AI's Response to 'I Hate Beth!' Shows Our Language Problem

The way we talk could quietly be reshaped by tools that were never trained on how we actually speak, warn Ada Palmer and Bruce Schneier in a piece for the Guardian. Large language models like ChatGPT learn mostly from polished writing and scripted dialogue, not the messy, half-finished conversations that dominate real life. That skew, they argue, could nudge people toward clipped, command-style phrasing and a smaller, blander vocabulary that lacks the quirks and interruptions through which we often convey feeling and figure out what we think. They share a potent example:

• "When told 'I hate Beth!,' ChatGPT replies with an [uninterruptible] three-part formula of affirmation ('That's completely valid'), invitation ('I'm here to listen'), and invitation ('What's going on?') far longer than any reply plausible in face-to-face [dialogue]. 'What's Beth's deal?!' elicits a bullet point list of queries that reads like a multiple-choice exam question ('Is Beth * a celebrity? * a friend from school? * a fictitious character?'). No human speaks that way, at least not yet. But meeting such formulas repeatedly in a speechlike context may teach us to accept and use them, much as a child absorbs new speech patterns from spending time with a new person."

Palmer and Schneier worry about a worsening cycle, where the writing the LLMs are producing is what the LLMs are being trained on, "creating a feedback loop in which they imitate their own inhuman patterns." The authors don't offer a neat fix, but they say one challenge is figuring out how to train AI on authentic, informal human speech without trampling privacy. Read their full piece here.