The word “safety” ends arguments. That is its function in 2026, and it performs that function across every domain of contested expert authority. To invoke safety is to convert a coalition’s preferred conclusion into a moral imperative, to reclassify dissent as recklessness, and to place the burden of proof on anyone who questions the claim rather than on anyone who makes it. No other word in the vocabulary of expertise does this work so efficiently, and no other word is so rarely examined for what it means.
\nWhat it means depends entirely on who is defining it, for whom, and through what process. These questions are almost never asked in public discourse, because the word itself forecloses them. This essay maps how that foreclosure works, what it costs, and where in 2026 it is operating most consequentially.<\/p>\n
The Core Distinction<\/strong><\/p>\n Before the map, one distinction needs to be stated clearly and held throughout. The Template<\/strong><\/p>\n The mechanism is consistent enough across domains to state as a template, which can then be applied to any specific case. Mental Health: The Clearest Current Case<\/strong><\/p>\n The mental health domain offers the most fully developed current example of the template in operation, because it combines all the elements with unusual clarity: the institutional versus experiential gap, the hero system, the moral language, the suppressed adverse event data, and a political reckoning now underway that is creating natural experiments. The GRAS Architecture: Safety Certified by the Interested<\/strong><\/p>\n The food supply offers a different but structurally identical case, with the mechanism partly codified in law. The Generally Recognized as Safe designation, GRAS, allows companies to introduce new food chemicals without FDA review if qualified experts determine the substance is safe. The original intent was administrative efficiency for common substances. What evolved was a system in which companies hire their own experts, conduct their own reviews, and declare their own ingredients safe without notifying the FDA at all. AI Safety: Two Coalitions, One Word<\/strong><\/p>\n The AI safety domain is unusual in the template because two opposed coalitions are fighting over the same word simultaneously, each claiming that the other’s definition of safety is itself dangerous. This doubles the epistemic coercion rather than canceling it. The Recursive Loop and Its Breaking<\/strong><\/p>\n In a domain with genuinely good evidence and honest inquiry, ongoing disagreement about mechanisms, edge cases, and applications persists even when the central finding is secure. Premature closure of that disagreement, the declaration that the debate is over and further discussion endangers people, is the signature of coalition management rather than scientific maturity. Consensus achieved through the suppression of dissent rather than the resolution of disagreement is not epistemic achievement. It is epistemic coercion wearing achievement’s clothing. The Questions That Are Crying Out<\/strong><\/p>\n The template points to a consistent set of questions that the dominant safety coalitions across all these domains have structured themselves to not ask. Each one is an investigation priority. What the Map Shows<\/strong><\/p>\n Safety claims fail in a consistent pattern. The coalition defines safety in institutional terms that align with its interests. It selects the risks it tracks and suppresses the ones it does not. It treats its normative and procedural choices as purely technical ones. It uses the moral language of protection to convert those choices into imperatives. It enforces the definition through the standard mechanisms of epistemic coercion: information deprivation, normalization and stigmatization, legitimation and delegitimation. The word “safety” ends arguments. That is its function in 2026, and it performs that function across every domain of contested expert authority. To invoke safety is to convert a coalition’s preferred conclusion into a moral imperative, to reclassify dissent … Continue reading
\nInstitutional safety is what coalitions of experts, regulators, licensing bodies, and professional associations define as safe. It is standardized, documented, legally defensible, and aligned with established protocols. It is safe for the system: for the physician who followed the guideline, the regulator who approved the protocol, the institution that implemented the procedure. When something goes wrong inside institutional safety, the question is whether the procedure was followed correctly, not whether the procedure was the right one.
\nExperiential safety is what the person subjected to the intervention experiences. It is variable, context-dependent, often not captured by the metrics the system uses to evaluate itself, and sometimes directly opposed to what the institution defines as safe.
\nThe gap between these two things is where the most important unasked questions in contemporary expertise live. A practice can be institutionally safe because it is standardized, defensible, and authorized while being experientially unsafe for the person subjected to it. The system calls this gap a problem of compliance, of communication, of patient factors, of the necessary costs of treatment. What it rarely calls it is a problem with the definition of safety itself.
\nStephen Turner’s framework<\/a> explains why. The coalition controls not only what counts as admissible evidence but what counts as an admissible outcome. If the outcomes tracked by the system are the ones the system is designed to produce, and the outcomes not tracked are the ones the system is designed not to see, then the system can validate itself indefinitely regardless of what is happening to the people it processes. This is not conspiracy. It is what Turner calls the recursive loop: the coalition defines safety, delivers interventions, evaluates outcomes using its own metrics, and validates itself. Negative experiences can always be reframed as part of the process, as resistance, or as evidence of need for more treatment.
\nErnest Becker<\/a> adds a layer Turner does not. The hero system is not just a career structure. It is a meaning structure. Professionals within a safety coalition are not simply protecting their income when they defend the coalition’s definitions against challenge. They are protecting the justification for their professional existence. To admit that the protocols cause systematic harm, that the interventions retraumatize more than they heal, that the safety claims were not grounded in the evidence the system claimed, is not just a career risk. It is an existential one. This is why coalitions absorb anomalies rather than learning from them, why they reframe failures as the cost of success, and why the reckoning, when it comes, rarely produces genuine accountability from the people who sustained the false consensus.<\/p>\n
\nA coalition forms around a shared stake in a particular definition of safety. The coalition controls who produces the relevant evidence, who reviews it, what counts as the appropriate evidentiary threshold, and what the default assumption is when evidence is absent or contested. It presents these procedural and normative choices as purely scientific ones. It uses the word safety to convert those choices into moral imperatives, so that questioning the definition is framed as endangering people rather than as legitimate inquiry.
\nThe coalition then selects the risks it elevates and the risks it suppresses. Elevated risks are the ones the coalition’s interventions address. Suppressed risks are the ones the coalition’s interventions create. The asymmetry is not random. It follows the incentive structure: elevated risks justify the coalition’s authority and funding; suppressed risks would undermine both. Liability protection runs in the direction of the elevated risks. No one is sued for following the guideline.
\nMoral language<\/a> stabilizes the definition and recruits allies. The vocabulary of care, protection, help, and support makes resistance look like denial, irresponsibility, or cruelty. This is not merely rhetorical. It shapes what questions can be asked in a grant application, what a peer reviewer will accept, what a journal will publish, and what a clinician will say out loud in a departmental meeting.
\nWhen challenges to the definition appear, the coalition manages them through the three mechanisms Turner identifies. Information deprivation limits what gets studied: adverse event data is not collected systematically, long-term follow-up is not funded, and preregistered trials that might produce inconvenient results are not required. Normalization and stigmatization ensure that when inconvenient findings appear, they are framed as methodologically flawed, ideologically motivated, or dangerous to share publicly. Legitimation and delegitimation determine whose voice counts: only credentialed insiders can speak authoritatively, and the credentials are issued by the coalition itself.
\nThe result is the recursive loop, closed from the inside. Breaking it requires either an institutional reckoning the coalition resists by design, or an accumulation of observable reality that the normalization machinery can no longer contain. When the loop does break, the coalition pivots without accounting for what it suppressed or who it punished for being right early.<\/p>\n
\nThe dominant coalition in mental health safety defined safety operationally as entry into the system. Distress equals risk. Risk requires intervention. Intervention means professional treatment following established protocols. Not intervening, or intervening in ways not sanctioned by the coalition, was itself defined as unsafe. This definition served the coalition’s interests precisely: it justified expanding authority, increasing funding, and treating any resistance as evidence of the severity of the problem rather than as legitimate criticism of the solution.
\nWithin this framework, inpatient psychiatric units, coercive holds, mandatory screenings, restraints, and rapid affirmation pathways for gender-distressed youth were all safety measures by definition. The evidence that these interventions sometimes produce the harms they claim to prevent was not admissible within the coalition’s evidentiary framework. Adverse event tracking in psychiatric settings has been systematically weak. Post-discharge outcome data has been sparse and rarely preregistered. Research on retraumatization in inpatient settings, on post-discharge suicide spikes, on regret and desistance rates, and on the iatrogenic effects of coercive intervention has been underfunded, underreported, and when produced, treated as politically motivated rather than scientifically relevant.
\nThe gender medicine case has become the sharpest point of political contestation. For years, the justification for rapid affirmation pathways and medical interventions in gender-distressed youth was, explicitly, safety: these young people are at risk of suicide, and denial of affirmation is dangerous. This framing did the work of placing any questioning of the pathways outside admissible discourse. To ask about evidence quality, about long-term outcomes, about regret and desistance rates, was to be indifferent to child suicide. The moral language closed the epistemic question before it could be asked.
\nThe Cass Review in Britain and subsequent systematic reviews commissioned after 2024 policy shifts found what the coalition’s definition of safety had placed outside admissible reality: very low certainty evidence that these pathways improve long-term mental health outcomes, and documented iatrogenic harms including infertility, sexual dysfunction, bone density loss, and persistent psychological difficulty. The Supreme Court upheld state restrictions. Federal policy shifted. Natural experiments are now underway in states with different policies, producing the comparative outcome data the prior coalition actively resisted generating.
\nThis is the pattern Turner identifies in the history of science: suppressed observations do not win by defeating the coalition on its own terms. They accumulate until the framework can no longer contain them, and then the coalition pivots. What the pivot does not include is accountability for who was punished for raising the questions earlier, or honest examination of why the adverse event data was not collected in the first place.<\/p>\n
\nSince 2000, nearly 99 percent of new food chemicals entered the American market through self-affirmed GRAS determinations, according to Environmental Working Group analysis of FDA records. Fewer than one percent went through the formal petition process Congress intended as the primary route. The experts who certify the safety are paid by the companies that profit from the determination. The determinations are often proprietary. The FDA has not considered this an admissibility problem, because the coalition that defined food chemical safety included the regulatory culture that built and maintained the system.
\nThe institutional safety claim here is the GRAS determination itself: qualified experts have reviewed this substance and found it safe. The experiential reality is that hundreds of substances banned in Europe, Canada, and Australia because of plausible health risks have circulated freely in the American food supply for decades under self-affirmed GRAS status. The substances do not change at national borders. The coalitions controlling the definitions do.
\nIn March 2025, HHS Secretary Kennedy directed the FDA to explore eliminating the self-affirmation pathway. The FDA’s own description of the system was that it was deeply flawed. That description represents a coalition shift, not a scientific discovery. The underlying evidence about these substances had been available, in various forms, throughout the period when the coalition was calling the system adequate. What changed was not the science. What changed was who controlled the definition of admissible reality.<\/p>\n
\nThe alignment-focused safety coalition defines the primary risk as catastrophic or existential harm from advanced AI systems that pursue goals misaligned with human values. Its admissible evidence is formal modeling of potential failure modes, benchmarks designed to detect dangerous capabilities, and arguments from first principles about the trajectory of increasingly capable systems. Its moral language is human survival. Its institutional position is partly secured through the AI Safety Institute and related regulatory bodies. Dissent from this definition is framed as recklessness about civilizational risk.
\nThe sovereign-growth coalition defines the primary risk as regulatory capture by incumbent AI companies using safety claims to raise barriers against competitors, and as government or corporate censorship enabled by safety frameworks that restrict what AI systems can say or do. Its admissible evidence is economic analysis of regulatory costs, examples of safety claims that served incumbent interests, and arguments about the value of open development and competition. Its moral language is anti-tyranny and democratic access to technology. Dissent from this definition is framed as naivety about who benefits from restricting AI.
\nEach coalition has the incentive structure Turner describes: the safety concerns it elevates justify its institutional position and funding, while the safety concerns it suppresses would undermine both. Neither has much institutional incentive to say that the genuine uncertainty about what advanced AI systems will do is distributed across the gap between the two coalitions in ways that current research cannot resolve. That honest position does not sustain a grant application, a regulatory proposal, or a legislative testimony.
\nThe questions that fall outside both coalitions’ admissible reality are revealing. Whether current AI systems might have morally relevant inner experience is treated as either obviously absurd or dangerously irresponsible to raise publicly, depending on which coalition’s framing one accepts. Neither response engages the question. Whether AI systems trained on human-generated content at massive scale are subtly reshaping the epistemic environment in ways their developers do not fully understand is difficult to publish on, because it implicates the interests of the companies funding the research. The FDA’s own AI tool is already shaping regulatory document review. Whether and how that changes regulatory judgment is not an admissible research question within the institutions that might investigate it.<\/p>\n
\nThe Becker layer explains why the loop is so hard to break from inside. The professional who has spent twenty years building a clinical practice around a particular safety protocol, who has trained students in it, who has testified to legislatures defending it, who has published research validating it, does not experience a challenge to that protocol as an invitation to revise a theory. They experience it as a threat to the meaning structure of their professional life. The hero system is not just a career. It is the answer to the question of what one’s work has been for. Coalitions sustain false consensus not only through external enforcement but through the internal psychological necessity of their members.
\nThis is why the reckoning, when it comes, takes the form of pivot rather than reckoning. The coalition does not say: we were wrong, we suppressed the people who were right, we owe them an accounting, and here is what we will do differently. It says: the evidence has evolved, our understanding has deepened, we are updating our guidance. The pivot reframes what happened as normal scientific progress rather than as a failure of the epistemic machinery. The people who raised the questions early and were punished for it remain punished. The people who sustained the false consensus retain their institutional standing. The asymmetry is structural, not incidental.<\/p>\n
\nWho is collecting standardized adverse event data, and who is not? In mental health, systematic harm monitoring from psychotherapy, inpatient treatment, and pharmacological intervention is astonishingly weak given the scale of the interventions and the vulnerability of the populations. In food chemistry, the GRAS system was explicitly designed to avoid collecting the data that would challenge it. In AI, post-deployment outcome tracking on actual users is largely proprietary. The absence of data is not neutral. It is the information deprivation mechanism operating at the design stage.
\nWhat risks are the coalition’s metrics counting, and what risks are they not counting? Every safety protocol encodes a theory about what harms matter. Institutional mental health safety counts acute crisis events. It does not count retraumatization, dependency, or resilience erosion from over-protection. Food chemical safety counts acute toxicity. It often does not count endocrine disruption, developmental effects, or cumulative exposure across a diet of ultra-processed foods. AI safety debates count the risks each coalition has a stake in. Neither counts what the other is looking at.
\nWhat would the evidence look like if the definition of safety were reversed? If the burden of proof ran the other direction, demanding that interventions demonstrate benefit rather than demanding that challenges demonstrate harm, what fraction of current safety-labeled practices would survive? This is not a rhetorical question. It is the difference between the American and European approaches to chemical safety, and it produces systematically different food and drug supplies in comparable wealthy democracies.
\nWho bears the cost when the false consensus holds, and who bears the cost when it breaks? The people subjected to retraumatizing psychiatric protocols bear the cost of the false consensus holding. The clinicians who designed and defended those protocols bear no cost when the consensus breaks and guidance is updated. The families who consumed food chemicals under self-affirmed GRAS determinations bear the cost of the prior definition of safety. The companies that profited from those determinations and the regulators who permitted them bear no cost when the chemicals are eventually restricted. This asymmetry is not a side effect of the mechanism. It is the mechanism. It is what makes the coalition stable, and what makes it dangerous.<\/p>\n
\nWhen observable reality accumulates past the point of management, the coalition pivots without accounting for what it suppressed. The lesson it draws from the failure is that communication needs improvement, or that more research is needed, or that guidelines will be updated. The lesson it does not draw is structural: that the definition of safety was serving the coalition’s interests rather than the people in whose name safety was invoked.
\nIn 2026, across mental health, food chemistry, AI governance, environmental health, and pharmaceutical regulation, the most consequential question is not whether a given claim is true. It is what the claim is forcing us to ignore in order to keep the coalition together. The answer to that question, in case after case, is the harm the coalition’s own interventions are producing, measured by the standards the coalition has structured itself not to apply.
\nSafe for the system. Not always safe for the person. The gap between those two things is where the investigation needs to go.<\/p>\n","protected":false},"excerpt":{"rendered":"