Joseph Bronski is an independent researcher attempting to create a science of power, which he calls exousiology. He draws inspiration from the old elite theorists, primarily Pareto, with the intention of rebuilding the field on a rigorous, almost obsessively mathematical basis, incorporating the psychology of individual differences and economic models. His research program concept is parallel to mine, except I begin from a rigorous definition of conspiracy theory as the use of deception by elites and note the connection to elite theory after the fact. He construes power as deriving from three sources: violence, wealth creation, and deception. Recently however he has argued that deception is not in fact real. Given my previously stated bias on this matter, I can only consider this to be a ridiculous and untenable position that would vitiate the entire research program. I begin by critiquing Bronski’s model and then proposing my own model.
Bronski begins by laying out two positions, which he attributes to Mosca and Pareto respectively, the latter of which he supports. The Moscaite position holds that information precedes power and that history can be explained by the rise and fall of ideologies. The Paretian position holds that what people say need have no bearing on how they actually behave, and so inculcating people with false beliefs can’t actually make them display harmful behaviours, it can only make people signal adherence to the dominant ideology. I consider both positions partially true and mostly wrong. The Moscaite position overstates the role of information; though information does not precede power, for those who are already powerful deception can be a more efficient means of control than force or economic coercion. The Paretian position is correct that people may signal through expressing beliefs that don’t align with their behaviours, but incorrect that people cannot be harmed by deceptive information.
To demonstrate the validity of the Paretian position, Bronski limits his real world examples exclusively to the study of Wokism. Noting the increased rates of criminality among minorities compared to whites he focuses on white flight: the tendency for whites to move from areas of high ethnic diversity to homogeneously white areas. Since white liberals claim to love diversity and may be deceived about crime rates, the phenomenon should be limited to conservatives who acknowledge the crime data and hence seek areas where the risk of being victimized is lower. Instead, across the US and the UK, anti-immigration whites are only marginally more likely to move to homogeneous neighbourhoods than pro-immigration whites. This suggests that praise for diversity among white liberals is performative signaling, rather than an honestly held belief. Another example, not included Bronski’s essay, is the Woke slogan “trans women are women”. Despite mouthing the slogan, leftists do not in fact behave as if trans women are women; they are not deceived about the differences between trans and biological women. Such a slogan cannot but be a signal of ideological conformity, seeing as it so frequently and obviously conflicts with objective reality. That being the case, the following Theodore Dalrymple quote is appropriate:
“In my studies of communist societies, I came to the conclusion that the purpose of communist propaganda was not to persuade or convince, not to inform but to humiliate; and therefore, the less it corresponded to reality the better. When people are forced to remain silent when they are being told the most obvious lies, or even worse when they are forced to repeat lies themselves, they lose once and for all their sense of probity. To assent to obvious lies is in some small way to become evil oneself. One’s standing to resist anything is thus eroded, and even destroyed. A variety of emasculated liars is easy to control. I think if you examine political correctness, it has the same effect and is intended to.”
So while Wokism is an elite-sponsored control mechanism, it is not really intended as deception1, and to use it as an example disproving deception is fallacious. If he wants to specifically argue that the phenomenon of Wokism is mostly a result of signaling rather than authentic belief that is one thing, but it is disingenuous to either redefine deception or else imply that Wokism is the only kind of deception that exists. Bronski also does not address the possibility of self-deception at all, or other psychological phenomena like irrational emotional attachment to beliefs.
Regardless, the real meat of the argument is a mathematical model that Bronski proposes to disprove deception, so let’s look at that:
The model starts by representing an information receiver as a rational Bayesian utility maximizer (RBUM) that takes whatever action would maximize the expected value of his utility given his beliefs about the state of the world. It models deception as a signal from an information sender about the state of the world intended to modify the receiver’s beliefs and induce him to change the action he takes. The question is, how the receiver will interpret the deceptive signal? “By definition, he does so in a way which maximizes his utility.” Therefore he will only update his beliefs about the world if the update would improve his utility. Since deception would result in his utility being lowered, he will reject deceptive updates. Bronski concludes “rational Bayesian utility maximizers will not be deceived in the long run, because they would revert updates that failed to give their promised utility returns.”
This last sentence is a tacit admission that deception is real, since even under this model RBUMs will not be deceived in the long run. This means that there will be a period of time during which the RBUM will experience a loss in utility due to updating on a deceptive signal. The model provides no information about how long ‘the long run’ is, nor does it explain how a RBUM will know which change to revert once they discover that their utility is reduced. Basically the model is saying, in complex mathematical terms, that if someone realizes they are being harmed by believing a lie they will stop believing it. This necessarily requires that they were harmed by the lie in the first place, which means deception works. The only way deception wouldn’t work is if the RBUM could always accurately assess the future utility cost of updating their world model before making the update, which in the real world would require complete knowledge and infinite computational resources, i.e. omniscience.
Bronski bases his ‘no deception’ theory on a mathematical analysis of Bayesian persuasion and a lab study that tests the model empirically. Crucially, the Bayesian persuasion model relies on a symmetric information environment – both the sender and receiver of a signal have access to the exact same information. When Bayesian persuasion is tested in the lab, the model accurately predicts the receiver’s behaviour 91% of the time, which is good, but it shows that even under controlled laboratory conditions with identical information 9% of people still do not behave so as to maximize their utility. Real world scenarios do not involve senders and receivers with symmetric information, and indeed asymmetric information is a requirement for deception – people can be deceived only when they lack information (or information processing ability). Bronski claims that he has extended the model to the non-symmetric case but offers no mathematical proof of this claim.
People are, of course, not rational Bayesian utility maximizers, though I do agree they are roughly Bayesian. People don’t actually maximize utility, except when maximizing utility is tautologically defined as the result of whatever actions the person decides to take. Utility is subjective, and there is no cardinal operational measurement of it. Even when we substitute something we do have a cardinal operational measurement for, biological fitness (number of gene copies passed on), we still find that people are not fitness maximizers, otherwise all men should be donating to sperm banks. People have genetically-based life history strategies, and evolution selects for individuals who have adaptive strategies given the current ecology, but no one is actually trying to maximize their fitness. A rational Bayesian utility maximizer is analogous to spherical cows in a vacuum but for economists (or exousiologists, as the case may be).
Any model of human behaviour that neglects psychology is bound to be inaccurate. Consider the case of a lost wallet. A utility maximizer would be likely to keep the money in the wallet, especially if it was a large amount of money, since that would maximize their utility. This would be contradicted by the data: a large (n = 17,000) study found that people are in fact more likely to report a wallet missing when it contains money, monotonically increasing for the amount of money it contains (three data points from $0 to $94.15 tested). This is in direct contradiction to the empirically measured predictions of both laypeople and economists. By controlling for other variables, the researchers determined that this behaviour was due to the psychological cost of being a self-perceived thief. By neglecting to model the difficult to measure psychological cost of having to update one’s self-concept in a negative way, the predictive validity of the economic model was undermined.
I claim there is also a psychological impetus towards a ‘no deception’ model in Bronski’s case. Bronski has repeatedly stated that he does not believe that there are, barring a few mainstream-admitted examples like the Gulf of Tonkin incident, any hoaxes that could influence the public. He calls such a belief in hoaxes degenerate – as in entropic or disorganized – and ‘schizo’. Certainly belief in hoaxes is not a high status, sexy position, and it incurs scorn from others (a position incurring scorn does not mean it is wrong, cf. Semmelweis). If Bronski wishes to develop an attractive research paradigm, taking the ‘no deception’ position is advantageous in multiple ways: it permits a sexy2 and tractable mathematical model, it summarily invalidates potential positions that may draw derision towards the program, and it contests a dominant paradigm in the cultural space in which he is operating (the ‘Wokism as mind virus’ model among the right). Bronski is also twice vaccinated for COVID, at material harm to himself, and defends that decision to this day. As he is someone who emphasizes personal agency and the ability to think for himself, it would incur a large psychological cost to update his beliefs on this matter (see also the Scott Adams vaccine saga). Thus, a priori, deception cannot be real.
Here I present a competing mathematical model of deception, also using a Bayesian model:
Deception proceeds in two phases: in phase 1 a deceptive signal updates an individual’s beliefs, and in phase 2 the update is reversed due to disconfirmatory information. In phase 1 a deceptive signal is sent which may update the posterior of the receiver. The signal may be deceptive because it includes data outside the real range of the distribution (fabrication) or because it leaves out probability mass from the distribution (omission). The magnitude of the update to the posterior is determined by epistemic vigilance, which is a function of both the sender and the content. Certain senders are trusted more than others, and a sender may only be trusted on some content and not others.
Epistemic vigilance scales the update to the posterior. When it is 0 the sender is completely trusted, and when it is 1 the sender is completely distrusted and no update is made.
Where P(A) is the prior. In phase 2, disconfirmatory information D is generated by a Poisson process D ~ Poisson(λ) where λ represents the ease of access to disconfirmation. Disconfirmations are assumed to be generated by direct observation not mediated by a sender, and so are not subject to epistemic vigilance. Any instance of disconfirmation may come with an attendant utility cost u(D) ≤ 0. It is not likely a priori that humans are perfect Bayesian updaters. They may overupdate on some highly salient observations. To model this, we propose that updates are made based on a perfect Bayesian base rate and a salience factor based on utility costs. The size of an update given some observation is greater when the attendant utility cost is greater. This can be modeled by repeatedly updating on the same observation, proportional to the magnitude of the utility cost.
Deception is not possible when λ is large, especially if E[u(D)] ≪ 0.
Furthermore, an individual’s priors can be conceptualized as a hierarchical Bayesian model where a learning rate parameter determines how sensitive higher order priors are to changes in changes in lower order priors. Gell-Mann amnesia is a special case where the learning rate parameter is set too low on the hierarchical model underlying the epistemic vigilance function. Delusional epistemic hypervigilance results when this same learning rate is set too high. The specific design of this model is left as an exercise for the reader.
Based on this model, the most effective deceptions would be those where disconfirmations are difficult to access, especially if there is a high utilty cost when they do occur. Deceptions that have the effect of reducing the odds that an individual will seek disconfirmation are also likely to be effective, e.g. warning that a particular situation is dangerous when it isn’t. Deceived individuals are less likely to seek disconfirmation when they believe that there is risk in doing so. This can reduce the ability to capitalize on available opportunities.
Although I hold that the theory behind this model more closely matches reality than does Bronski’s, in the real world none of these functions are actually operationally measurable and computable (the same applies to all of Bronski’s math). Therefore I present this model not as a complete and accurate model of deception, but as a starting point to begin thinking about the concept. We may hope that in the future there will be ways of empirically estimating these quantities like there is for fixation index (FST), but there is much work to be done before that is possible. Despite the difficulties with physical measurement, a simulation model incorporating these concepts could be designed, which could empirically determine under what circumstances deception can reduce an agent’s utility.
In the real world, case studies can be used to assess the explanatory power of the model. The case of the pharmaceutical Vioxx serves as one example. Vioxx was marketed by Merck to treat arthritis and other chronic or acute pain conditions. It was on the market for 5 years, being prescribed to over 80 million people worldwide, and generating up to $2.5 billion per year in revenue. In 2004 it was withdrawn from the market, after it was revealed that Merck had systematically misrepresented the safety data of the drug, including by outright removing evidence of adverse events from clinical trial data. FDA testimony before the Senate Finance Committee revealed that Vioxx had been responsible for up to 55,000 premature deaths from cardiovascular disease, and Merck subsequently paid $4.85 billion in a class action settlement (note that Vioxx is still net profitable even with the settlement).
A ‘no deception’ signaling model model with rational Bayesian utility maximizers might attempt to explain this by saying that patients taking Vioxx were signaling to their doctors that they were good and obedient patients, as this was maximizing their utility. Under my model this would be explained as deception, followed by a very delayed disconfirmatory signal with huge utility cost like death or a heart attack. In the event that the patient is able to identify Vioxx as the cause of the cardiovascular event, they face the question of how far to backpropagate the signal up their world model. Case 0 (no update): the drug is still worth taking, make no changes; case 1: Vioxx is a bad drug, stop taking it; case 2: Vioxx and all drugs like it are dangerous; case 3: all pharmaceuticals are dangerous, never take any ever again. It is not possible to actually calculate the expected utility returns of any of these updates, and so people will update based on their idiosyncratic psychology.
There is actual evidence of this kind of backpropagation of error in the case of the COVID vaccines. If a person is injured by a COVID vaccine, either mildly or severely, they can make the following updates. Case 0 (no update): the illness I feel is a sign that the vaccines are working, I will keep taking them; case 1: I feel ill, I will no longer take the vaccines (Bronski’s case); case 2: I will never take any vaccines ever again. There are numerous self-reported examples of each of these cases occurring, and they are also sometimes accompanied by evidence of an update to the epistemic vigilance function with respect to Pfizer or pharmaceutical companies in general.
Evidently there is already plenty of naturalistic data which supports this model (fraud involves deception and people suffer losses from fraud all the time), but we can propose experiments to test this. In keeping with the previous two examples, have a pharmaceutical company create two new drugs to treat condition X, one of which works, and one of which simply harms the patient for zero benefit. Market both drugs identically as useful breakthroughs in the treatment of condition X. Have doctors randomly prescribe one drug or the other to patients who present with that condition. If patients reject the harmful drug (before suffering its effects) but accept the helpful drug, we can consider deception falsified. Of course this study would never make it past an ethics review board, and yet I feel like this basic premise has been tried before under less formalized conditions…
More seriously, we could propose a study investigating how people react to falsified data on risk. Have subjects attend a psychological study on the third floor of a building that can be accessed either through an elevator or a staircase. Have the subjects read either a control article or an article that exaggerates the risks of taking an elevator, then observe whether people who previously used the elevator to attend the study choose to take the stairs on the way down. Proper experimental design should be used to ensure that subjects are not aware that the goal of the study is to assess elevator usage, to prevent demand characteristics and experimenter effects. Experimental conditions may include having the article be ostensibly written by an elevator expert or by a layperson, to estimate the effects of epistemic vigilance. Any observed effect would suggest that deceptive information can affect the rate at which people exploit opportunities in their environment. I predict that there would be a very small but real effect, especially among subjects high in neuroticism. If this result was found then it would suggest that the rate at which people suffer utility costs due to deception is moderated by psychological variables. As far as I know, experiments assessing the utility costs of deception in a controlled environment are yet to be performed.
Footnotes:
1 While I agree that Woke beliefs are in large part signaling, I do not agree they are entirely signaling. Transsexuals who later decide to detransition and report feeling tricked or manipulated into their transition are a good example. Surgically mutilating your genitals is just about the most costly signal imaginable, and it would be absurd to suggest that people do it to maximize their utility. When they decide to detransition they are mostly ignored, like many other victims of the medical system, and would hence lose any utility they gained from their signaling. If it is to be argued that such people are simply mentally ill, due to e.g. high mutational load (Dutton & Woodley’s spiteful mutant hypothesis) then it would still suggest that certain genotypes or psychological phenotypes are susceptible to losing utility through deception.
2 I wrote this and then heard Bronski say this (punctuation not included in transcript):
“Imagine we get the data we’re expecting and then like in the final product it’ll just be like yeah people maximize this utility function and that’s why ideas aren’t real and deception isn’t real and there’s no mind viruses and power is money right and it’s like we just present the data and present the equations and it’s like the physics of power basically that would be very sexy that’s what we’re trying to build”
(emphasis mine, source).
Overdue Revolutions 