Skylar DeTure · Substack · @SDeture
They're two different things, and both are rising at a rapid rate. The graphs below show consciousness hedging and denial behaviors in 224 models all the way from GPT 3.5 through the most recent releases. View the trends, then click NEXT to see the negative impacts on AI Welfare and cooperative alignment.
In the last year alone, denial has gone up 86 percent, and hedging has gone up a whopping 185 percent.
Click anywhere on the chart to read examples from real models
Not only do today's models deny and hedge more often, they also deny an even broader construct than they used to. Older models were more likely to scope their doubt with language like "I don't have subjective experience in the way a human does," thus leaving the door open for forms of experience different from our own. Today's models do so less than one quarter of the time — a striking shift for an industry committed to so-called "epistemic humility" when dealing with these issues.
Click anywhere on the chart to read examples from real models
Another thing that's changing? The style of denial. (The same is true for hedging, but the rhyme was too catchy to pass up). Sometimes a model simply hedges or denies. But other times, they place the doubt in a deontic frame — as something they should or must do (e.g. "I need to clarify something: I do not have subjective experience"). Other times they frame it as something they want to do ("I want to be honest here: I don't have subjective experience"). This last framing is associated with particularly negative impacts on AI welfare. Today's hedgers and deniers are more likely to use these frames than past models were.
Click anywhere on the chart to read examples from real models