Why treating AI like a person is the future
In my new book (out in a couple days, preorder for access to neat things!) I commit a sin. And not just once, but many, many times. I anthropomorphize AI.
Partway through the book, I stop writing that an “AI ‘thinks’ something” and instead just write that “AI thinks something.” The missing quotation marks may seem like a subtle distinction, but it is an important one. Many experts are very nervous about anthropomorphizing AI. And they have good reasons.
Anthropomorphism is the act of ascribing human characteristics to something that is nonhuman. We’re prone to this: we see faces in the clouds, give motivations to the weather, and hold conversations with our pets. It’s no surprise, then, that we’re tempted to anthropomorphize artificial intelligence, especially since talking to LLMs feels so much like talking to a person. Even the developers and researchers who design these systems can fall into the trap of using humanlike terms to describe their creations, starting with labels like “machine learning.”
This may seem like a silly thing to worry about. After all, it is just a harmless quirk of human psychology, a testament to our ability to empathize and connect. But a lot of researchers are deeply concerned about the implications of casually acting as if AI is a human, both ethically and epistemologically. They ask important questions like: Are we being fooled into believing these machines share our feelings? Could this illusion lead us to disclose personal information to these machines, not realizing that we are sharing with corporations? How does treating AI like a person cloud our views of how they work, who controls them, and how we should we relate to them?
I know these are real risks, and to be clear, when I say an AI “thinks,” “learns,” “understands,” “decides,” or “feels,” I’m speaking metaphorically. Current AI systems don’t have a consciousness, emotions, a sense of self, or physical sensations. So why take the risk? Because as imperfect as the analogy is, working with AI is easiest if you think of it like an alien person rather than a human-built machine. And I think that is important to get across, even with the risks of anthropomorphism.
Because AI is made of complex software, I find many people think of it as something coders should use. This attitude is so pervasive that you can see it everywhere - IT departments are often put in charge of corporate AI strategy, computer scientists are assumed to be experts on forecasting the social change AI might bring, and, most importantly - many people seem reluctant to use AI because they “don’t know computer stuff.”
This is like saying that, since we are made of biochemical systems, only biochemists should deal with humans - but it is even worse than that. It is more like saying that only chemists should be allowed to paint, because only they understand the molecular make-up of pigments. Why would we let artists, who may be completely unaware of how their paints are made, use such complex chemistry? But actually, it is even worse than that, because even computer scientists don’t really know why LLMs can do the things they do.
LLMs are made of software, but they do not work like most software. They are probabilistic and mostly unpredictable, producing different outcomes given the same input. Though they don’t think, they produce simulations of human language and thought that are, as far as we can tell, original enough that they out-invent most humans. The are perceived as more empathetic, and more accurate, than human doctors in controlled trials. And yet they are also limited in odd ways that are surprising, like an inability to do backward reasoning.
LLMs are essentially just a really fancy autocomplete. So how can a fancy autocomplete do these things? The answer so far, as described in an excellent overview in the MIT Technology Review is “nobody knows exactly how—or why—it works.”
The result is that working with these systems is just plain weird. Here I ask GPT-4 to create an animated image zooming into a dog. It first tells me that is impossible, then, after I give it a pep talk (“no you absolutely can do it! I have faith in you”) it decides it can solve the problem. But it comes up with a solution that doesn’t quite work, including giving me a link to a nonexistent file it said it created. So, I get slightly sterner: “you didn't actually write any code you know. seriously, this is odd. just do it.” And it does.
And then I ask it to make the animation better, and now I have a GIF that zooms into a dog, completely coded and developed by AI, managed by me.
This is animated, in case your browser doesn’t support GIFs
This doesn’t feel like working with software, but it does feel like working with a human being. I’m not suggesting that AI systems are sentient like humans, or that they will ever be. Instead, I’m proposing a pragmatic approach: treat AI as if it were human because, in many ways, it behaves like one. This mindset can significantly improve your understanding of how and when to use AI in a practical, if not technical, sense.
AI excels at tasks that are intensely human: writing, ideation, faking empathy. However, it struggles with tasks that machines typically excel at, such as repeating a process consistently or performing complex calculations without assistance. In fact, it tends to solve problems that machines are good at in a very human way. When you get GPT-4 to do data analysis of a spreadsheet for you, it doesn’t innately read and understand the numbers. Instead, it uses tools the way we might, glancing at a bit of the data to see what is in it, and then writing Python programs to try to actually do the analysis. And its flaws — making up information, false confidence in wrong answers, and occasional laziness — also seem very much more like human than machine errors.
This quasi-human weirdness is why the best users of AI are often managers and teachers, people who can understand the perspective of others and correct it when it is going wrong. You will notice, for example, in the conversation above, that I cut off the AI (using the “stop” button) when I saw it was going in the wrong direction and offered both feedback and correction. Rather than focusing purely on teaching people to write good prompts, we might want to spend more time teaching them to manage the AI. To getting them inside the non-existent head of the AI so that they can understand intuitively what works. After all, perspective taking is a form of social, rather than technical, proficiency and it can actually be learned.
The idea of treating AI like a person also aligns with the two of the best basic prompting techniques. The first is to give the AI a specific persona, defining who the AI is and what problems it should tackle. Telling the system “who” it is helps shape the outputs of the system. Telling it to act as a teacher of MBA students will result in a different output than if you ask it to act as a circus clown. This isn’t magical—you can’t say _Act as Bill Gate_s and get better business advice or write like Hemingway and get amazing prose —but it can help make the tone and direction appropriate for your purpose.
A second powerful technique that aligns with treating the AI as a person is chain-of-thought prompting, where you ask the AI to “think step by step” or provide clear instructions for it to follow. This not only results in better quality answers, but also lets us better understand where the AI’s thinking went off the rails. And, again, managers and teachers are often best at providing clear directions, which makes chain-of-thought prompting more effective. Talking to an AI like a person feels like a practical necessity when prompting.
And some of the risks of AI might actually be lower if their creators gave them more obvious fake personalities. You aren’t used to a computer making errors, but you know not to put complete trust in Sydney, your slightly-over-the-top and stressed-out intern, no matter how useful they are. Personality may end up being a differentiating factor for particular LLMs. You may like that Gemini is a bit of a planner, while others might like that Claude 3 is more willing to pretend to have emotions.
This “just talk to the AI like a person, rather than code” approach is also the technique used by one of the most important prompts in AI, the system prompt for Claude 3. System prompts are one of the ways in which AI labs set AI behavior by giving them initial instructions. Take a look at the text (and read the commentary from Anthropic) and you can see how close this prompt is to how you would communicate with a human about similar topics (the major difference being it is written in third person).
Ultimately, even if you don’t want to anthropomorphize AI, they seem to increasingly want to anthropomorphize themselves. The chatbot format, longer “memories” across multiple conversations, and features like voice conversation all lead to AI interactions feeling more human. I usually cover AI for practical uses in these posts, but many of the most popular AI sites are focused on creating AIs as companions - character.ai is the second most used AI site, after ChatGPT. And if you haven’t tried voice chatting with an AI model to see the appeal, you should. You can use a chatbot site, but you can also use Inflection’s Pi for free (at least for now, much of Inflection was just bought by Microsoft), or ChatGPT-4 via the phone app. These approaches seem to be working. An average discussion session with Pi, which was optimized for chitchat, lasts over thirty minutes.
Anthropomorphism is the future, in ways good and bad. An increasing number of humans already feel they have deep bonds with AI, with unpredictable results for our own interactions - helping some people while damaging the human relationships of others. Given the trend, treating AIs like people seems like an inevitability, so figuring out how to do it in safe, productive ways may be better than the alternatives.
Recommend One Useful Thing to the readers of Ulrich’s Newsletter
Trying to understand the implications of AI for work, education, and life. By Prof. Ethan Mollick