We need a better word for when an AI agent “acts.” To act implies intent, volition, and, well, agency. What’s a more accurate way to describe when a human instructs an AI agent to perform a function and it does that?
0 Boosts
0 Likes
How we talk about tech matters. Talking about AI as acting on its own lets companies swindle investors and escape responsibility for AI harms. The “AI agent” framing is a marketing tactic and a liability dodge.
@tiffanycli If I were looking for non-anthropomorphizing language about LLMs, the first place I'd look for inspiration is the article referenced in the post below.
The second place I'd look, in case a bit more of a humorous framing were in order, is a recent Cal Newport talk, based on which I've explained to non-technical clients: If you train a dog to trim grass with a weed whacker, and strap that weed whacker to the dog and start it up, and the dog runs off to chase a squirrel and injures someone in the process, the responsibility rests not with the dog, but with you who strapped the weed whacker to the dog.
If you're discussing legal liability or anything else academic, technical, professional, or otherwise serious, I'd go with such language as Prof. Bender and Dr. Inie recommend.
https://dair-community.social/@emilymbender/116811128152569506
@tiffanycli If I were looking for non-anthropomorphizing language about LLMs, the first place I'd look for inspiration is the article referenced in the post below.
The second place I'd look, in case a bit more of a humorous framing were in order, is a recent Cal Newport talk, based on which I've explained to non-technical clients: If you train a dog to trim grass with a weed whacker, and strap that weed whacker to the dog and start it up, and the dog runs off to chase a squirrel and injures someone in the process, the responsibility rests not with the dog, but with you who strapped the weed whacker to the dog.
If you're discussing legal liability or anything else academic, technical, professional, or otherwise serious, I'd go with such language as Prof. Bender and Dr. Inie recommend.
https://dair-community.social/@emilymbender/116811128152569506
I understand that the point of delineating agentic AI from more simplistic LLMs is that AI agents can perform complex functions, including “deciding” how to proceed when given a broad prompt that would require multiple subprocesses to complete. But there’s still a human giving instruction.