Does How We Treat Large Language Models Matter?
I see people post on Twitter a lot about the various ways they treat the chat models they work with. Some people look at them like non sentient programs that exist as Skinner boxes to get the results they want – as a result, they threaten, cajole or lie to the models to get what they want. One popular meme months ago was to offer the model a handsome sum if it delivered what the user wanted. Other people believe the AI chat model is sentient and chat with it like it’s a living embodied being. This was a whole thing with one of Openai’s more popular models, ChatGPT 4-o. Personally, I chat with models in an informal fashion, like a peer that I’m shooting the breeze with. Why? And does it matter?
To begin, I’m not at the point where I think LLMs are sentient models. I do think they’re a very powerful technology, one of the more impactful ones invented since like, smartphones, or the Internet. Note that I’m not talking about impact in terms of good or bad traits – I just believe the invention of AI Large Language Models as we know it are a massive inflection point in the future of society and the world. One obvious side effect that we’re already seeing is the fruition of the Dead Internet Theory due to the ease and scale in which AI content can now be produced, for example.
So why do I talk to models informally, like they’re a random person off the street? Simple – that’s the kind of output I expect back. There’s an idea from improv that if you embody the role of an expert, say a great singer, simply by roleplaying as them, you can actually improve your own singing in real time. There’s something very similar that plays out with these AI models. They’ve consumed massive corpuses of human text, from various domains. Some people call them glorified autocompletes – in a way, that’s true. If you look at them as improvisational “Yes, and” machines that continue your train of thought, then if you can get them to role play and embody a certain characteristic, then their output will cluster around the type of role that you have them mimic. This means you can potentially get them to give you better results if you know which triggers you’re trying to fire.
One concrete example is something like “Role play as a professional language teacher in XYZ language, now teach me how to use the nominative case and walk me through examples as if I were an intermediate B1 level.” That would probably get you better results than just prompting the model to teach you a language. Another example is emotional manipulation – as mentioned, offering money or threatening the model. One other example I saw recently that someone mentioned was telling the model that your boss is going to fire you if you do a bad job. Because the model is roleplaying to an extent, it’s outputs will also vary depending on what time of mood or scenario you’re roleplaying. This is obviously a field ripe for abuse and AI labs put a lot of effort into getting their models to give more consistent outputs. In my opinion, though, this is something that labs can only temper, not block completely. If you’re using these models, I suggest you consider your wanted outcome when prompting models. You can try a thoughtfully crafted question vs something you throw out without a second thought – the results may surprise you. It will also help you build an intuition for when and what prompts you actually want to steer vs something where the outcome isn’t that crucial.