Semantic Ablation

It was about a year ago that I signed up for a paid OpenAI subscription. It is no exaggeration to say that the return on investment I get for those 20 USD per month far outweighs almost anything I have ever bought in my life. Its usefulness extends across domains, from academic administration to travel advice, and I am convinced that it has made me a better teacher, supervisor – and perhaps even runner.

In the beginning, much of the concern revolved around “hallucinations”: the tendency of generative AI models to make up facts that are not true. Over time, as the models have improved and I have become more attentive to the kinds of mistakes they make, I have grown less worried about hallucinations and more concerned about what might be called “semantic ablation”. If hallucination is the model seeing what is not there, semantic ablation is the model quietly erasing what is.

The process is subtle. You paste a jagged paragraph into the machine – something slightly overdetermined, perhaps overly metaphorical, maybe even a bit too fond of its own terminology – and ask for “polishing”. What comes back is smoother. Cleaner. More readable. And yet something has been lost.

The rare word is replaced by a more common synonym. The technical term becomes “accessible”. The structure is straightened into a respectable, well-tempered five-paragraph march. Nothing is wrong. But neither is anything quite alive.

Statistically, this makes perfect sense. Large language models are trained to move toward the center of probability distributions. The tail – where idiosyncrasy, precision, and intellectual risk often reside – is shaved off in the name of likelihood and helpfulness. The result is not error, but regression to the mean.

And perhaps that is the deeper danger. Not that the machine invents fantasies, but that it gently encourages us to abandon complexity. Not that it deceives us, but that it smooths us. A civilizational drift toward the middle, where friction is minimized and originality becomes statistically inconvenient.

Used carefully, these systems are immensely helpful when it comes to clarifying thought. Used unreflectively, they may erode it. The question is not whether we can write with AI, but whether we can do so without allowing our sentences – and eventually our thinking – to suffer semantic ablation.