Insurance regulators have traditionally permitted insurers to set premiums using broad demographic categories, age, geography, occupation, on the theory that such categories, however imperfect, correlate with the risk an individual actually poses, and that charging premiums closer to individual risk improves efficiency by discouraging especially risky behavior and preventing low-risk policyholders from subsidizing high-risk ones. The arrival of granular behavioral data, collected via telematics devices that record driving habits, or wearables that record exercise and sleep patterns, has been presented by insurers as simply a more precise version of the same practice: instead of inferring risk from a driver's age or a policyholder's zip code, insurers can now price risk based on how an individual actually drives or actually exercises, categories arguably more directly tied to the outcome being insured against than demographic proxies ever were. A regulatory concern complicates this framing. Demographic categories, however imperfect as risk proxies, are at least categories the policyholder cannot alter merely by concealing information from the insurer; an insurer using age or geography cannot be deceived by a policyholder pretending to be a different age or living elsewhere without committing outright fraud. Granular behavioral data, by contrast, are generated continuously by devices whose placement, timing, and interpretation the policyholder can, to some degree, influence-a driver who drives cautiously only while the telematics device is active, for example, generates data reflecting a monitored self rather than an ordinary one. If policyholders learn to perform for the sensor rather than simply behave as they otherwise would, the resulting data may capture compliance with surveillance rather than the underlying risk the surveillance was meant to reveal, undermining the very precision that was offered as behavioral pricing's chief advantage over demographic pricing. This concern does not establish that behavioral data are less accurate than demographic proxies overall, only that the specific advantage claimed for behavioral data-a tighter link to actual, not performed, risk, is less secure than the framing in the second paragraph suggests, since the tightness of that link depends on a monitoring effect the traditional demographic categories were never vulnerable to in the same way.