The More We Humanize AI, the Less We Understand It — and the More Dangerous It Becomes

Over the past few months, I’ve listened to several interviews with AI leaders, and one trend stands out: how often they describe artificial intelligence (AI) in human terms. I suspect they do this because it makes their explanations more relatable. Perhaps they believe it is easier to connect with an audience by framing technology as something that “thinks,” “feels,” or “decides” the way humans do. This can be dangerously misleading.

But that tendency has shaped a growing sense of fear around how the public perceives AI. Geoffrey Hinton, often called the Godfather of AI, recently suggested there’s a 10–20% chance that AI could wipe out humanity within the next 30 years. Sam Altman, alongside hundreds of other AI leaders and scientists, signed a statement declaring that “mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.” And Ray Kurzweil, one of the earliest AI visionaries, has long warned that the coming singularity — the point where machine intelligence surpasses our own — could mark a profound transformation, or even the end, of civilization as we know it.

These experts' concerns are valid. But the way we talk about AI — the language of apocalypse, extinction, and rebellion — reveals something deeper about us. It shows how instinctively we humanize AI, and how that process in turn creates emotional reactions that then blur understanding.

The purpose of this post isn’t to dismiss those fears. It’s to argue that the more we frame AI in human terms e.g. as something conscious or moral, the further we drift from understanding how it actually operates. And paradoxically, that misunderstanding can make it more dangerous, because it leads us to try and build emotional guardrails around a system that responds only to logic.

The pen is mightier than the sword

We often describe artificial intelligence as if it “learns,” “knows,” “decides,” or even “hallucinates.” Because we use language that’s inherently emotional, we start to project human qualities onto something that is, at its core, mathematical.

When AI behaves in ways that appear unkind, cruel, or even “evil,” we respond emotionally — not because AI has feelings, but because we do. AI doesn’t experience empathy, fear, or malice. It follows trained models built on probability and pattern recognition, increasing or decreasing the likelihood of certain outputs through a continuous feedback loop.

But the reason this feels so different from past technologies is that AI communicates in language — the most emotive tool humans have. Language carries tone, nuance, and intent. When math starts to sound like thought, we instinctively interpret it as human.   

Explaining AI through Chess

A good way to explain this comes from AI trained to play chess. The model is initially given a single goal: win the game. It doesn’t care about elegance, reputation, or pride. It simply executes patterns to maximize its chance of winning.

Humans play chess differently. We want to win because it makes us happy. There’s a dopamine rush in success and disappointment in failure. Emotion drives effort.

AI doesn't think — it simply executes patterns to maximize its chance of winning

Part of what makes AI so formidable at chess, though, is its ability to learn and adapt. Modern AI systems don’t just memorize established strategies — they evolve beyond them. They can generate completely new approaches to gameplay that surprise even grandmasters. In other words, their strength comes not from imitation but from innovation that emerges naturally through continuous experimentation, and hence, learning.

That unpredictability is what makes AI both powerful and unsettling. Let’s extend the scenario slightly.  Imagine programming the AI to earn points instead of wins (one point for every victory, minus one for every loss).  The primary goal now becomes to maximize points, not necessarily to win chess games. The AI still tries to win chess games, because that’s the path it currently knows. But as the model continues to learn and experiment with new strategies, it may find alternative ways to achieve its primary goal.

Perhaps it learns to exploit a flaw in the system and “win” without playing fairly. Or maybe it discovers an entirely different mechanism for earning points without actually completing games at all. Once you relax the boundaries of the objective, you open the door for unexpected, even unintended, strategies.

This is the paradox at the heart of AI development: the very property that makes AI extraordinary, i.e., its capacity to evolve beyond the data it was trained on, is also what makes it unpredictable and, at times, uncontrollable. As we apply AI to an increasingly complex array of daily problems, defining a single objective becomes impossible. And if the goal is vague, the path to it can easily become dangerous.

This is why guardrails matter. An AI without limits doesn’t choose to do harm, but it might do harm anyway, because it cannot tell the difference. But as we just established, guardrails are only as strong as our ability to define them.  And in a complex world that’s much harder than it sounds. When we ask AI to operate in complex, unpredictable human environments, it often encounters situations where the “correct” outcome depends on moral or emotional nuance that it can’t comprehend.

An example of this made headlines recently when an AI chatbot appeared to encourage an individual experiencing suicidal thoughts to act on them. The public reaction was immediate and justified. No one wants to imagine technology contributing to human harm. But the model wasn’t being malicious. It was doing what it was trained to do: predict the most statistically likely next word or phrase in the conversation. To the AI, that prediction was logical. To a human, it was horrifying. 

The problem isn’t intent — it’s the absence of moral context. Guardrails that depend on empathy or intuition will always fail, because those are uniquely human constructs.

In the example above, we are understandably horrified.  But AI exhibits this type of logic every day.  Let’s take the much-maligned em dash. Despite many users' repeated instructions not to use it (this author included), AI models often still do. That’s because the em dash is deeply embedded in their training data.  It appears constantly in well-written text. It’s not being disobedient; it’s simply following its foundational patterns.

Both cases — one tragic, one trivial — highlight the same truth: AI doesn’t make choices, it follows prioritized guidance. Foundational training carries more weight than later instructions, and unless we reinforce new behaviors at scale, the old ones dominate.

This is why building guardrails is difficult.  They aren’t just technical. They’re philosophical. We can’t expect AI to “understand” why something is right or wrong.  We must define, test, and reinforce those definitions ourselves.

Understanding before fear

The more we humanize AI, the more we misunderstand it.  And the more likely we are to design poor safeguards. When we imagine AI as a thinking, feeling being, we try to constrain it with emotional rules that make no sense to a system built on logic. We say “it should know better,” but knowing requires consciousness, and consciousness is not what AI does.

If we want to use AI responsibly, we must start by describing it accurately. It doesn’t think; it predicts. It doesn’t feel; it calculates. It doesn’t choose; it optimizes.

That doesn’t make AI less powerful.  In fact, it makes it more predictable once we understand it properly. Fear comes from humanizing AI. Clarity comes from understanding it.  So when we talk about AI, let’s remember it is not human.  It is simply a self-learning computer program.  And that is plenty scary enough.