Trauma-Informed Design

Building systems that hold the people inside them, instead of extracting from them, in an age where intelligent systems increasingly decide what those people see, do, and conclude.

Future Proof Intelligence. Research. No. VI. MMXXVI

Abstract

Trauma-informed practice is usually filed under care. It belongs under design. Over four decades, clinical and built-environment practitioners have produced a rigorous body of knowledge about what an environment is permitted to do to a human nervous system: it must be safe, predictable, trustworthy, and transparent; it must offer genuine choice and agency; it must be able to repair the harm it causes; and above all it must not re-injure the people who depend on it. These are not soft preferences. They are structural constraints, grounded in the neuroscience of stress and the documented physiology of unpredictability and loss of control.

This paper takes those constraints seriously and moves them out of the clinic. As intelligent systems increasingly mediate attention, work, and decision making, they have become the dominant environment most people now inhabit. Almost all of that environment is designed on the opposite principle. Extraction is the default: attention is harvested, uncertainty is manufactured, agency is narrowed, and the human is left to absorb the consequences of failures the system arranged. We argue that trauma-informed design is the precise counter-discipline to extractive design, that its principles translate cleanly into the grammar of software and institutions, that the 2026 regulatory frame is beginning to make extraction a liability rather than a strategy, and that a system built to hold rather than extract is not a kinder version of the same machine. It is a different machine. Future Proof was built underneath that distinction.

1. The discipline nobody files under design

1.1 A change of question

The most consequential idea in trauma-informed practice is not a technique. It is a change of question.

The older clinical question is "what is wrong with this person." It treats the person in front of you as a problem to be diagnosed and the environment as neutral. The trauma-informed question is "what happened to this person, and what is the environment doing to them now." It treats the person as a nervous system that has learned, often correctly, that certain conditions are dangerous, and it treats the environment as an active participant in either the injury or the recovery. The environment is never neutral. It is part of the diagnosis.

That sentence is the whole paper compressed. If an environment is part of the diagnosis, then the environment is something that can be designed well or designed badly, and the difference is measurable in the body of the person inside it. This is a design claim. It has simply been written down, until now, in the language of care.

The United States Substance Abuse and Mental Health Services Administration set out the canonical articulation in 2014, in its document on the concept of trauma and guidance for a trauma-informed approach. It defines a trauma-informed approach through four assumptions, often called the Four Rs. An organisation or a system realises the widespread impact of trauma and the possible paths to recovery. It recognises the signs of trauma in the people it touches. It responds by integrating that knowledge into its policies, procedures, and practices. And it actively resists re-traumatisation: it does not, through its own ordinary operation, re-injure the people who come to it already carrying injury.

That fourth R is the load-bearing one for everything that follows. Resisting re-traumatisation is not an aspiration. It is a constraint on the system's permitted behaviour. It says, in effect: there are things this environment is not allowed to do to the people inside it, no matter how efficient those things are, because doing them re-opens a wound the environment did not create but is now responsible for not re-opening.

1.2 The six principles, read as design constraints

The same framework sets out six principles. They are usually presented as values. Read them instead as a specification, because that is what they are.

Safety. People feel physically and psychologically safe. The setting itself is safe, and interactions inside it produce a felt sense of safety. Crucially, safety is defined by the people served, not asserted by the institution. A system cannot declare itself safe. Safety is an outcome measured in the other person's nervous system.

Trustworthiness and transparency. Decisions are made and conducted with transparency, with the goal of building and maintaining trust. The process by which difficult decisions get made is itself visible. Trust is treated as something the system must continuously earn and can easily forfeit, not a default it is owed.

Peer support. The lived experience of others who have been through the same thing is treated as a primary vehicle for safety, hope, and recovery, not a supplement to expert delivery.

Collaboration and mutuality. Power differences between those who run the system and those who depend on it are levelled deliberately. Healing, and good decisions, happen in relationship and in shared work, not in a hierarchy delivering to a passive recipient.

Empowerment, voice, and choice. The system is built to recognise and build on the strengths of the people in it. They have real decision-making power over what happens to them. Choice is not a single forced option presented as a choice. It is genuine optionality with genuine consequences the person controls.

Cultural, historical, and gender awareness. The system moves past stereotypes and bias, responds to the specific histories people carry, and recognises that some of the trauma in the room is historical and structural, not individual.

Strip the care vocabulary and look at what these six things describe. They describe an environment that is predictable, that is honest about its own mechanics, that does not concentrate power asymmetrically against the people who cannot leave, that gives those people real control over what is done to them, and that is responsible for not re-injuring them. That is a design specification. It happens to be the design specification almost no large intelligent system currently meets.

1.3 Why this is design and not sentiment

There is a reflex, when a paper begins with the word trauma, to expect sentiment and to discount it accordingly. There is also a more careful objection worth meeting head on, because a paper that argues for rigour cannot itself be loose. The popularisation of trauma over the past decade, of which Bessel van der Kolk's The Body Keeps the Score is the most prominent example, has been criticised, including in a 2025 review in BJPsych Bulletin that evaluated a large set of the book's specific claims, for overreaching on certain mechanistic assertions about trauma-induced brain damage and the unique efficacy of particular body-based treatments. This paper does not rest on the contested claims. It does not need them. Its argument runs on the conservative, well-replicated core that survives the critique: that the nervous system appraises threat in an environment rapidly and largely below conscious control, that sustained or repeated activation of the stress response carries a cumulative physiological cost, and that predictability and control reliably modulate that cost. Those propositions are not the disputed frontier of trauma science. They are its settled foundation, drawn from stress physiology and human factors rather than from any single popular book, and they are sufficient for everything claimed here. Where the popular discourse overreaches, the discipline does not, and this paper stays on the side of the discipline.

With that boundary drawn, the discount on the word trauma is still a mistake, and the reason it is a mistake is physiological.

Trauma-informed practice is not the claim that systems should be nicer. It is the claim that the nervous system processes the environment before the conscious mind does, that an environment which signals threat imposes a measurable and cumulative biological cost on the people in it, and that this cost degrades exactly the capacities the environment was supposedly built to support: attention, judgement, memory, the ability to act well under pressure, the willingness to tell the truth about what is going wrong. A frightened nervous system is not a moral problem. It is a performance problem, an accuracy problem, and a safety problem, and it is one the designer caused.

This is why the discipline has to be read as design. It is not asking the designer to care more. It is telling the designer that the environment is an input to human cognition and behaviour, that the input is being set whether or not anyone is paying attention to it, and that setting it badly is a defect in the system, not a defect in the user. The next section is the evidence for that claim, and the bridge from the clinic to everything else.

2. The body reaches the room before the mind does

2.1 The built environment as the first test case

The discipline of trauma-informed design did not stay in the clinic. Over the past two decades it moved into architecture, where the claim could be tested against physical space: permanent supportive housing, homelessness services, domestic violence shelters, therapeutic residential care for young people. These are environments full of people whose nervous systems have already learned that certain conditions are dangerous. They are, in effect, an instrument sensitive enough to read what a building does to a human before the human has a conscious thought about it.

The premise the field works from is direct: the body reacts to space before cognition processes it. A person walking into a room registers the lighting, the sightlines, the exits, the acoustics, the degree of control they have over their own immediate surroundings, and the predictability of the layout, and the body has already responded, has already either braced or settled, before any of it has been consciously named. If that is true, then the building is not a backdrop to the human experience inside it. The building is part of the experience, and it can be engineered toward bracing or toward settling.

The practical principles that have emerged are specific enough to falsify. Diffuse, warm light rather than harsh overhead glare, kept toward the warmer end of the spectrum. Connection to the natural world, because exposure to nature reliably lowers measured physiological stress. Prospect and refuge, an old idea in environmental psychology: people settle when they can observe their surroundings without themselves being exposed, and they brace when the geometry is reversed. Predictability of structure and finish, because a space whose logic the body can learn becomes a space the body can stop monitoring. Dedicated places to self-regulate. And, running through all of it, control: the degree to which a person can act on their own immediate environment rather than only be acted on by it.

None of this is decoration. Each item is a lever on the human stress response, set by a designer, whether the designer knew they were setting it or not.

2.2 Allostatic load: the cost has a name

The reason these levers matter is that the cost they govern is real, cumulative, and named. The name is allostatic load.

Allostasis is the body's process of maintaining stability through change: mobilising the stress response to meet a demand, then standing it down. Allostatic load, a concept developed in the stress-physiology literature most closely associated with Bruce McEwen, is the wear and tear that accumulates when that machinery is activated too often, held on too long, or not switched off when the demand passes. The stress response is not the problem. The stress response held chronically is the problem. The hypothalamic-pituitary-adrenal axis and the sympathetic-adrenal-medullary axis exist to handle acute threat and then return the system to baseline. An environment that keeps signalling low-grade threat keeps that machinery partly engaged, and the cost compounds, in the cardiovascular system, in metabolic regulation, in cognition, in mood.

The phrase to hold onto is this: an environment has a metabolic cost, and the designer of the environment sets that cost. A space that is unpredictable, that cannot be controlled by the person in it, that withholds information the body needs to feel safe, is not merely unpleasant. It is, in a precise physiological sense, expensive to occupy. The bill is paid by the occupant, in a currency the occupant cannot opt out of, and it accrues whether or not anyone consents to it or even notices it.

2.3 Predictability and control are not comforts. They are variables.

Two variables do most of the work in determining what an environment costs the body: predictability and control.

The stress-physiology literature is consistent on the direction of the effect. Uncertainty, the inability to predict what is coming, drives the stress response. So does the absence of control, the inability to influence what happens to you. Stress can be modelled as a response to predictive uncertainty: a state in which the organism cannot tell how to get back to a condition compatible with its needs. Robert Sapolsky's long body of work on social hierarchies and stress points the same way: it is not exertion that most reliably loads the system, it is subordination without control or predictability. The organisms that fare worst are not the ones working hardest. They are the ones who cannot predict what will happen to them and cannot do anything about it.

Read that sentence again with a screen in mind instead of a savannah. An interface that changes its behaviour without warning, that does not explain why it did what it did, that cannot be steered by the person using it, that surfaces consequences the person could not have anticipated, is not a usability problem in the conventional sense. It is an environment that has been engineered, deliberately or by neglect, to maximise exactly the two variables that load a nervous system. The fact that the environment is made of pixels rather than concrete does not change the physiology. The body reaches the room before the mind does, and the room is now mostly made of software.

This is the bridge the rest of the paper crosses. Trauma-informed design is not a clinical speciality that happens to use the word design. It is the general study of what an environment is permitted to do to the people who cannot leave it, grounded in the measurable biology of how environments act on bodies. The environment most people now spend most of their waking attention inside is no longer a building. It is a system: a feed, a workflow, an institution, a model deciding what they see and what they are offered and what they are told. The discipline applies there with full force. The next section translates it.

3. From rooms to systems

3.1 The translation is exact, not metaphorical

It would be easy to treat the move from buildings to software as an analogy, a suggestive comparison to be enjoyed and then set aside. It is not an analogy. The physiology does not know the difference between a corridor with no sightline and a system that changes state without telling you. Both are environments. Both are processed by the body before the mind. Both load or unload the same machinery. The translation from the six principles to the grammar of systems is therefore not poetic licence. It is a one-to-one mapping, and each mapped principle yields a test the system either passes or fails.

What follows is that mapping. The aim is not a checklist. It is to show that every principle developed in care has a precise systems counterpart, that the counterpart is engineerable, and that most systems fail it not because the principle is unknown but because failing it is more profitable in the short term.

3.2 Safety, as a system property

In the clinic, safety is defined by the person served, not asserted by the institution. The systems translation is severe and worth stating plainly: a system does not get to declare itself safe. Safety is the absence of a credible threat in the experience of the person inside the system, measured there and nowhere else.

The operational test is whether the system can harm the person through its ordinary, intended operation, not only through error or abuse. A system that, when working exactly as designed, escalates the person's stress, exposes them to consequences they could not foresee, or extracts from them in ways they would refuse if they could see them clearly, is not a safe system with occasional failures. It is an unsafe system performing correctly. This distinction is the entire difference between a bug and a design posture, and trauma-informed design refuses to let the second hide behind the first.

Figure 1. Safety as designed versus safety as asserted. In the asserted model, the system states a safety claim, ships, and treats harm reports as edge cases against an assumed-safe baseline. In the designed model, the baseline assumption is reversed: the system is treated as capable of harm through normal operation until its normal operation has been examined from inside the affected person's experience, and the burden of proof sits with the system, not the person reporting the harm.

3.3 Trustworthiness and transparency, as a system property

In care, trust is continuously earned and easily forfeited, and the process by which difficult decisions are made is itself visible. The systems counterpart is the requirement that the system be legible about its own mechanics to the person it acts on.

This is sharper than the familiar call for explainable interfaces. Legibility here means three concrete things. First, the person can find out why the system did what it did to them, in terms they can act on, not a generic disclosure written for a regulator. Second, the incentives of the system are not concealed from the person whose behaviour those incentives are shaping; a system that profits from the person's continued engagement and does not make that legible is, by this standard, untrustworthy by construction, regardless of how well it performs. Third, when the system changes the rules, the change is announced before it takes effect, not discovered afterwards by the person who has already been governed by the new rule. A system that quietly changes what it does to people, and lets them find out by being harmed, has forfeited trust in exactly the way the clinical literature describes, and the fact that the change was logged in a release note does not repair it.

3.4 Choice, voice, and agency, as a system property

The clinical principle is empowerment, voice, and choice: genuine optionality with genuine consequences the person controls. The systems failure mode is precise and ubiquitous. It is the false choice: a decision presented as the person's, engineered so that one option is the only viable one, while the appearance of choice transfers responsibility for the outcome onto the person who never really had it.

A trauma-informed system is held to the standard that the choices it presents are real. A choice is real when the non-preferred option is genuinely available, genuinely survivable, and genuinely understood by the person before they choose. By this standard, an interface that makes the privacy-protecting option slower, uglier, and harder to find than the data-surrendering option is not offering a choice. It is manufacturing consent and calling it agency. The clinical literature has a blunt word for an environment that removes a person's real control over what is done to them while maintaining the form of consent: it calls that environment coercive, and it treats coercion as re-traumatising. The word does not lose its force when the environment is an onboarding flow.

3.5 Predictability, as a system property

Predictability moved, in the built-environment work, from a comfort to a measured variable, because the body pays a metabolic price for an environment it cannot anticipate. The systems counterpart is the requirement that the system be steady: that its behaviour be inferable from its past behaviour, that state changes be signalled, that the person never has to brace because they cannot tell what the system will do next.

This is the principle most directly violated by systems that optimise engagement, because unpredictability is engaging. Variable, intermittent, unanticipated reward is the most powerful schedule of reinforcement known, which is precisely why it is the most extractive. A feed that is predictable is a feed that can be put down. A feed engineered so the next item is never predictable is a feed engineered to keep a nervous system partly mobilised, scanning, unable to stand down. That is not a side effect of the design. In the systems built to maximise time on platform, it is the design. Trauma-informed design treats engineered unpredictability not as a growth tactic but as the deliberate manufacture of allostatic load in another person's body, and asks the designer to own that description of what they have built.

3.6 Collaboration over hierarchy, as a system property

In care, the deliberate levelling of the power gradient between those who run the system and those who depend on it is a principle, not a courtesy, because an uncontrollable asymmetry is itself a stressor. The systems counterpart concerns dependence. The more a person depends on a system they cannot influence, cannot exit at acceptable cost, and cannot hold to account, the more that system is, structurally, in the position the clinical literature warns about: the position of an institution acting on someone who has no power over what is done to them.

The design response is not to pretend the asymmetry away. It is to build deliberate counterweights into the system: real exit, real recourse, real influence over the terms, the ability to be heard by something that can actually change the system's behaviour. A system that accumulates dependence while removing every counterweight is not efficient. It is, by the discipline's own definition, building the precondition for institutional harm and then standing on the side of the power.

3.7 The AI-specific failure: the system that does not remember you

There is one systems failure mode that has no clean analogue in the building and that the arrival of conversational and agentic AI makes urgent. It is the per-interaction reset: the system that meets the person as a stranger every time, retains nothing of what was established, and requires the person to rebuild context, re-establish what they need, and re-explain who they are at the start of every encounter.

In a clinical setting this failure has a name and it is taken seriously. Making a person who carries injury re-narrate the injury repeatedly, to a system that does not hold what they already disclosed, is itself recognised as a re-traumatising act, which is why continuity of the relationship is treated as a structural feature of trauma-informed care rather than an administrative convenience. The systems translation is direct. A system that resets its understanding of the person on every session, that makes the person prove again what was already proven, that cannot carry a commitment it made forward into the next interaction, is not merely inconvenient. It is imposing the specific load the discipline identifies: the demand to re-establish safety from zero, repeatedly, with a counterpart that has no memory of having been trusted.

This is precisely the seam where the trauma-informed argument and the identity-layer argument meet, and it is worth marking here rather than only at the end. An intelligent system without continuity of the person across interactions cannot satisfy the discipline, because it structurally cannot accumulate trust, cannot honour a prior commitment, and cannot avoid making the person start over. Continuity is not a feature that improves the experience. It is a precondition for the system being capable of holding at all. The orchestration layer, the part that routes and executes the work, is indifferent to this. Only a layer built to carry who the person is, and what was established with them, across the resets can meet the requirement. Section 6 returns to this as the structural conclusion. It is named here so the reader sees it is not a late pivot but a property the analysis required from the moment the building became a system.

3.8 The summary mapping

Figure 2. The six principles as system tests. Safety: can the system harm the person through correct operation, judged from inside that person's experience. Trustworthiness: is the system legible about its mechanics, its incentives, and its changes, before they bind. Choice and agency: are the presented options real, with the non-preferred option survivable and understood. Predictability: is the system's behaviour inferable, with engineered unpredictability treated as manufactured load. Collaboration over hierarchy: are there real counterweights to the dependence the system accumulates. Repair, taken up in Section 6: when the system harms someone, does it have a designed, owned, non-discretionary way to make it right. A system passes the discipline only when it passes all of them. Almost none do, and the next section is why.

4. The extractive default

4.1 Extraction is the path of least resistance

If the trauma-informed specification is this clear, the obvious question is why so little of the built digital world meets it. The answer is not that designers are cruel. The answer is that extraction is the path of least resistance, and that for roughly two decades the dominant business model of consumer software actively rewarded building the exact inverse of every principle in Section 3.

The critique is not new and it is not fringe. It was made from inside the industry. Practitioners who had worked at the centre of consumer technology, among them figures associated with the Stanford persuasive-technology tradition and later the Center for Humane Technology, described the attention economy as extractive by construction: a competition not to serve the user's intent but to capture and hold the user's attention, a race, in the phrase that became its epitaph, to the bottom of the brain stem. The mechanisms they catalogued read like Section 3 inverted point by point. Intermittent, unpredictable reward, deployed precisely because unpredictability holds a nervous system. Notifications engineered to interrupt, because the interrupted mind returns. Infinite scroll, which removes the natural stopping cue a predictable environment would provide. Choice architecture arranged so the extractive option is the easy one and the protective option is the friction. None of this required malice. It required only an objective function that rewarded time and engagement, and the gradient did the rest.

4.2 Dark patterns are the discipline run backwards

The clearest evidence that extraction is a coherent design posture and not an accident is that it has a literature, a taxonomy, and now a body of law. Deceptive design patterns, commonly called dark patterns, are interface choices that get a person to do something they did not intend: to buy, to subscribe, to keep a subscription they wanted to end, to surrender data they would have withheld, to consent to something the design made it irrational to read.

Set the dark-pattern taxonomy beside the trauma-informed specification and the relationship is exact. The trauma-informed principle of real choice is inverted into the forced or obstructed choice. The principle of transparency is inverted into the hidden cost and the buried disclosure. The principle of predictability is inverted into the bait and switch. The principle of agency is inverted into confirmshaming and the manufactured urgency that overrides deliberation. A dark pattern is not the absence of design. It is trauma-informed design executed in reverse, with the same knowledge of the human nervous system, turned to the opposite end. This matters because it disposes of the defence that holding systems is merely aspirational. The same precision that builds extraction can build its inverse. The knowledge is identical. Only the objective function differs.

4.3 The worked example: consent, fatigue, and the manufacture of compliance

The single most instructive worked example of the discipline run backwards is the one almost every reader has experienced thousands of times: the consent interface.

Consent is, on paper, the purest expression of the trauma-informed principle of choice. It is the formal moment at which a person is asked, explicitly, what may be done to them. Data-protection law in the European Union requires that consent be freely given, specific, informed, and unambiguous, and the European Data Protection Board has been explicit that deceptive design can vitiate it. The principle and the law are aligned with the discipline. The implementation, almost everywhere, is the discipline inverted.

The mechanisms are catalogued and measured. Asymmetric design, in which acceptance is a single prominent button and refusal is buried, low-contrast, or routed through additional screens. Click fatigue, in which accepting everything takes one action and protecting oneself takes many. Pre-selected options that enable what the person did not choose. The aggregate effect, documented in studies of large samples of websites, is that a substantial majority of consent interfaces deploy at least one deceptive pattern, and the human consequence has its own name in the literature: consent fatigue, the state in which a person, asked to make the same defensive decision hundreds of times under deliberately uneven friction, stops reading, stops resisting, and clicks the path the designer made easy, then a related state, banner blindness, in which the person ceases to perceive the prompt as a decision at all.

Read that sequence through Section 2. The person is placed in an environment that repeatedly demands a decision, makes one option costly and the other free, and offers no way to make the demand stop on protective terms. Repeated, uncontrollable, low-grade demand with an asymmetric exit is a textbook recipe for the precise physiological state the discipline exists to prevent, and the documented behavioural result, surrender through exhaustion, is exactly what the recipe predicts. The system did not obtain consent. It manufactured compliance by loading the person until resistance became more expensive than capitulation, and then recorded the capitulation as a free choice. This is the discipline in reverse with unusual clarity, because here the law itself names what the design defeated. It is also why the regulatory turn described next is not an overreach. It is the law catching up to a harm that was engineered in plain sight.

4.4 The Ironies of Automation: the system arranges the human's failure

There is a second extractive pattern, older than the consumer attention economy and now central to AI-mediated work, and it is the systems analogue of re-traumatisation in its purest form.

In 1983 Lisanne Bainbridge published "Ironies of Automation," which became one of the most cited papers in human factors and remains, by the field's own assessment, largely unresolved four decades later. Its argument is short and devastating. When a system automates the easy, routine parts of a task, it does not simply relieve the human. It removes the human's practice, pushes the human out of the loop of understanding, and reserves for the human exactly the rare, hard, high-stakes moments when the automation fails, demanding that the human now take over a situation they have not been tracking, with skills the automation has been quietly eroding, under time pressure the automation created, and then holds the human responsible for the outcome.

Read that against the fourth R, resist re-traumatisation. The system creates the conditions for the human's failure, exposes the human to the consequences of that failure, and then locates the fault in the human. This is, structurally, what an institution does when it re-injures the person who came to it for help and then treats the injury as the person's pathology. Bainbridge described it in control rooms and cockpits. It is now the default architecture of human-AI collaboration: the model handles the routine, the human is kept nominally in charge for accountability and edge cases, the human's competence and context decay precisely because the routine has been taken away, and when the rare failure arrives the human is positioned to absorb it. A trauma-informed reading of automation does not ask whether the human is in the loop on an org chart. It asks whether the system has been arranged so that the human is set up to fail and then made to own the failure. In most current deployments it has been.

4.5 Extraction is becoming a liability, not only a posture

For most of the period in which extractive design was built, it carried no external cost. That has begun, materially, to change, and the change is part of why this paper is timed as it is.

In the European Union, the Digital Services Act prohibits, in its Article 25, the design of online interfaces that deceive or manipulate users or otherwise materially distort or impair their ability to make free and informed decisions. The Digital Services Act is in force, its enforcement powers are live, and the penalties available reach a percentage of global turnover large enough to be strategic rather than incidental. As of 2025 and into 2026, formal proceedings touching deceptive design have been opened against several of the largest platforms. Beyond the existing instrument, the European Commission has tasked the preparation of a Digital Fairness Act aimed specifically at dark patterns, addictive design, and the manipulative profiling of vulnerable consumers, signalling that the regulatory direction is toward treating extractive design as a category of harm in its own right rather than a marketing technique. In the United States, the Federal Trade Commission stated as early as 2021 that dark patterns can constitute unfair or deceptive practices under existing law.

Running alongside this, the European Union's Artificial Intelligence Act becomes fully applicable on 2 August 2026. Obligations for providers of general-purpose AI models took effect on 2 August 2025; the substantive obligations for high-risk systems and the Commission's enforcement powers apply from 2 August 2026, with an extended transition to 2 August 2028 for certain high-risk systems embedded in already-regulated products under the simplification agreed through the AI omnibus. The Act includes prohibitions on manipulative and exploitative AI practices and transparency obligations on AI-mediated content and interaction. The detail that matters for this paper is not any single clause. It is the direction. The roots of the system that will govern AI-mediated environments are hardening now, in 2026, and they are hardening in the direction of treating manipulation, opacity, and the engineered overriding of human decision-making as liabilities to be evidenced against. Extraction is moving from a default to a documented exposure. The discipline that was filed under care is, quietly, becoming the discipline that will be filed under compliance, and the operators who built on it early will be the ones who do not have to retrofit it under enforcement.

5. Holding as an engineering specification

5.1 The positive construction

Critique is the easy half. The discipline is only worth the name if it can say, concretely, what a system that holds rather than extracts actually does, at the level of architecture rather than sentiment. This section is that construction. It treats the trauma-informed principles not as ethical garnish applied to a finished product but as first-order system requirements, on the same footing as latency, correctness, and cost, and earlier in the build than any of them.

The organising idea is a single inversion. In an extractive system, the human absorbs the system's costs: the system externalises its uncertainty, its failures, its incentive conflicts, and its cognitive load onto the person, because the person is the cheapest place to put them and cannot refuse delivery. In a system that holds, the system absorbs them. The system carries the uncertainty so the person does not have to brace. The system carries the cost of its own failures rather than transferring them to the person and then blaming the person. The system makes its incentives legible rather than making the person pay the price of not knowing them. Holding is not softness. It is the deliberate, engineered decision about where in the system the unavoidable costs are going to land, and the decision that they will not land, by default, on the nervous system of the person who cannot refuse them.

5.2 Predictability as a first-class requirement

A system that holds is, before anything else, steady. Its behaviour is inferable from its history. When it must change, it signals the change before the change binds, in a form the affected person actually receives, not a form that technically discloses while practically hiding. State is visible: the person can always tell what mode the system is in and what it is about to do. Surprises are treated as defects, including the surprises that increase engagement, because the discipline does not grant an exception for profitable harm.

This is the requirement most in tension with growth-optimised design, and naming the tension is the point. A team that has internalised the discipline does not ask "how do we make this more engaging." It asks "what is this system doing to the load on the person, and is any of that load being manufactured rather than inherent to the task." Manufactured load is not a feature with an unfortunate side effect. It is the defect, named correctly.

5.3 Repair as a designed, non-discretionary behaviour

The single most under-built property in almost every system is repair. Extractive systems are often competent at prevention theatre and almost universally incompetent at repair, because repair costs the system something and the system was designed to externalise its costs. Trauma-informed practice treats this exactly backwards from the industry norm: it assumes harm will occur, because any sufficiently complex system acting on humans will sometimes harm them, and it judges the system not primarily on whether it avoided all harm but on what it does the moment harm has occurred.

Designed repair has properties that distinguish it from a complaints process. It is non-discretionary: it triggers on the harm, not on the person's willingness to fight for it, because requiring the harmed person to fight is itself a second injury. It is owned: a specific part of the system is accountable for the repair, not a queue. It is proportionate and legible: the person can see that the system has acknowledged the harm, named it accurately, and changed something so that the same harm is less likely to recur. And it is fast enough to matter, because a repair that arrives after the nervous system has already drawn its conclusion about the environment is not a repair, it is a record. A system with engineered repair is not a system that fails less. It is a system that, when it fails, does not convert a failure into a betrayal. That distinction is the entire subject of the next section.

5.4 Agency as a structural property, not a setting

In a holding system, the person's control over what is done to them is structural, not a preference buried three screens deep. The protective option is at least as easy as the extractive one; where they cannot be made equally easy, the system carries the friction rather than placing it on the protective path. Defaults are set to the configuration the person would choose if they had full information and unlimited time, because the system, not the person, is the party that has full information, and a default is a decision the designer made on the person's behalf whether or not the designer admits it. Exit is real: the person can leave with their dignity, their data, and their continuity intact, because a system that can only retain people by making leaving punishing has already failed the collaboration principle and is relying on the asymmetry the discipline exists to counterweight.

5.5 The economics of holding

A specification that cannot survive contact with an objective function is a wish, not a discipline. The honest objection to everything in this section is economic: holding appears to cost more than extracting, because extracting externalises its costs and holding internalises them, and a system that internalises costs its competitors externalise looks, in a quarterly frame, slower and dearer. The objection is real and must be answered in its own terms rather than waved away with virtue.

The answer has three parts and none of them requires the reader to value anything they do not already value. First, the cost asymmetry is largely a timing illusion. Extraction does not eliminate its costs. It defers them and relocates them, into churn that has to be repurchased, into trust that has to be rebought after each erosion, into the brittle dependence that holds users only until a less punishing alternative appears, and, increasingly, into regulatory exposure that converts to a direct liability on a published timetable. A system that holds pays its costs visibly and early. A system that extracts pays them invisibly and later, with interest, and reports the gap in between as margin. The discipline does not abolish the cost. It refuses to misreport it.

Second, holding compounds where extraction depreciates. Trust that is continuously earned rather than continuously spent is one of the few assets in a system that appreciates with use, because each honoured interaction lowers the cost of the next one. Extraction has the opposite curve: each extractive interaction raises the user's guard, raises acquisition cost, and shortens the half-life of the relationship. Over a long enough horizon the holding system is not the more expensive one. It is the one whose unit economics improve while the extractive system's decay, which is precisely why extraction must keep accelerating to stand still.

Third, and decisively for the period this paper is written in, the externalisation route is closing. The mechanism that made extraction cheap, the ability to put the cost onto a party who could not refuse it and onto a future that had not yet arrived, depended on no one pricing the cost back to the system. The 2026 regulatory frame is, in plain terms, the beginning of that pricing. When the cost of extraction is forced back onto the system that produced it, the apparent economic advantage of extraction does not shrink. It inverts. Holding is not the expensive option that a principled operator chooses against their interest. It is, on the horizon that is now arriving, the cheaper one, and the discipline's economic claim is simply that the people who saw this early built on the correct side of the inversion before it was obvious.

5.6 The specification, stated once, plainly

Figure 3. The holding specification. A system that holds: carries uncertainty so the person does not have to brace; is legible about its mechanics and its incentives before they bind; offers choices that are real, with the protective option no harder than the extractive one; behaves predictably and treats engineered unpredictability as a defect; counterweights the dependence it accumulates with real exit and real recourse; and, when it harms someone, has a designed, owned, non-discretionary, fast, and legible way to make it right. Every item is testable. None of it requires the system to be soft. All of it requires the system to decide, in advance and on purpose, that the costs of its operation will not be silently transferred to the people who cannot refuse them. That decision is the whole discipline. Everything else is implementation.

6. The institution as a nervous system

6.1 The largest environment is not the interface

Everything so far has treated the system as the thing the person touches: the feed, the workflow, the model, the screen. But the largest environment a person inhabits is rarely the interface. It is the institution behind it. The interface is downstream of an organisation's incentives, its tolerance for telling itself the truth, and its conduct when something it did caused harm. A trauma-informed analysis that stopped at the screen would be treating a symptom. The discipline has to reach the institution, and at the institutional level it has two of its most rigorous and least sentimental bodies of evidence.

6.2 Institutional betrayal: dependence plus failure is itself an injury

The psychologist Jennifer Freyd developed the concept of institutional betrayal: wrongdoing perpetrated by an institution upon the individuals who depend on it, including the institution's failure to prevent harm, or to respond supportively to harm, that occurs within its own context. It is an extension of betrayal-trauma theory, and its central finding is the one that matters for design. The injury is not only the original harmful event. A distinct and measurable injury is inflicted by the trusted institution's failure to respond, by the experience of being harmed and then, by the very body one depended on for protection, being disbelieved, minimised, processed, or ignored. Betrayal by a depended-upon institution is not a lesser harm than the precipitating event. In many cases it is the larger one.

Translate that into system terms and it indicts a near-universal design assumption. Most systems are built to minimise the rate of harmful events and treat the response to harm as an afterthought: a support queue, a form, a policy page. The institutional-betrayal evidence says this is precisely inverted. The response is not the afterthought. For a person who depends on the system and cannot easily leave it, the quality of the response is a primary determinant of whether the encounter becomes an injury at all. A system that handles failure with opacity, delay, denial, and the quiet transfer of fault onto the harmed person is not a system that merely failed. It is a system that converted a failure into a betrayal, and betrayal, the evidence is clear, is the part that does the lasting damage. This is the deepest justification for the repair requirement in Section 5. Repair is not customer service. It is the mechanism by which a system avoids inflicting the second, larger injury on the people who trusted it.

6.3 Institutional courage: repair is a designable institutional behaviour

Freyd named the antidote institutional courage: a depended-upon institution's commitment to seek the truth and act morally despite unpleasantness, risk, and short-term cost. The concept is operationalised, not aspirational. It includes responding supportively to those harmed and to those who report harm, conducting honest self-study, practising transparency about the institution's own conduct rather than only its successes, and making reparation where reparation is owed. Empirical work has found that institutional courage buffers against the harm of institutional betrayal and is associated with better outcomes for the people inside the institution.

The design claim that follows is the load-bearing one for this section. Institutional courage is a set of behaviours, and behaviours can be designed into a system rather than left to the character of whoever happens to be on duty. An institution that depends on individual virtue to do the right thing when it has caused harm has not solved the problem; it has gambled it on personnel. A trauma-informed institution builds the courageous behaviour into the structure: the response to harm is owned and non-discretionary, the self-study is scheduled and adversarial rather than self-congratulatory, the transparency extends to the institution's own failures and not only its wins, and the reparation is a designed pathway rather than a negotiation the harmed person has to win. Courage, engineered into the structure, stops being a hope and becomes a property.

6.4 Psychological safety: the same root, the internal face

Amy Edmondson's work on psychological safety completes the picture from the inside. Psychological safety is the shared belief that a group is safe for interpersonal risk: that a person can ask a question, admit a mistake, raise a concern, or challenge a direction without fear of humiliation or reprisal. Edmondson's research, and the large body of work that followed it, established that this is not a comfort variable. Its effect on performance grows precisely as uncertainty and complexity rise, because in an uncertain environment the things that prevent disaster, the early warning, the admitted error, the named risk, the dissenting judgement, only reach the system if the people holding them are not too braced to speak.

This is the same root as everything else in the paper, observed from inside the institution rather than from the position of the person it acts on. A braced nervous system does not produce its best work, does not surface what it knows, and does not tell the truth about what is going wrong, whether the brace is caused by an unpredictable interface, a coercive policy, or a culture that punishes the messenger. Trauma-informed design of the external system and psychological safety inside the institution are not two topics. They are one principle, applied to the people the system acts on and to the people who build it, and a system cannot reliably hold the first while bracing the second. An organisation that extracts from its own people will, under pressure, extract from everyone else, because it has trained itself that the cheapest place to put a cost is onto whoever cannot refuse it.

6.5 Why this cannot be solved at the application layer

Before stating the structural conclusion it is worth disposing of the most natural objection an engineer will raise: that everything described here is an application concern, solvable feature by feature inside each product by teams that have read the discipline. This is the same mistake the field has already made once with security and once with privacy, and it failed both times for the same reason it will fail here.

Trauma-informed properties are not local. Safety judged from inside the person's experience, predictability across the systems a person actually moves through, legibility of incentives, the avoidance of forced re-narration, repair that is owned rather than dispersed: every one of these is violated not by a single bad screen but by the seams between systems, and the seams are exactly where no application team has authority. A person does not experience one product. They experience the composition of many, none of which carries who they are across to the next, each of which resets them, each of which can honour its own consent and still hand the person to a context that does not. The harm accrues in the gaps, and a property that must hold across systems cannot be guaranteed by any system. It can only be guaranteed by a layer beneath all of them that carries the continuous person and the commitments made to them, and that can be held accountable when the seams fail. This is why the discipline, taken to its conclusion, is a substrate question and not a feature one. It is also why the operators who treated it as a substrate question first did not arrive at this by adding ethics to a finished product. They started underneath.

6.6 The counterpart to the identity layer

This is where the paper arrives, and where Future Proof sits underneath it.

The companion thesis to this one, the identity layer, makes a structural argument about the AI era: that orchestration, the routing and execution of intelligent work, is becoming commodity, and that the scarce, durable, defensible layer is identity, the continuity of who a person or an institution is, what they have committed to, and what may be done in their name across systems that will not remember unless something is built to remember. Trauma-informed design is the exact counterpart of that thesis, and the two only make sense together. The identity layer answers the question who is the continuous self that this system is accountable to. Trauma-informed design answers the question that immediately follows and is otherwise never asked: what is this system permitted to do to that self. An identity layer without a trauma-informed specification is a precise record of a self that the system is still free to extract from. A trauma-informed specification without an identity layer is a principle with no continuous subject to protect, reset every session, accountable to no one across time. They are halves of one structure.

Future Proof is built underneath that structure rather than beside it. The recurring pattern across this work, an identity layer that sits over the orchestration layer rather than competing with it, certification and insurance approached as the roots of a system that is hardening into a standard, an ecosystem designed to regenerate the participants in it rather than to extract from them, is the same pattern this paper has been describing the whole way down, expressed at the level of infrastructure rather than interface. An AI-era trust standard, taken seriously, is trauma-informed design made auditable: it is the question of what a system may do to the continuous self, written down precisely enough that an institution can be held to it and an investor can price the difference between a system that holds and a system that has merely not yet been caught extracting. None of this is a product description and none of it should read as one. It is the observation that the discipline this paper reconstructs from the clinic, the built environment, the stress-physiology literature, the human-factors canon, and the institutional-betrayal evidence converges on a single requirement, and that the requirement is being built underneath the AI era by the people who saw it as a substrate question and not a feature one. The reader who has followed the argument this far should finish it with the sense that this was already being operated at the root, quietly, before it had a name in this register. That is the correct sense. The discipline is old. Its application to the systems now deciding what people see, do, and conclude is what is new, and the timing of that application, against a regulatory frame hardening through 2026, is not incidental.

7. Implications

The argument is general. Its consequences are specific, and they differ by where the reader stands.

7.1 For institutions

For an institution, the operative finding is the institutional-betrayal result, and it reorders the priority list. Most institutional risk management is built to lower the probability of harmful events. The evidence says the larger, more durable injury, and therefore the larger durable liability, is the institution's conduct after harm has occurred: opacity, delay, denial, the transfer of fault onto the harmed party. An institution that is excellent at prevention and incompetent at repair is not a low-risk institution. It is an institution that has optimised the smaller variable and left the larger one to chance. The design instruction is to build the response to harm with the same rigour, ownership, and non-discretionary structure as the prevention of it, and to treat that response as infrastructure rather than as public relations. As the 2026 regulatory frame hardens around manipulation, opacity, and the overriding of human decision-making, the institutions that built repair and legibility into the structure early will be evidencing an existing practice. The institutions that did not will be retrofitting one under enforcement, which is both more expensive and, by this paper's own argument, less credible to the people it is supposed to protect.

7.2 For investors

For an investor, the finding is that extraction and holding are not points on a quality spectrum. They are different machines with different risk profiles, and the difference is becoming legible to regulators precisely as it becomes priceable. An asset whose growth depends on engineered unpredictability, manufactured choice, and the silent transfer of cost onto users is an asset whose core loop is, increasingly, a documented exposure rather than a moat. The diligence question is not whether a system is engaging. It is whether the engagement is inherent to the value delivered or manufactured against the user's nervous system, because the first compounds and the second is a liability waiting for an enforcement date. The durable position, and the one that becomes more valuable as the frame hardens, is the system that holds: it does not have a retrofit cost in front of it, it does not have a trust cliff under it, and it is the rarer asset, because extraction was the path of least resistance and most of the market took it.

7.3 For operators

For an operator, the finding is the most actionable, because it is the most controllable. Every principle in Section 5 is testable without a regulator, a budget cycle, or anyone's permission. The single highest-yield move is to relocate trauma-informed requirements from the end of the build, where they arrive as ethics review and lose every argument to a shipped deadline, to the start of it, where they sit as system requirements alongside correctness and latency and shape the architecture rather than apologising for it. The diagnostic question an operator can ask of any feature, in any review, costs nothing and reveals almost everything: where does this put the cost, and is any of the load it creates manufactured rather than inherent to the task. A team that asks that question in every review has internalised the discipline more deeply than a team with a policy document, because the discipline was never a document. It is a default about where costs are allowed to land.

7.4 For the people inside these systems

For the person inside the system, the finding reframes an experience that is usually mislabelled, and the reframing has force precisely because it is physiological and not moral. The fatigue, the bracing, the low background dread of environments that change without warning, extract without disclosing, and offer choices that are not choices, is not a personal deficit, a failure of discipline, or a character flaw to be managed with better habits. It is allostatic load, imposed by a designed environment, paid by a body that did not consent and largely cannot opt out. Naming it correctly does two things. It moves the fault from the person to the design, where the evidence says it belongs. And it converts a vague unease into a specific demand the person is entitled to make of the systems they depend on: be predictable, be legible, give me real choices, and when you harm me, repair it without making me fight you for it. That demand is not unreasonable, and it is not new. It is the oldest finding in the discipline, addressed for the first time to the systems that now set most of the conditions of a life.

8. Coda

Every system encodes a decision about where its costs are allowed to land. The decision is made whether or not anyone admits to making it, and for most of the systems now mediating attention, work, and judgement, it has been made the same way: the costs land on the person, because the person is the cheapest place to put them and the only party that cannot refuse delivery. Unpredictability is manufactured because it holds attention. Choices are narrowed because narrowed choices convert better. Failures are externalised because externalised failures are someone else's. None of it required cruelty. It required only an objective function and a gradient, and the gradient runs downhill toward the human every time the human is the path of least resistance.

Trauma-informed design is the discipline that refuses the gradient on purpose. It is not the claim that systems should be kinder. It is the older and harder claim that an environment is part of the diagnosis, that the body reaches the room before the mind does, that the cost of an environment is real and named and paid in a currency the occupant cannot opt out of, and that a system therefore carries a responsibility for what it does to the people who cannot leave it, a responsibility that does not lapse because the harm was efficient or the externalisation was profitable. A system built on that refusal is not a gentler version of the extractive machine. It is a different machine, with a different objective function, a different risk profile, and a different relationship to the nervous systems it touches.

The systems being built now, in the AI era, will set the conditions of attention, work, and decision for a long time, and the roots of the standard that will govern them are hardening this year. The only question that ultimately matters about any of them is the one the discipline has been asking since the clinic: not what can this system do, but what is it permitted to do to the person who cannot refuse it. The systems that were built underneath that question, rather than asked it after the fact, are the ones that will still be standing when it is no longer optional to have an answer.

References and Notes

The following sources are real and verifiable. They are the load-bearing references for the claims made above. Where a claim rests on a body of work rather than a single document, the canonical source is named.

1. Substance Abuse and Mental Health Services Administration (SAMHSA), SAMHSA's Concept of Trauma and Guidance for a Trauma-Informed Approach, 2014. The canonical articulation of the Four Rs (Realise, Recognise, Respond, Resist re-traumatisation) and the six key principles. Available via the SAMHSA Library and the National Child Traumatic Stress Network.

1. SAMHSA, 6 Guiding Principles to a Trauma-Informed Approach (infographic and guidance), archived via the United States Centers for Disease Control and Prevention document repository.

1. Bainbridge, L., "Ironies of Automation," Automatica, Vol. 19, No. 6, 1983. The foundational human-factors paper on automation, skill atrophy, the out-of-the-loop problem, and the displacement of failure onto the human operator.

1. McEwen, B. S., and colleagues, work on allostasis and allostatic load, including "Physiology and Neurobiology of Stress and Adaptation: Central Role of the Brain," Physiological Reviews, 2007, and the broader allostatic-load literature on the cumulative physiological cost of chronic stress activation.

1. Sapolsky, R. M., body of work on the neuroendocrinology of stress, social hierarchy, control, and predictability, including the synthesis in Why Zebras Don't Get Ulcers and associated primary research on subordination, controllability, and disease risk.

1. Freyd, J. J., and Smith, C. P., research on institutional betrayal and institutional courage, including the Project on Institutional Courage, and Smith, C. P., and Freyd, J. J., "Institutional Betrayal," American Psychologist, 2014. Empirical extension: research finding institutional courage buffers institutional betrayal, published in PLOS ONE, 2022.

1. Edmondson, A. C., "Psychological Safety and Learning Behavior in Work Teams," Administrative Science Quarterly, Vol. 44, No. 2, 1999, and The Fearless Organization, 2018. The foundational and applied statements of psychological safety, including the finding that its effect strengthens under uncertainty and complexity.

1. Trauma-informed design of the built environment: practitioner and peer-reviewed literature including the scoping review on trauma-informed design of supported housing through the lens of neuroscience (PubMed Central) and the applied principle sets published by architectural practices specialising in supportive housing and dignified design. Recurring principles: diffuse warm light, connection to nature, prospect and refuge, predictability of structure, dedicated spaces for self-regulation, and occupant control.

1. Center for Humane Technology and associated practitioners, public statements and analyses of the attention economy, intermittent reinforcement, infinite scroll, and engineered unpredictability as extractive design mechanisms. The persuasive-technology lineage and the "race to the bottom of the brain stem" framing.

1. European Union, Regulation (EU) 2022/2065, the Digital Services Act, including Article 25 on the prohibition of deceptive and manipulative online interface design; enforcement in force with penalties referenced to global turnover. European Commission communications on the preparation of a Digital Fairness Act addressing dark patterns and addictive design. United States Federal Trade Commission policy statement (2021) treating dark patterns as potentially unfair or deceptive practices.

1. European Union, Regulation (EU) 2024/1689, the Artificial Intelligence Act. Full application 2 August 2026; general-purpose AI model obligations from 2 August 2025; high-risk obligations and Commission enforcement powers from 2 August 2026; extended transition to 2 August 2028 for certain embedded high-risk systems under the AI omnibus simplification. Provisions on manipulative and exploitative practices and transparency of AI-mediated interaction. Source: European Commission, Shaping Europe's digital future, AI Act pages, and the published implementation timeline, as of 2026.

1. van der Kolk, B., The Body Keeps the Score: Brain, Mind, and Body in the Healing of Trauma, 2014, cited here only as the most prominent popularisation of trauma in public discourse, and read alongside its scholarly critique, including a 2025 evaluation in BJPsych Bulletin assessing a large set of the book's specific claims. This paper deliberately does not rely on the contested mechanistic assertions and rests only on the conservative core of stress physiology and human factors that the critique leaves intact.

1. Empirical and regulatory material on deceptive consent design: studies of dark patterns in cookie and consent interfaces across large website samples (including joint work by the Karlsruhe Institute of Technology and the IT University of Copenhagen), and European Data Protection Board guidance that deceptive design can defeat the requirement that consent be freely given, specific, informed, and unambiguous under the General Data Protection Regulation (Regulation (EU) 2016/679). The concepts of consent fatigue and banner blindness are drawn from this literature.

A note on method. Where this paper makes a structural claim it does not attach a fabricated statistic to it. The neuroscience of allostatic load, the human-factors evidence on automation, the institutional-betrayal findings, and the psychological-safety research are cited at the level of their canonical sources rather than reduced to single decontextualised figures, because the structural truth is more durable than a borrowed number and does not degrade when the number does. Regulatory dates are stated as of 2026 and were verified against the European Commission's published material; readers consulting this paper later should confirm against the current text of the instruments named.

Future Proof Intelligence . Research . No. VI . MMXXVI