engineering (Hammer, 1990, Harvard Business Review 68(4), 104-112). Martinko and Gardner brought learned helplessness into management in 1982. Tucker and Edmondson documented the workaround loop. Argyris described single-loop and double-loop learning. David and the J-curve literature explained why general-purpose technologies need reorganisation.

Against that background, the loop at the heart of systemic coping syndrome is essentially the capability trap seen from the employee's side, and I do not claim it as a discovery. What I think the term adds is narrower:

1. The link to AI's distributional evidence. On skill and prior performance, the benefit gradient and the adoption gradient run in different directions, and the barriers that workers name are organisational.
2. The symptomatic fix has become self-service. Re-engineering and robotic process automation were projects: someone had to fund them and sign them off, which at least put a decision-maker in front of the question of whether the step should exist. Spreadsheets and macros were self-service, but only for people who could build them, and mostly for structured data. Generative AI, where it is allowed at all, lets any individual automate their own workaround in ordinary language, without a project, a budget or a sign-off, while deleting the step still needs permission. The Danish survey's authors think ChatGPT's unusually fast adoption likely reflects this kind of take-up, "driven by the take-up of individual workers independent of firm-wide initiatives". When coping faster is self-service and correcting is not, the perverse case above becomes more likely than it was for earlier technologies.
3. The cost model. The cost is the forgone capacity to compound, which is larger when the technology is non-rival and has network effects, as AI artefacts do.

If you prefer precision to memorability, "organisationally induced helplessness" ties more cleanly to the existing literature. I keep the coined term because it names the part the older terms leave out: not just lost agency, but the replacement of organisational learning with organised endurance.

What exists and what does not

• What exists: the published studies cited above, each of which measured what it says it measured in its own setting. The findings about who benefits and who adopts are empirical.
• What does not exist: any measurement of systemic coping syndrome itself. There is no validated scale and I have not run a study. The connection between the benefit evidence and the adoption evidence is my argument, not a finding of any of those papers, and none of them uses my term.
• What is analogy: the use of the 2016 neuroscience. Animal passivity under shock is a model that makes the organisational question sharper, not evidence about offices.

The strongest objections

"The people who gain most are the people least able to check the output." I think this is the strongest objection, and it is partly right. The same BCG study found that on a task outside the model's capability frontier, consultants using AI were 19 percentage points less likely to get the right answer. A junior who cannot tell good output from plausible output is a real risk, and an organisation that restricts juniors' use of AI may be practising sound risk control rather than coping. Where I come down: the objection justifies supervised use, with review of junior AI work by someone who can check it, and it does not justify exclusion, because exclusion forfeits the gains inside the frontier along with the risks outside it. The two can be told apart in practice. Risk control is targeted at specific high-risk tasks, comes with a review path, and is revisited as tools and people improve. Coping-driven restriction is blanket, has no review path and is never revisited. If most restrictions in practice turn out to be the first kind, this part of my argument is wrong.

"This is just learned helplessness, or the capability trap, with a new name." It overlaps with both, and I have said where. The difference from learned helplessness is the feedback loop: helplessness is a state that persists, while systemic coping syndrome is a process in which the coping work keeps the cause alive. The difference from the capability trap is the AI-specific part above. That difference matters practically, because it says the intervention is to remove workarounds and shorten the loop between proposal and change, not to rebuild anybody's confidence.

"The evidence on who benefits comes from narrow tasks and may not generalise." That is fair. Customer support chats, consulting tasks and short writing assignments are not the whole of work, and there is a well-known counter-case: in a field experiment with Kenyan entrepreneurs, high performers gained substantially more from an AI assistant than low performers did (Otis, Clarke, Delecourt, Holtz and Koning, 2024, "The Uneven Impact of Generative AI on Entrepreneurial Performance"). My reading, not the authors', is that where the work has no one to check it and no structure around it, the gradient can reverse. My claim does not need the benefit gradient to hold everywhere. It needs it to hold often enough that the people being left out include many of the people with the most to gain, and the current evidence supports that much.

"You sell AI, so of course you think organisations are failing to adopt it." Correct, which is why I have tried to make the central claims falsifiable rather than persuasive, and why none of the remedies below involves a purchase. The predictions below would each embarrass me if they failed.

Predictions

Each of these needs a measure and a horizon to be testable, so here they are.

1. Workaround load predicts non-adoption. Measure workaround load as the share of staff time spent on recurring tasks that the staff themselves say should not need to exist, and measure it before AI tools are made available, so that the tools cannot have changed it. Within a single organisation, holding job exposure constant, teams with the highest prior workaround load will show the lowest AI use twelve months after the tools arrive, and what use they do show will be concentrated on tasks they had already listed as workarounds.
2. High-coping organisations automate steps rather than delete them. Classify organisations as high or low coping on the same measure before AI is introduced, and count the steps in a sample of core processes at that point and again 12 to 24 months later. In the high-coping group, the step count will not fall, and may rise, while AI use rises. In the low-coping group it will fall.
3. Control beats training. Within an organisation, randomly assign teams either to receive one fast, visible yes to a small change that they proposed, or to spend the same amount of time in AI training. At three months, AI use will be higher in the first group. Prediction 3 bets directly against the training explanation, which is the most natural rival.

What to do about it

If systemic coping syndrome is organisational, the remedies have to be organisational too.

1. Count the coping. Ask every team a single question: what do you do every week that should not need to exist? Cass Sunstein has argued for something similar for firms and public institutions alike, "sludge audits" that measure the burden that processes place on the people who go through them (Sunstein, 2022, "Sludge Audits", Behavioural Public Policy 6(4), 654-673). The answers are both the diagnosis and the AI roadmap, and they will be densest among the people the organisation usually asks last. Do not turn the count into a target. The moment that teams are rewarded for reporting fewer workarounds, they will stop reporting them, and you will have taught the organisation one more way to cope.
2. Sort the workarounds before removing any. Some are the real process and should be written down. The rest are compensating for something upstream, and those are the ones to remove.
3. Shorten the loop between a proposal and a change. The thing being learned is control. The most powerful intervention is a fast, visible yes to a small proposal from someone who has stopped expecting one.
4. Protect the slack. Time freed by AI should be reinvested in removing the next workaround, deliberately, rather than absorbed by more coping. Otherwise the compounding never starts.
5. Replace prohibition with governance. Name one person who owns the AI rules, require review of junior AI work by someone able to check it, and give every restriction a date on which it will be revisited. A restriction with no owner and no review date is part of the syndrome.
6. Build shared artefacts, not private habits. Prompts, templates, agents and checked workflows that the whole organisation can use are where non-rivalry and network effects pay off. A firm where everyone uses AI privately has the tools and none of the compounding.
7. Remove steps, do not just accelerate them. Before automating any task, ask whether it should exist. If AI makes a pointless step painless, the step will live forever.

Organised endurance

The organisations that need AI most are not, for the most part, the ones that have rejected it. They are the ones that are too busy coping to see what it is for, and whose people have so rarely seen a proposal turn into a change that they no longer make proposals. What they lose is not one tool or one productivity gain. It is the capacity to compound that a learning organisation would have had.

The good news in Maier and Seligman's reversal is that agency can be learned. The bad news is that an organisation has to give its people something to learn it from. That is a leadership decision, and it is available today.

About the author: Eduardo Aguilar Pelaez is CTO and co-founder at Legal Engine Ltd. He writes on formal methods, AI agents, and the discipline of building systems that survive being walked away from.