AI safety warnings raise threat appraisal sky high. Without coping appraisal, they produce fear control, not safety.
I kept noticing the same pattern across every AI safety announcement I read this year. An open letter warns that AI systems capable of catastrophic harm are coming soon and we are not ready. A think tank publishes a scenario where advanced AI causes mass unemployment before anyone has time to retrain. A leading researcher testifies to Congress about extinction risk. Every one of these is making a threat appraisal argument. And almost none of them follow it with a coping appraisal. The structure is Protection Motivation Theory, Rogers 1975 and 1983, applied to artificial intelligence risk, except nobody is using the name.
I wrote earlier about why fear alone does not make people secure in the cybersecurity context. The mechanism is the same here, and the stakes might be higher. PMT explains protective behavior through two parallel appraisals. Threat appraisal asks how severe the consequences would be and how vulnerable I am to them. Coping appraisal asks whether the recommended protective action will actually work and whether I can do it. Fear motivates protective behavior only when both appraisals are positive. High threat without high coping does not produce protection. It produces what Boss and his colleagues, applying PMT to information security in 2015, called fear control. The person manages their emotional fear rather than the actual danger. They deny the risk, they avoid the information, they decide the warnings are overblown.
Now look at the AI safety discourse through this lens. Threat appraisal is running at maximum. The severity of the threat is described in language that borrows directly from nuclear risk: catastrophic, existential, irreversible. The vulnerability is universal. Every person, every organization, every government is exposed. The scenario is global and it could happen soon. The threat side of PMT could not be stronger. And the result has been a wave of public concern followed by a noticeable shift toward dismissal. The more apocalyptic the warning, the more the response from the broader industry and the public has shifted toward a kind of learned denial. The risks are overblown. The timeline is speculative. The people warning about extinction are the same people building the systems. This is fear control unfolding exactly as PMT predicts.
The coping side of PMT is where the AI safety picture falls apart. What is the recommended protective action in most AI safety warnings? Sign the letter. Pause training. Support a moratorium. These are not coping appraisals in the PMT sense, because they fail both efficacy questions. Response efficacy: will signing a letter actually reduce AI risk? The link between the action and the outcome is uncertain and indirect. Self-efficacy: can I personally do anything that changes the trajectory? The answer for most people and even most organizations is no. So the coping appraisal fails, the threat appraisal stays high, and the individual or organization slides into fear control. The mechanism is textbook.
Johnston and Warkentin demonstrated in 2010 that PMT predicts security policy compliance only when both appraisals are strong. Their finding transfers directly to AI safety messaging. Raising threat perception without raising coping perception is not just neutral. It can be counterproductive. You can make people less likely to engage with AI safety by making the problem appear too big for any individual response.
The difference between AI companies that produce actual danger control and those that produce fear control is visible if you know what to look for. Anthropic has built a safety culture around specific, repeated protocols: red teaming exercises with documented findings, deployment thresholds tied to capability evaluations, a responsible scaling policy that names concrete conditions for release. These are coping appraisals in action. They give employees and external observers a set of actions with a plausible link to outcomes. Response efficacy comes from the documented track record of internal testing catching problems before deployment. Self-efficacy comes from the fact that an individual researcher can run a specific test and see the result. The company is not free of fear control dynamics, but the structure pushes toward danger control.
The broader industry operates differently. The main output is warnings. Signed statements about existential risk. Frameworks that define levels of capability without defining what happens at each level. Benchmark papers that show a model can pass a dangerous test without specifying what organizational routine prevents that model from being deployed. The threat content is high. The coping content is minimal. And the result is what I think the PMT lens predicts: a field that produces high public anxiety about AI risk alongside very little organizational change in how AI systems are actually built and released.
I think this is why regulation might be the most effective intervention available. Boss et al. drew on the extended parallel process model to show that only danger control produces adaptive security compliance. Regulation introduces a form of coercive coping appraisal. It mandates specific actions with specified consequences for noncompliance. A company that must file a safety case before deploying a model cannot default to fear control about that model. The regulation forces a coping response even when the organization lacks intrinsic efficacy beliefs. The response efficacy and self-efficacy are not delivered by the regulation itself. They are constructed through the organizational routines that compliance demands. This is not a perfect solution. Coercive coping can produce shallow compliance that satisfies auditors without changing behavior, exactly like the security training problem I wrote about before. But regulation at least creates a forcing function for coping appraisal that the current voluntary threat-appraisal approach does not.
My opinion is that the AI safety field needs to study Boss et al. 2015 directly. Not because it maps neatly onto security behavior, but because PMT predicts the current state of AI safety discourse with disturbing accuracy. High threat perception. Low coping efficacy. Widespread denial dressed up as reasoned skepticism. The field is producing fear control at scale and calling it awareness. The papers are all on the threat side. The signed letters are all on the threat side. The congressional testimony is all on the threat side. And the result, which PMT told us about forty years ago, is that people who feel powerless against a threat stop believing in the threat rather than taking action against it. The field will not produce safety until it produces coping appraisal that matches the severity of the threat it describes.