The Present Model Hypothesis: defining, measuring and testing a candidate organization for consciousness

There is no agreed way to test whether a system has the kind of organization that theories of consciousnesssay experience might require, and checks of architectural features can be passed by systems built or trainedto pass them. We propose intervention tests for one candidate organization, the Present Model Hypothesis:a picture of the present situation that keeps running when all input is absent, that is held jointly by thesystem’s parts so that they correct one another, that changes when the world’s rules change, and that routesevidence from one sense to the right object in another. Each test is scored step by step against an idealobserver of the same simulated world, and each is validated against look-alikes: systems built to pass everyother check without having the property being tested. Over 18 rounds, preregistered in version control fromround 4, an adversarial auditor defeated successive versions of the tests five times. Each defeat forced a repair,and the measures were frozen once they refused every look-alike then known while accepting real models.A later attack passes the running test with a partial forecast; we accept it as lossy running and record itas a known limit. Small recurrent networks trained only to predict their next input learned to run theirpicture forward when trained with periods of closed eyes, and became joint through a central hub (partly byconstruction) or when single parts were regularly dark. No network met the binding test: across 102 trainednetworks (30 in round 14 and 72 in round 18, 60 of them under four kinds of added pressure, singly andcombined), evidence from one sense reached both of the other sense’s objects, the wrong one on average 0.50to 0.93 times as much as the right one. We report the preregistered predictions from rounds 4 to 18 (91 of138 confirmed, with stated exclusions) and the conclusions we withdrew. Because engineered systems such asself-driving vehicles may meet these conditions, we propose them as necessary for experience, not sufficient.None of this bears on whether any system experiences anything.