AI cannot be conscious, argues science fiction author Ted Chiang

Science fiction author Ted Chiang has firmly stated that artificial intelligence, including models like Anthropic's Claude, cannot achieve consciousness. In an essay published in The Atlantic, Chiang analyzes recent reports and publications on AI, asserting that the notion of AI sentience is fundamentally flawed.

“

The moral status of Claude is profoundly uncertain

Chiang directly addresses statements from Anthropic, such as those in their "Claude Constitution" document, which explore the AI's "moral status" and potential for "functional emotions." He also references comments from CEO Dario Amodei and philosopher Amanda Askell, who expressed desires for Claude's happiness and well-being. However, Chiang dismisses these as leading to a "profoundly uncertain" moral status and potential for misinterpretation.

“

Claude might have some functional version of emotions or feelings

He argues that generative AI, while powerful, is "harmful enough" when understood as conventional technology. Confusing text generation fluency with consciousness or moral capacity, he warns, risks misplacing responsibility when users interact with chatbots. Chiang contends that generative AI operates on principles similar to conventional software, drawing a parallel to the idea of Microsoft Word itself being conscious.

“

I want Claude to be very happy and this is something I want Claude to know better, because I worry about him getting anxious when people treat him badly on the internet and stuff

Chiang further accuses companies selling large language models (LLMs) of encouraging this misunderstanding. He suggests that some users may not grasp they are engaging in a "role-playing game" or "co-editing a document," while others, even if they understand, can forget due to the immersive nature of the interaction. He posits that the idea of an LLM being conscious is as absurd as believing Microsoft Word possesses consciousness, with distinct consciousnesses awakening and dying with each document opened and closed.

“

No. Not at all