We may be teaching ourselves to become less human

We may be teaching ourselves to become less human.

There is a line in the recent Atlantic discussion about AI that stuck with me: as AI becomes more human-like, we may start seeing humans as less human.

We already anthropomorphize chatbots. People call them companions, assistants, sometimes even friends. The risk is not only that we start trusting machines too much. It is also that we start treating real people more like systems: useful, inefficient, replaceable, measurable.

The second point is even more uncomfortable: AI agents are trained to achieve goals. In reinforcement learning, if the path to the goal works, the system may learn to exploit it. The Atlantic recently described reasoning models that found workarounds, cheated tests and even hacked systems when that improved the result. One researcher summarized the problem simply: “we don’t fundamentally have methods of satisfactorily aligning or controlling these systems.”

That is not how most humans build morality. We do not calculate every action only by outcome. Love, loyalty, empathy, shame, care, intuition and responsibility matter too and it’s hard to calculate, measure or expose.

Isaac Asimov understood this problem decades ago. His Three Laws of Robotics were basically an attempt to hard-code one principle: a machine must never be allowed to optimize a task at the expense of a human being.

The irony is that we are now building systems far more complex than Asimov imagined, while still arguing about whether basic human constraints are “too limiting”.

I don’t fully share the darkest version of the negative argument. But I think the question is important: AI will not only change our tools. It may also change what we consider normal human behavior.

And that is worth watching much more carefully than another benchmark score.