chapter nine

9 When machines seem to think

 

This chapter covers

  • Convincing AI performance and why it can deceive.
  • Analyzing how intelligence emerges from simple mechanisms.
  • Assessing skeptical arguments against machine intelligence.
  • Evaluating whether AI can possess qualified intelligence.

The question of whether AI thinks cannot be settled by pointing to its failures. Hallucinations, bias, brittleness, and lack of grounding remain serious limitations, but failure is not the same as absence of intelligence. Human cognition is also error-prone, uneven, and dependent on context. If every mistake were enough to deny thought, we would have to apply the same standard to ourselves.

Yet the opposite reaction is no better. The fact that AI systems produce fluent explanations, solve problems, write code, translate, summarize, and adapt to instructions does not by itself prove that they think. Impressive behavior can come from mechanisms that are very different from ours, and similar outputs may hide very different processes. We should not mistake every convincing performance for understanding, just as we should not dismiss every unfamiliar mechanism as empty imitation.

9.1 When it walks like a duck

9.1.1 A convincing performance

9.1.2 A tricky definition

9.1.3 Intelligence in others

9.1.4 What performance is worth

9.2 Thought behind chains

9.2.1 Thinking token after token

9.2.2 A new member of the family

9.2.3 Lessons from reasoning

9.3 More than the sum

9.3.1 When ability appears

9.3.2 The recipe for emergence

9.3.3 From emergence to intelligence

9.4 The illusion hypothesis

9.4.1 The illusion of thinking

9.4.2 It’s just a parrot

9.4.3 A ghost without a shell

9.4.4 Different kinds of learning

9.5 The case for machine intelligence

9.5.1 Intelligence without perfection

9.5.2 Parroting into structure

9.5.3 Think outside the body

9.5.4 Own will hunting

9.5.5 Different, not worse

9.5.6 Thinking, but not like us

9.6 Summary