Elephants Don't Play Chess

The target of the argument

By 1990 the dominant strand of artificial intelligence had been moving in one direction for roughly thirty years, and Brooks argued that a different route was available ^[1]. The tradition he set himself against put the abstract manipulation of symbols at its centre. The difficulty, as the paper states it, was that this symbol manipulation was rarely grounded in physical reality ^[1]. A system could reason about the world in formal notation while having almost no working connection to a world.

The AIWS History of AI project describes the paper as proposing a ground-up approach to developing artificial intelligence, explicitly in contrast with classical AI ^[2].

The alternative: interaction first

The proposal is methodological rather than a new formalism. The paper explores a research methodology in which ongoing physical interaction with the environment is treated as the primary constraint on designing intelligent systems ^[1]. Instead of deciding in advance which internal representations a mind must contain and then building them, the designer starts from a body in a world and lets the demands of acting there shape the architecture.

The vocabulary of the paper points the same way. Its listed keywords are situated activity, mobile robots, planning, subsumption architecture and artificial intelligence ^[1]. Situated activity means treating intelligence as something studied in the context of a body acting in an environment rather than in isolation from it. Subsumption architecture is a way of composing robot control out of simple behaviours arranged in layers, where a higher layer can take over from a lower one when it needs to, instead of a single central model of the world directing everything.

The evidence offered

Brooks did not present the approach as a promise about the future. The abstract states that the methodology had recently produced significant successes on a par with the most successful classical efforts ^[1]. The paper also outlines plausible future work that could lead to vastly more ambitious systems ^[1].

Who wrote it

According to ScienceDirect's author note, Brooks was an associate professor at MIT and a member of the Artificial Intelligence Laboratory, where he led the mobile robot group ^[1].

Why it is remembered

Two things make the paper a reference point. First, it framed the disagreement as one about grounding: symbols need something to be about, and physical interaction is offered as the source of that something ^[1]. Second, it tied that point to machines that ran. The claim that the alternative approach had matched classical results was a claim about robots in a laboratory, not about a thought experiment ^[1]. The AIWS History of AI project records the paper among the events of the field's history, dated to 1990 ^[2].