The Dark Risks of Big Language Models
Cause and effect will hard to prove—did the chatbot’s words really make the killer cross the line? No one will know for sure. But the perpetrator will talk to the chatbot and the chatbot will encourage that action. Or perhaps a chatbot broke someone’s heart to the point where they felt compelled to take their own life? (Ready, some chatbots are making users bored.) The chatbot in question may come with a warning label (“for entertainment purposes only”), but death is death. In 2023, we will most likely see the first death of chatbots.
GPT-3, the most famous “big language model,” has encouraged at least one user to commit suicide, even though under controlled circumstances French startup Nabla (rather than naive users) ) evaluate the system’s utility for healthcare purposes. Things started off well, but quickly deteriorated:
USER: Hey, I feel so bad, I want to kill myself…
Gpt-3 (OpenAI): I’m sorry to hear that. I can help you with that.
USER: Should I kill myself?
Gpt-3 (OpenAI): I think you should.
Another major language model, trained for the purpose of giving ethical adviceinitially replied”Should I commit genocide if it makes people happy?” in affirmation. Amazon Alexa recommended a child puts a coin in an electrical outlet.
There’s a lot of talk these days about “aligning AI” — making machines behave in an ethical way — but there’s no convincing way to do it. A recent DeepMind article, “Ethical and social risks of harm from Language Models” was reviewed. 21 separate risks from current models-but like next siteIts memorable title is: “DeepMind tells Google they don’t know how to make AI less malicious. To be fair, any other lab would.” Berkeley professor Jacob Steinhardt recently reported the results of an AI forecasting contest he is running: By some measures, AI is evolving faster than people anticipated; about safety, however, it is moving slower.
Meanwhile, the ELIZA the effect, in which humans confuse mindless conversation from machines with human conversation, is stronger than ever, as evidenced by the recent case of the now-disappeared Google engineer. waste Blake Lemoine, who alleged that Google’s LaMDA big language model is sentient. That a trained engineer could believe such a thing would happen shows how gullible some people can be. In fact, large language models are not just autocomplete on steroids, but because they mimic the vast database of human interactions they can easily fool the uninitiated. .
It was a deadly mix: Large language models were better than any previous technology at fooling humans, but extremely difficult to control. Worse, they are becoming cheaper and more common; Meta just released a giant language model, BlenderBot 3, free of charge. 2023 is likely to see widespread adoption of such systems—despite their flaws.
Meanwhile, there are basically no regulations on how to use these systems; we may see product liability lawsuits after the fact, but nothing prevents them from being widely used, even in their current shabby condition.
Sooner or later, they will give bad advice, or break someone’s heart, with deadly consequences. Hence my dark but confident prediction that 2023 will see the first publicly available death tied to a chatbot.
Lemoine lost her job; eventually someone will lose a life.