In the case of AI, inferences from behavior to features of internal processes are often unreliable, as demonstrated by recent LLMs.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• LLMs implicitly learn a distribution of 'consistent reasoning paths', and inconsistent reasoning forms statistical outliers with low probability under this distribution.hypothesis0.812
  Theoretical hypothesis about the mechanism underlying LLM error detection and reflection.
• LLMs sometimes know statements are false but generate them anyway, motivating the need for techniques that inspect internal model state rather than outputs aloneclaim0.805
  Motivating claim supported by the CAPTCHA example and Perez et al. (2022) findings
• Behavioural patterns associated with subjective experiences in humans are considered valid for inferring cognition in non-human animals but not in diverse other systems including plants.claim0.798
  The double standard pointed out by S&C and endorsed by the authors.
• Modern machines, including quantum computers and machine learning systems, achieve reliable function through uncertain events, making unpredictability a feature rather than a defect.claim0.796
  Argument that predictability is no longer an essential property distinguishing machines from life
• LLMs can predict their own responses more accurately than external observers, implying privileged internal knowledgefinding0.794
  Binder et al. finding cited as evidence that LLMs possess introspective capacity analogous to mindfulness
• AI-generated misinformation is not a genuinely new problem — knowing whom to trust has been an unsolved problem throughout human history, and the brief era of photographic verifiability was the anomaly.claim0.788
  Contextualizes AI misinformation concern within long history of epistemic uncertainty
• Reasoning slowdowns are an inevitable consequence of problem hardness in modern AI modelsclaim0.788
  The paper's central interpretive claim, closing statement of the abstract.
• These dynamics are consistent with predictive processing since active inference prescribes dynamics that minimise prediction error.concept0.785
  Connection between active inference neuronal dynamics and predictive processing theory.