Reddit Debates Whether LLMs Think

The useful takeaway is not whether these systems are conscious, but whether they reliably produce reasoned outputs under testable constraints. For builders, that makes evals, grounding, and verification loops far more important than philosophy.

• –“Lies” are usually hallucinations or confabulations, not intent; the model is optimizing for plausible next tokens, not truth
• –Safety rules and system prompts can shape tone and refusal behavior, but they do not turn generation into human-style thought
• –Reasoning models can appear to self-correct, yet they still break on adversarial logic, long-context drift, and weak grounding
• –The practical metric is task success: code review, search, tool use, and calibration matter more than whether the model “thinks”
• –“Engineered intelligence” is a fair description of the current stack, but it’s still a behavior claim, not evidence of free-thinking