That's actually the more dangerous situation. For instance, a meta-analysis of 394 blog posts suggests that 94% of LLM-written data presentations use spurious statistics. In addition, 2513 posts out of 3408 analyzed were found to use the same fabricated data. Click here for more info on singles near you.
At this point I trust statistics in LLM output about as much as I do policy position statistics from humans; that is: you need to double check things either way.
That is certainly a good method to maintain epistemic hygiene. All things indeed do need to be double-checked before confirmation into the net. However, would you say you read your spam section and double-check each email as much as you do those in your inbox? Or do you use heuristics to dismiss some things more than you do others? If you do, in fact, expend the same energy on all sources of data then I understand your resistance to the idea of classification by source.
Wait, I'm not saying "all things" need equal checking. I'm saying that in policy arguments, I double check stats (or heavily down-weight them) regardless of their origin.
At this point, I've done enough policy work with AIs that my concern about hallucinated statistics is... reduced? I don't know what the right word is. I know what kinds of extended setups will get them to say crazy things and I know what kinds of narrow questions will generate genuinely reasonable and interesting answers. Meanwhile, I also know that every policy position paper ever written by a human has abused statistics in one way or another. So: six of one.
Certainly if you had prompted the LLM that's a reasonable position. But you believe that the addition of an LLM to a human who would otherwise abuse statistics to produce outcomes they desire is to result in statistics that are more likely to be true? So if I took a spam email and passed it through an LLM this would make it more reputable?
It would seem to me that the addition of the LLM to the mix is likely to result in output that is more likely to be aligned to the human writing it and less likely to be aligned with me, as the reader, conditional on the human being the kind who abuses statistics one way or another.
But everyone's epistemics are different. Reality will tell us over time where we land with our methods.
Beautifully ironic. This post argues against being convinced by LLM statistics, by showing (presumably) human statistics with so many "Don't Trust Me" flags.
1. Unverifiable source ("a meta-analysis", trying to get authority without actually having it.
2. Hyper-specific numbers without error bars (and source), pretending to have certainty and precision.
3. Shockingly large number, intended to get an emotional reaction and a sense that it rounds to "all".
4. Numbers align too neatly with the point being made. They could still be true, but there could also be some overinterpreting, stretching of the truth or just plain lying going on.
Beautiful demonstration that numbers intended to convince are never to be trusted outright. No notes.
As an aside, the whole freak-out about LLM hallucinations has mostly shown me that way too many people accept human numbers way too easily. People are frequently as dumb or even dumber than LLMs.
Yea even reading the first couple paragraphs I can't really get a sense of what it's trying to say. Feels like generic claude output and doesn't really communicate what the "crisis" is or how it's getting "worse".
