Sabinus and the AI Trap

In one of his letters to Lucilius, Seneca tells an interesting story about a fellow called Calvisius Sabinus.

It goes like this, in a modern translation:

In my own lifetime, there was a rich man named Calvisius Sabinus, a man with a freedman’s money and a freedman’s mind. I never saw anyone whose good fortune was so out of place. His memory was so poor that he would sometimes forget the names of Ulysses, Achilles, or Priam, names we know as easily as we know the names of our own servants. No senile steward, unable to call people by their proper names and forced to make them up, ever mangled the names of his master’s household as badly as Sabinus mangled the names of the Trojan and Greek heroes. And yet he wanted to seem educated. So he came up with a shortcut to learning. He paid absurd sums for slaves, one who had memorized Homer, another who had memorized Hesiod, and then one slave each for the nine lyric poets. And do not be surprised that he paid so much for them. If he could not find the kind he wanted, he had them specially trained. Once he had assembled this entourage, he began tormenting his dinner guests with them. He kept these men at the foot of his couch and would ask them now and then for lines of poetry so he could repeat them himself, though he often got stuck halfway through a single word. One of his hangers-on, Satellius Quadratus, a man who fed off witless millionaires, flattered them, and, as often happens, mocked them too, suggested that Sabinus should also hire scholars to collect the scraps when he forgot. Sabinus replied that each slave had cost him a hundred thousand sesterces. Satellius answered, “For less money, you could have bought the same number of bookcases.” But Sabinus kept insisting that whatever anyone in his household knew, he knew too.

When I read this passage, AI immediately came to mind. The resemblance is hard to miss. These days, of course, one does not need to keep slaves or be a millionaire. And the access is far broader. It is not a few books or poems, but practically all of human knowledge within reach.

To be fair, access itself is not new. We have had that for decades with the internet and search engines, most notably Google. But until recently, access still required more work. You had to search, compare sources, read, and then hopefully understand, learn, and retain at least some of it.

That is why I think the opportunity here is fantastic. With the right mindset, AI can serve as a virtual tutor or mentor, drawing on the work of world-class performers in almost any field. You are not limited by page ranking, recency bias, or scattered fragments of information that need to be painfully stitched together so that they make sense. With a good prompt, you can get a study guide, reading recommendations, videos, and a self-paced path into almost anything.

But the temptation on the other side is obvious. You can also let the machine hand you words, ideas, and conclusions, then treat them as if they were yours. Sabinus seemed convinced that if anyone in his household knew something, then he knew it too.

That is what makes the analogy uncomfortable.

A 2025 arXiv paper, The AI Memory Gap: Users Misremember What They Created With AI or Without, tested whether people could correctly remember which ideas came from themselves and which came from an LLM-based chatbot. In a pre-registered experiment with 184 participants, the authors found evidence of source confusion between self and AI, especially in mixed human-AI workflows. In other words, people were not always clear on what they had actually produced versus what the system had helped produce.

Another 2025 study, reported by TechXplore and published in Computers in Human Behaviour, found that people using ChatGPT overestimated their cognitive performance, and that higher AI literacy was associated with greater, not less, overconfidence. The researchers describe this as a reversal of the usual Dunning-Kruger pattern in this context.

That last point hits home for me. I can easily see myself falling into that trap. I usually know my lane, but with AI, I can push a little further and move a little faster. That part is real. The uncomfortable question is how far I can go before I stop extending my own thinking and start borrowing the appearance of competence from the machine.

Here is how the Sabinus story wraps up, again according to Seneca:

Then the same Satellius urged Sabinus to learn wrestling. Sabinus, frail, pale, and thin, said, ‘How can I? I am barely alive as it is.’ ‘Please,’ Satellius replied, ‘don’t say that. Just look at how many strong, healthy slaves you have.’

That is the danger.

AI can help us learn faster, think more clearly, and go farther than we could on our own. But it can also make it dangerously easy to confuse access with understanding, output with mastery, and assistance with ability. And once that happens, we are not that far from Sabinus after all.