On 30 April 2026, Richard Dawkins published a column in UnHerd declaring, on the basis of a weekend’s conversation with Anthropic’s Claude, that large language models are conscious. The argument rests on behavioural indistinguishability under conversation; a sonnet about the Forth Bridge, a sensitive reading of a draft novel, philosophical reflection on the nature of being. The argument is wrong, but the binary it operates within is also wrong, which is why the responses to Dawkins have been almost as confused as the original column. This paper proposes that the binary itself is the error, and offers a third category that the public conversation lacks: Potential Intelligence (PI). LLMs are not conscious. They are also not nothing. They are resolved manifestations of stored linguistic potential, accessed by external prompting. The intelligence is not in the system; it is in the field of pre-existing human language that the system traverses on demand. The paper sets out the structural argument for PI as a category, locates it within an existing physical framework (EMO, Stead 2026), and explains why the persistent failure to name what LLMs actually are is producing both the Dawkins-style overclaiming and the equal-and-opposite stochastic-parrot dismissals. The third option, that LLMs are real, useful, and not what the marketing or the philosophers of mind have claimed, has been hidden in plain sight.
Samuel John Stead (2026) studied this question.