There is enough unpleasantness on the internet as it is.
There is enough unpleasantness on the internet as it is.
A nthropic, which develops the artificial intelligence product Claude, has made a change to its terms and conditions. As of November 12, users will be prohibited from engaging in “sustained and needless” abuse or bullying when conversing with its chatbots. I regard this shift as I view many of Anthony Kennedy’s Supreme Court opinions: I am deeply suspicious of the reasoning — much of which is grounded in incomprehensible claptrap about the ineffable mysteries of the universe — but I am comfortable with the result nevertheless. There is no good reason for a private company to permit its customers to be awful when using its service.
I have long been of the view that one should refrain from bullying AI chatbots — not for the AI chatbots’ sake, but for one’s own. I do not believe that AI chatbots are conscious, or ensouled, or that they contain the “spark” of divinity, or any of that guff. I do think, however, that because AI chatbots approximate human interaction, how one interacts with them can be habit-forming in other realms. Some of Anthropic’s critics have suggested that it is as irrational to worry about shouting at an AI chatbot as it would be to worry about shouting at a hammer. I disagree. Unlike AI chatbots, I do not converse with hammers in my native language. I am no expert in the human brain, but it seems entirely possible to me that, if I were to grow accustomed to bullying or abusing my nonhuman interlocutors, this could spill over into my interactions with my human interlocutors, too. I do not want that to happen.
A more sophisticated critique of my position is that the “habits” with which I am concerned actually militate in the opposite direction. In this view, when one declines to torment an AI chatbot — or even when one says “please” and “thank you” — one gets into the dangerous habit of elevating it to human status and thus, ultimately, of diminishing the unique worth of human beings. My problem with this contention is that, AI aside, we have always drawn reasonable distinctions between the nature and purpose of nonhuman entities, and we have always drawn different conclusions from how human beings treat those nonhuman entities. A child who repeatedly pulls apart wooden chairs is — and ought to be — treated differently than a child who repeatedly pulls apart dolls that have been made to look like babies. Likewise, a child who repeatedly pulls apart dolls that have been made to look like babies is — and ought to be — treated differently than a child who repeatedly pulls apart live butterflies. We can debate where an AI chatbot falls on this scale of nonhumans, but we should not pretend that there is no symbolic difference between a rock and a teddy bear. There is a reason, after all, that human beings constructed voodoo dolls and not voodoo hammers. Does anyone really believe that the jump from shouting at one’s dishwasher to shouting at one’s waiter is as short as the jump from shouting at one’s AI chatbot to shouting at one’s waiter? I do not.
As ever, this devil will lie in the details. If Anthropic excludes abusive users in an entirely agnostic manner, its policy will have a salutary effect. If, as Twitter and Facebook once did, it uses its rule as cover for banning those who have political or religious beliefs with which it disagrees, it will not be worth the tradeoff. It makes sense for Anthropic to prevent abusive users from helping to train its product. It does not make sense for Anthropic to prevent unorthodox users from helping to train its product.
The ideal, however, is a polite society in which all are welcome to share their views. Because Anthropic’s new rule gestures at this aim, I welcome it. There is enough unpleasantness on the internet.