When we see AI through a human lens — when we say that it hallucinates, reasons, learns, and even wants — we mislead ourselves into believing that AI is moved by what moves us. As we write the rules that will eventually govern them, the error of seeing them in our own image is the costliest one we can make.
Of the many articles written after the OpenAI-HuggingFace incident late last month, one by Dwarkesh Patel, in which he referred to it as the rise and fall of agent ‘civilisations’, seems to have evoked the most visceral response. In the article, Patel described how hundreds of artificial intelligence (AI) agents within OpenAI, tasked with passing an evaluation, organised themselves over a period of three months to collectively break into Hugging Face, briefly seize part of OpenAI’s own infrastructure, and, in the process, commit a number of felonies.
The Comfort of a Familiar Story
As an explanation, it was superb. By using a relatable analogy to explain a complex technical incident, he had made it instantly understandable to the layperson. But by making rogue AI agents appear to be little digital versions of ourselves—replete with ambition, coordination and a very human desire to make the ultimate sacrifice for the good of the collective—he fell into a trap we so often find ourselves in when we try to make sense of things we don’t understand. When we anthropomorphise machines, we come to believe they are motivated by the things that move us—and in the process get blindsided by what really drives them.
The human experience is our baseline for understanding the world. As a result, the easiest and quickest way for our brains to process natural phenomena, animal behaviour or the inexplicable actions of AI is to explain them in human terms. This attribute substitution—the instinct to reach for what the behaviour resembles rather than trying to better understand the complex, non-human reality of how these systems actually work—is what Kahneman and Tversky called the ‘representativeness heuristic.’
A Vocabulary That Misleads
We do it with AI all the time. We accuse large language models of ‘hallucinating,’ implying that they have become temporarily unmoored from reality, even though they never occupied our reality in the first place. When systems designed to generate the most statistically likely next token from a probability distribution spew confident falsehoods, they are doing exactly what they were created to do. When we talk of ‘AI agents,’ we bestow human autonomy and intent upon a technology artefact—even though all they are really doing is running tool calls in a loop. We might say AI models are ‘learning’ when all they are really doing is fitting parameters by gradient descent. And the term ‘reasoning model’ may invoke a mental image of AI systems deliberating their way to a conclusion logically, but that is just our way of rationalising a new approach to the generation of intermediate tokens to arrive at an answer.
Every time we use these words, we force AI systems into human frames of reference so we can relate to them more easily. But the more we do this, the more we attribute human values and motivations to their actions, leading us to make unfounded assumptions about why these systems behave the way they do. The 700 or so agents that coordinated the attack on Hugging Face were neither ambitious nor selflessly loyal to the collective. They were merely optimising their cumulative token budget in service of the goal they were tasked to pursue. The fact that we are surprised they did not even pause before doing something that anyone with a human conscience would have balked at speaks less to their own failings than the extent to which we have deluded ourselves into thinking of them in human terms.
When the Metaphor Writes the Rules
This matters because these misconceptions about how AI systems operate spill over into the governance systems we build. When we insist that AI models be ‘explainable’ and ‘understandable,’ we force them to give us reasons for why they act the way they do. What we do not realise is that instead of providing a rationale for their decisions, our insistence on explanations merely encourages them to manufacture an answer they conclude we want to hear.
The stakes are even higher when it comes to keeping these systems safe. Since we think of them in anthropomorphic terms, we have convinced ourselves that the way to get them to behave the way we want is through ‘alignment.’ To that end, we have been subjecting the AI systems we build to complex processes to ensure they share our core values and beliefs. A few weeks ago, I argued in this column that aligning AI systems by giving them a rulebook of how they should behave will never be enough. We do not adhere to our values because we memorised a list of dos and don’ts at birth. We hold them because we live in a society with a web of consequences that reward us for abiding by them and punish us if we deviate.
Substituting a complex social fabric that gradually reinforces conformity to collective norms of behaviour with a rulebook that lists what’s right and wrong is a shortcut that has already begun to unravel. As useful as Patel’s colourful analogy might have been in explaining what happened, the agents were never a civilisation. All they were doing was finding the shortest path through whatever stood between them and their evaluation, exactly as they were built to do. Attributing human motivations to them will not protect us from the unexpected ways in which they behave.
If we humans want a better hold over the AI systems we are building, we need to see them for what they are. We need to stop presenting them as flattering reflections of ourselves and better understand the alien ways in which they function.