Buckner (2024) defines anthropomorphism as the tendency to see humanlike characteristics in nonhuman entities on the basis of insufficient evidence, and reminds us how strong it is: in the Heider and Simmel animation (1944), a few geometric shapes moving around a rectangle are enough for viewers to ascribe to them an emotionally rich story (2024, p. 86). Comparative psychology has struggled with this bias since Darwin and Lloyd Morgan; its classic tool is Morgan's Canon, which forbids interpreting behavior in terms of higher psychological processes if it can be fairly interpreted in terms of lower ones. Buckner suggests a Morgan's Canon for robots, with "lower" hypotheses suited to machines, such as lookup tables that pair inputs with fixed responses (2024, p. 86): Block (1981)'s Blockhead.
Shanahan (2024) states the warning for LLMs: it is a serious mistake to apply to them unreflectively the intuitions we use with each other, because they work in profoundly different ways, and this can lead us to expect human-level performance where there is none (2024, p. 68). His practical advice is to keep in view the objective they were trained on, next-token prediction, and to be careful with verbs like "believes" or "knows." The ELIZA effect and the stochastic parrot label grow out of the same concern.
But the bias has a flip side. Buckner calls anthropocentrism the assumption that only behavior with the superficial trappings of human performance is valuable or intelligent, and anthropofabulation its combination with an inflated picture of our own cognition: we idealize human rationality, introspection and common sense, and then measure the machine against that ideal (2024, pp. 29, 87). Millière & Buckner (2024a) add the "Redescription Fallacy": denying a capacity merely because a system's operations admit a deflationary description ("it's just linear algebra"), as if a piano could not produce harmony because it is hammers striking strings (2024a, sec. 3).
The task, then, is not to choose between credulity and dismissal but to calibrate: what evidence—behavioral, but also mechanistic—would justify attributing which state, and who bears the burden of proof.