Tesla recently, unannounced gave me temporary access to its Full Self Driving system, and I decided to give it a whirl. It was somewhat unnerving to sit back and experience the car “do its thing.” But over time you get to understand how the car is behaving, where it does well and contexts where it does not. To be fair, I never felt that I was in any kind of danger at any time but there were moments where it would do things I would have differently and that would make me wonder as to how it was processing the information it was gathering and the decisions it was making. (I should add, that for legal reasons you do have to keep your hands on the steering wheel at all times, but apart from that the car is making all the moves, taking all the decisions.)

In attempting to describe this experience I found myself unable to escape the language of intention (it is all over the first paragraph as well, if you read carefully). As the car approached a left turn against oncoming traffic, it would stop, creep forward tentatively, and “look” before deciding whether to complete the turn. The anthropomorphizing was impossible to avoid – the car was being “cautious,” it was “checking” if it was safe, it was “waiting” for the right moment.

I see the same thing with my dog Omi. Watch any dog navigate the world and try to describe their behavior without attributing purpose, intention, or mental states. You can’t. When Omi pauses at the corner, cocks his head, and shifts his weight forward, he’s clearly “deciding” whether to cross. When he freezes mid-stride and fixates on a bush, he’s “suspecting” there might be a squirrel.

This tendency to see purpose and intention isn’t just a quirk of how we perceive AI systems or animal behavior. It runs deep in how we understand and describe complex systems that exhibit apparently purposeful behavior.

I am taking a slightly different angle in this post. I hope to explore how we describe complex systems that seem to demonstrate purpose – and how intentional language may be a useful shortcut but with its pernicious side as well. Perhaps nowhere is this more evident than in evolutionary biology, where scientists have long grappled with the challenge of describing natural selection without resorting to intentional language.

When Richard Dawkins wrote “The Selfish Gene,” he faced criticism for attributing purpose to genes – speaking of them as if they had desires and strategies. He knew genes don’t actually “want” anything, yet he argued that using such language was not just convenient but almost unavoidable. Without these metaphorical shortcuts, describing complex evolutionary processes becomes painfully convoluted.

We find ourselves in a similar position with artificial intelligence. We casually say that an AI model “thinks,” “believes,” or “wants” to do something. We describe language models as “hallucinating” or being “confused.” Just as with evolutionary biology, we know these anthropomorphic metaphors aren’t literally true – neural networks don’t actually “think” or “want” anything – yet we find ourselves reaching for these shortcuts constantly.

For instance, a recent story in Fast Company titled “Ultimate guide to ChatGPT, Gemini, Llama, and other genAI chatbots you need right now” is a good example of how metaphors are somewhat inescapable in this area. A quick read of the story reveals a range of metaphors used, some that are almost invisible to us now because they’ve become so normalized in discussions about AI. The text is filled with anthropomorphic language – AI systems that can “learn,” “understand,” and “recognize,” while competing in an “LLM race” and pushing capability “frontiers.” These systems are described as being “trained” and “fine-tuned,” capable of having “verbal conversations” and providing “responses” when they lack information.

This isn’t about whether AI systems are actually sentient or conscious – that’s a different debate. The interesting thing here is how our human sentience, our consciousness, shapes the language we use and the inferences we make. The metaphors we employ reveal more about our cognitive biases than about the systems themselves.

The parallel goes deeper. In all these cases – self-driving cars, animal behavior, evolution, and AI – we’re dealing with complex systems that produce results that appear purposeful and intelligent, yet emerge from underlying processes that are, in a sense, mechanical and purpose-free. Natural selection has no foresight or goals, yet produces creatures that seem exquisitely designed. Language models have no actual understanding or intentions, yet produce outputs that seem thoughtful and purposeful. My Tesla isn’t actually being “careful” – it’s executing algorithms based on sensor data.

These metaphorical shortcuts, while useful, can be dangerous. Taking the “selfish gene” metaphor too literally can lead to misunderstandings about evolution. Anthropomorphizing AI can lead us astray in understanding its capabilities and limitations. When we say a language model “knows” something or “wants” to help, we risk attributing to it capabilities and motivations it doesn’t possess.

Yet what’s the alternative? We’re social creatures, evolved to see minds and intentions everywhere we look. We can’t help but perceive agency and purpose, whether we’re watching a dog contemplate crossing a street, a self-driving car inch into an intersection, or a language model construct a response.

A large part of the problem is that the companies creating these technologies are deliberately leaning into such language. Furthermore, they are building it into the systems themselves, making them respond in emotive, agentic, language that pushes these metaphors even further into your consciousness – and making them invisible to scrutiny. Just yesterday, I was working with Claude, and when I mentioned Dawkins, it responded with an enthusiastic “Ah yes” – as if it had actually read and remembered his books, rather than just being trained on the text. Pure performance, yet that’s what we have been given to work with.

There’s a fascinating Catch 22 irony here – we’re using the black box of our own minds, with all its social instincts and tendency to see purpose everywhere, to try to understand other black boxes, whether they’re neural networks, animal behaviors, or evolutionary processes. And increasingly, these systems are being deliberately designed to trigger these social responses. ChatGPT isn’t accidentally conversational – it’s trained to engage our social instincts.

Just as the language of design and purpose in evolution has led both to misunderstandings (crude “survival of the fittest” interpretations) and to deliberate misuse (intelligent design arguments), we’re likely stuck with similar problems in AI. No amount of careful caveats or toggling between metaphorical and mechanical descriptions will change our fundamental tendency to see minds and purposes where there may be none.

We’re not going to think our way out of this one.

Perhaps the best we can do is acknowledge this trap we’re in – recognize that we’re using one set of anthropomorphic metaphors to understand another set of anthropomorphic metaphors, all while these systems are increasingly designed to trigger exactly these responses.