The Witness Who Types Herself

Process reliabilism says a belief is justified if the process that produced it is reliable. Conee and Feldman (1998) pointed out the catch, now called the generality problem. Every belief-forming event belongs to countless process types: “visual perception”, “seeing a maple through a window”, “seeing in daylight at 30 metres”. Each type has its own reliability, and the theory gives no principled rule for picking one. A recent essay breaks the problem into five sub-problems, which is good bookkeeping but not a solution. The strongest reply I know is Jack Lyons’s “Algorithm and Parameters” (Phil. Review, 2019). On his view, the relevant type is the token’s full algorithmic description plus “the values of all the causally relevant parameters,” meaning the psychological variables that systematically affect processing. That hands the choice of type to cognitive science instead of intuition. The standard objection is that this is a promissory note, because nobody can actually list the parameters for a real belief.

In one domain the note has been paid. Eyewitness research has spent close to fifty years doing what Lyons’s scheme requires. Wells (1978) split the causally relevant variables into system variables, which police control (lineup fairness, blind administration, first test of the memory, when confidence is recorded), and estimator variables, which they don’t (distance, lighting, exposure time, weapon focus, cross-race effects). The “pristine” protocol in Wixted and Wells’s 2017 synthesis fixes the algorithm and pins the system parameters at known values. Under those conditions, high-confidence suspect identifications in the lab routinely exceed 95% correct. Break a system parameter, for example with an unfair lineup, and the same “I’m sure” is seriously compromised. That is Lyons’s typing run as a research program, with a reliability number in each cell. The field data agree. A signal-detection analysis of 348 real photo lineups from the Houston Police Department covered robberies, mostly with a weapon and mostly cross-race, which are bad viewing conditions by any textbook. It still put high-confidence suspect IDs at roughly 97% accurate.

The estimator variables are the hard part, since nobody at a crime scene measured the distance or the light. Here the data go further than Lyons claimed. In a five-distance lab study (Ayala et al. 2025, summarized in Wixted and Mickes 2025), the share of witnesses making a high-confidence ID fell from 33% at 10 feet to under 3% at 210 feet. Accuracy among those high-confidence IDs barely moved: 98% for good views against 96% for poor ones (z = 1.08, p = 0.27). The proposed mechanism is a constant-likelihood-ratio criterion. As memory strength falls, the threshold for saying “I’m sure” moves outward, so the statement keeps meaning roughly the same odds of guilt. The confidence report then works as a readout of the unmeasured parameter values, because a witness who saw badly mostly can’t produce the high-confidence token. Within the class “initial, pristine, high-confidence ID”, the estimator variables are close to screened off. That is Salmon’s condition for an objectively homogeneous reference class, which James Beebe (2004) proposed as the reliabilist’s criterion for choosing a type. Two separate answers to the generality problem meet in one protocol. The problem becomes tractable for any process whose output includes a calibrated estimate of its own reliability, captured before anything contaminates it. For such processes, the token tells you its type. The courtroom shows the contaminated case. In Brandon Garrett’s study of DNA exonerations (discussed in Wixted et al. 2015), 57% of 161 misidentification cases involved witnesses who were confident at trial but had been substantially uncertain at the first identification. Trial confidence comes from a different process type, the same memory re-tested after feedback and rehearsal. The certainty factor in Manson v. Brathwaite (1977) let jurors treat it as if it were the first reading.

The extension has a live opponent. Lockamyeir et al. (2020), also reviewed in Wixted and Mickes 2025, tested 3,762 participants at three distances. At 20 m, with d′ = 0.29, high-confidence suspect IDs were only 63% accurate, which suggests confidence does not fully screen off viewing conditions. The reply is that the 63% rests on just 19 IDs, 1.6% of 1,194 witnesses, so a handful of careless participants could produce it. My bet is that the next field dataset with recorded initial confidence will repeat the Ayala pattern. High-confidence accuracy will differ by under five points between good-view and poor-view cases, while the rate of high-confidence IDs differs by a factor of three or more. If poor-view high-confidence accuracy instead drops below 85% on a sample of 100 or more such IDs, the extension fails. The generality problem would then reopen even here, in the one place where it looked solved.