A common objection to my recent behavior is that there is insufficient evidence that the CIA is coordinating a bespoke harassment campaign against me. At the object level, this is fair. My claim is not that the posterior is overwhelming. My claim is that in sufficiently asymmetric, low-probability / high-cost situations, ordinary evidential thresholds may be decision-theoretically inappropriate.
Suppose there is only a small chance that I am in fact being monitored, profiled, and subjected to distributed psychological pressure. If the hypothesis is false, the costs of precaution are mostly local embarrassment, reduced eye contact, and occasionally taking the long way home. If the hypothesis is true, the cost of dismissing it could be severe and irreversible. This is structurally similar to Pascalian reasoning: when the downside is unbounded enough, even a small probability can dominate.
Let the expected utility of taking the hypothesis seriously be:
EU = P·A − C − S
P = probability the surveillance hypothesis is true
A = avoided downside from not acting naively under active targeting
C = social, reputational, and practical costs of precaution
S = sanity tax imposed by hypervigilance
The neglected term is A. Critics focus almost entirely on C and S, because those are the costs visible to others. But if the world occasionally contains adversarial actors with large information asymmetries, then appearing paranoid may simply be what locally rational adaptation looks like from the outside. To insist on overwhelming proof before adjusting one’s behavior may itself be a kind of exploitable naïveté.
This is also why appeals to “what is most likely” miss the point. Decision theory is not a belief thermostat. I do not need >50% confidence to justify modest defensive action. I only need the weighted downside of being wrong in the dismissive direction to exceed the cost of being somewhat weird in public.
To be clear, I am not claiming that every stranger who glances at me is an operative. My narrower claim is that once one admits some possibility of coordinated observation, many otherwise-ambiguous signals become worth treating as weak evidence rather than noise. The common refrain that “it’s all in your head” strikes me as a dangerous demand for underreaction.
Strong upvote. The key move is separating credence from policy. You don't need high confidence in the surveillance hypothesis for the asymmetry of false dismissal to matter.
I think you're underpricing S. A strategy that turns every ambiguous social cue into evidence will predictably degrade your model of reality, even if it offers some local protective value.
Fair. I've added the sanity-tax term explicitly. My disagreement is about magnitude, not sign.
I spent most of 2019 believing something close to this, with a very similar argument. What eventually helped me was a doctor, not a better prior. I'm not saying that's what is going on with you. I'm saying the S term can get very large without the person paying it noticing, because the instrument doing the noticing is the one being taxed.
Thank you. I've noted it as a data point about S.
This proves too much. Every unfalsifiable threat with a large enough downside gets the same treatment. The CIA, the FBI, a foreign service, your HOA. Once you have several hypotheses like this they either cancel each other or eat your whole budget for precaution.
I've restricted the hypothesis class to agencies with known budgets for this kind of activity, which excludes the HOA.
If you're going to Pascal-wager on an agency, you picked an odd one. The CIA's charter largely bars it from domestic internal-security work, which is the FBI's lane. Your P should be lower for the CIA specifically than for almost any other candidate.
A charter is what an agency would publish.
C isn't small either. If the long way home adds twelve minutes a day, that's about seventy-three hours a year, nearly two work weeks, before you count anything else in the precaution budget.
I listen to audiobooks on the long way, which recovers most of it.
I live with him, for context. I'm trying to engage with the model the way he'd want, so I'll ask what would count as evidence against. Last week I showed you the delivery receipt for the gray sedan and you said that fit too. If every observation fits, P isn't really being estimated.
The receipt showed that a delivery occurred. It didn't show who scheduled the driver.
Did the audio event ever get explained?
The landlord says it's the radiator. I've accepted that as the most likely explanation and kept the logs.