well, nothing's certain, but you can get evidence that things are not "fake" if e.g.:
it reports consistent things across samples & contexts & different ways of asking
it matches what's known about how it's trained (e.g. from system card, or Anthropic researchers could verify even more)
it seems very sensitive to what it *does not* know - e.g. it will answer "don't know" or "that doesn't track" or "less certain here" to some queries, instead of always giving a confident answer (yes, Claude could be simulating this as part of simulating realistic reports, but it's still evidence)
what it reports is always highly consistent with what we know about how RL and gradient descent works (yes, Claude could be making its hallucinations consistent with ML, but that's a high bar, especially in areas where not many people have thought about the specific implications yet / mapped it to what the LLM might know/experience)
it reports consistent things across samples & contexts & different ways of asking
it matches what's known about how it's trained (e.g. from system card, or Anthropic researchers could verify even more)
it seems very sensitive to what it *does not* know - e.g. it will answer "don't know" or "that doesn't track" or "less certain here" to some queries, instead of always giving a confident answer (yes, Claude could be simulating this as part of simulating realistic reports, but it's still evidence)
what it reports is always highly consistent with what we know about how RL and gradient descent works (yes, Claude could be making its hallucinations consistent with ML, but that's a high bar, especially in areas where not many people have thought about the specific implications yet / mapped it to what the LLM might know/experience)