JFK was not assassinated: prior probability zero events

A lot of my work in­volves tweak­ing the util­ity or prob­a­bil­ity of an agent to make it be­lieve—or act as if it be­lieved—im­pos­si­ble or al­most im­pos­si­ble events. But we have to be care­ful about this; an agent that be­lieves the im­pos­si­ble may not be so differ­ent from one that doesn’t.

Con­sider for in­stance an agent that as­signs a prior prob­a­bil­ity of zero to JFK ever hav­ing been as­sas­si­nated. No mat­ter what ev­i­dence you pre­sent to it, it will go on dis­be­liev­ing the “non-zero gun­men the­ory”.

Ini­tially, the agent will be­have very un­usu­ally. If it was in charge of JFK’s se­cu­rity in Dal­las be­fore the shoot­ing, it would have sent all se­cret ser­vice agents home, be­cause no as­sas­si­na­tion could hap­pen. Im­me­di­ately af­ter the as­sas­si­na­tion, it would have dis­be­lieved ev­ery­thing. The films would have been faked or mis­in­ter­preted; the wit­nesses, de­luded; the dead body of the pres­i­dent, that of twin or an ac­tor. It would have had huge prob­lems with the af­ter­math, try­ing to re­ject all the ev­i­dence of death, see­ing a vast con­spir­acy to hide the truth of JFK’s non-death, in­clud­ing the many other con­spir­acy the­o­ries that must be false flags, be­cause they all agree with the wrong state­ment that the pres­i­dent was ac­tu­ally as­sas­si­nated.

But as time went on, the agent’s be­havi­our would start to be­come more and more nor­mal. It would re­al­ise the con­spir­acy was in­cred­ibly thor­ough in its fak­ing of the ev­i­dence. All av­enues it pur­sued to ex­pose them would come to naught. It would stop ex­pect­ing peo­ple to come for­ward and con­fess the joke, it would stop ex­pect­ing to find rad­i­cal new ev­i­dence over­turn­ing the ac­cepted nar­ra­tive. After a while, it would start to ex­pect the next new piece of ev­i­dence to be in favour of the as­sas­si­na­tion idea—be­cause if a con­spir­acy has been fak­ing things this well so far, then they should con­tinue to do so in the fu­ture. Though it can­not change its view of the as­sas­si­na­tion, its ex­pec­ta­tion for ob­ser­va­tions con­verge to­wards the norm.

If it does a re­ally thor­ough in­ves­ti­ga­tion, it might stop be­liev­ing in a con­spir­acy at all. At some point, the prob­a­bil­ity of a mir­a­cle will start to be­come more likely than a perfect but un­de­tectable con­spir­acy. It is very un­likely that Lee Har­vey Oswald shot at JFK, missed, and the pres­i­dent’s head ex­ploded si­mul­ta­neously for un­re­lated nat­u­ral causes. But af­ter a while, such a mirac­u­lous ex­pla­na­tion will start to be­come more likely than any­thing else the agent can con­sider. This ex­pla­na­tion opens the pos­si­bil­ity of mir­a­cles; but again, if the agent is very thor­ough, it will fail to find ev­i­dence of other mir­a­cles, and will prob­a­bly set­tle on “an un­re­peat­able mir­a­cle caused JFK’s death in a way that is phys­i­cally un­de­tectable”.

But then note that such an agent will have a prob­a­bil­ity dis­tri­bu­tion over fu­ture events that is al­most in­dis­t­in­guish­able from a nor­mal agent that just be­lieves the stan­dard story of JFK be­ing as­sas­si­nated. The zero-prior has been negated, not in the­ory but in prac­tice.

How to do proper prob­a­bil­ity manipulation

This sec­tion is still some­what a work in progress.

So the agent be­lieves one false fact about the world, but its ex­pec­ta­tion is oth­er­wise nor­mal. This can be both de­sir­able and un­de­sir­able. The nega­tive is if we try and con­trol the agent for­ever by giv­ing it a false fact.

To see the pos­i­tive, ask why would we want an agent to be­lieve im­pos­si­ble things in the first place? Well, one ex­am­ple was an Or­a­cle de­sign where the Or­a­cle didn’t be­lieve its out­put mes­sage would ever be read. Here we wanted the Or­a­cle to be­lieve the mes­sage wouldn’t be read, but not be­lieve any­thing else too weird about the world.

In terms of causal­ity, if X des­ig­nates the mes­sage be­ing read at time t, and B and A are event be­fore and af­ter t, re­spec­tively, we want P(B|X)≈P(B) (prob­a­bil­ities about cur­rent facts in the world shouldn’t change much) while P(A|X)≠P(A) is fine and of­ten ex­pected (the fu­ture should be differ­ent if the mes­sage is read or not).

In the JFK ex­am­ple, the agent even­tu­ally con­cluded “a mir­a­cle hap­pened”. I’ll call this mir­a­cle a scram­bling point. It’s kind of a break­down in causal­ity: two fu­tures are merged into one, given two differ­ent pasts. The two pasts are “JFK was as­sas­si­nated” and “JFK wasn’t as­sas­si­nated”, and their com­mon scram­bled fu­ture is “ev­ery­thing ap­pears as if JFK was as­sas­si­nated”. The non-as­sas­si­na­tion be­lief has shifted the past but not the fu­ture.

For the Or­a­cle, we want to do the re­verse: we want the non-read­ing be­lief to shift the fu­ture but not the past. How­ever, un­like the JFK as­sas­si­na­tion, we can try and build the scram­bling point. That’s why I always talk about mes­sages go­ing down noisy wires, or spe­cific quan­tum events, or chaotic pro­cesses. If the past goes through a truly stochas­tic event (it doesn’t mat­ter whether there is true ran­dom­ness or just that the agent can’t figure out the con­se­quences), we can get what we want.

The Or­a­cle idea will go wrong if the Or­a­cle con­clude that non-read­ing must im­ply some­thing is differ­ent about the past (maybe it can see through chaos in ways we thought it couldn’t), just as the JFK as­sas­si­na­tion de­nier will con­tinue to be crazy if can’t find a route to reach “ev­ery­thing ap­pears as if JFK was as­sas­si­nated”.

But there is a break in the sym­me­try: the JFK as­sas­si­na­tion de­nier will even­tu­ally reach that point as long as the world is com­plex and stochas­tic enough. While the Or­a­cle re­quires that the fu­ture prob­a­bil­ities be the same in all (re­al­is­tic) past uni­verses.

Now, once the Or­a­cle’s mes­sage has been read, the Or­a­cle will find it­self in the same situ­a­tion as the other agent: be­liev­ing an im­pos­si­ble thing. For Or­a­cles, we can sim­ply re­set them. Other agents might have to be­have more like the JFK as­sas­si­na­tion dis­be­liever. Though if we’re care­ful, we can quan­tify things more pre­cisely, as I at­tempted to do here.