Jacob Nestle
Find the Hidden Premise
Not long ago, a colleague was late to a virtual meeting. “Sorry,” he said when he hopped onto the call. “My dog got out.” We all nodded, took two minutes to tell stories of when our own dogs had attempted an escape of their own, and got back to work.
There are a lot of things he didn’t have to say. He didn’t have to say “My dog should not be wandering around the neighborhood,” “If I can, I should spend time trying to find him,” or “The time I spent trying to find him is the reason I’m late.” Those are hidden premises — things that were implied and everyone on the call understood. If we felt the need to list every premise of every sentence in everyday life, we’d never get anything done.
What my colleague did is normal. Human beings leave things out when we communicate. We omit words when the meaning is obvious, skip explanations that our audience already understands, and move from one thought to the next without formally identifying every premise that connects them. But sometimes that same technique, the same shorthand that saves us time in normal conversation, can lead us astray. If we fill in the hidden premise with our own expectations, we can be wrong without ever understanding why.
As Always, Aristotle Got There First
The Greeks were thinking about this problem more than two thousand years ago. Aristotle called enthymeme the characteristic form of rhetorical proof. A classic example is: “Socrates is mortal, because he is human.” The speaker leaves out the premise that “All humans are mortal,” knowing listeners will infer it. People rarely make arguments by laying out every step as if they were solving a proof on a chalkboard. Rhetoric rarely follows all the rules of philosophy. Speakers rely on things their audience already believes, knows, or is willing to assume.
A related idea is ellipsis, where words or parts of a sentence disappear because context supplies them. If someone asks, “Coming?” no one needs to interrupt and insist that the correct sentence is “Are you coming with us?” The missing words are doing no damage. They are missing because everyone involved already knows what they are.
The same thing happens with arguments.
Under normal circumstances, that’s perfectly fine. Most of the arguments we encounter in real life are not formal demonstrations built from premises that have all been stated and proven. They depend on shared beliefs and assumptions. Simplicity makes communication possible.
Sometimes, it gives a bad assumption somewhere to hide. Aristotle took the time to note the difference between philosophical syllogisms and “enthymeme” rhetorical syllogisms because we often mistake one for the other. Both are forms of argumentation, but they are not created equal.
Self-Awareness Is Self-Defense
Often, we find ourselves nodding along to sentences that seem perfectly reasonable. We’ll take one common example: “We shouldn’t trust that source, the author was biased.” There is a conclusion, then a reason given, but the argument rests on you assuming the bridge between them. We assume that the author intends a premise like “A biased source is not trustworthy.” The speaker uses an enthymeme to bring us into their argument, making us assume the premise and, in supplying it, follow their train of thought.
Unstated, the premise can grow or shrink in our minds to match our personal assumptions. In the “bias” example, the speaker’s intention could be a narrow or well-supported argument that the author allowed their bias to color their writing. It could also be an overbroad leap to the belief that we have to pretend every source is totally objective to trust it at all.
We can see how much depends on this hidden premise if we make the example more specific. Frederick Douglass was a biased author — against slavery. Thomas Jefferson was a biased author — in favor of American independence. Abraham Lincoln was a biased author — to defend the Union. Each of these beliefs will naturally color how they approach those issues, but that doesn’t mean they lose their value because of it. A sentence like “We shouldn’t trust Lincoln’s take on the Civil War because he believed in the Union” should raise eyebrows.
That doesn’t mean bias is irrelevant. We can and should question the texts we use and investigate the author’s intentions. But when we find ourselves supplying the argument, we have to be careful that the assumptions we bring are well supported.
Enthymeme is not necessarily malicious or manipulative. Remember, it’s rooted in the way that humans themselves talk. It’s natural. But it can also become the Trojan horse through which bad ideas begin to take over our thinking.
To avoid falling for a bad use of enthymeme, we have to know ourselves. We have to stop and ask ourselves what we’re assuming.
I think this is one of the more useful habits we can teach students because it gets them past the easiest version of disagreement. A student can say, “I disagree with the conclusion,” but that does not tell us much. If we ask what has to be true for the conclusion to follow, we can often find the real disagreement underneath it.
A student who cannot identify the deeper assumptions they are making becomes susceptible to poor reasoning both in and beyond the classroom.
We spend a great deal of time telling students to fact-check what they read. They should. If somebody says that George Washington wrote the Gettysburg Address, there is no need to search for a clever alternative interpretation. The claim is wrong. But a person can get every fact right and still make a poor argument. They can bring in dangerous, overbroad ideas and premises, and they might not even know that they’re doing it.
Fact-checking can tell us if a claim is true. But if you don’t know the premises, you can’t check them. Students need to be able to step back, realize what they’re bringing to the table, and prepare themselves to evaluate the whole argument, not just the visible parts of it.
You Need to Do This With AI
Large language models are extraordinarily good at producing plausible connections between ideas. That is one of the reasons they’re useful. Give an AI a pile of notes, a complicated subject, or an argument that needs stress-testing, and it can organize information at a speed that would have been unimaginable a few years ago. Just because the sentences fit together, however, doesn't mean the reasoning between them is sound.
AI might tell a student, “Industrialization dramatically increased national wealth, so it improved the lives of ordinary Americans.” That sentence is factually correct. Industrialization did increase American productive capacity and wealth enormously. But the conclusion depends on another assumption: that increases in national wealth translate into improvements in the lives of ordinary people in the way the answer implies. Wages, prices, working conditions, housing, health, hours of labor, regional differences, and who actually captured the new wealth all matter before we can make the jump from “the country became richer” to “ordinary Americans became better off.”
A student who reads and understands the first sentence is not wrong. But they do not understand what is actually happening if all they do is literally or mentally copy and paste that knowledge. The AI does not have to invent a statistic for the answer to be weak. It can cite real numbers and still move too quickly from those numbers to a conclusion. A student who memorizes everything in the answer might still miss important details. If the AI provides the bridge between evidence and conclusion every time, a student memorizes facts without learning judgment.
The current conversation about AI literacy can become too focused on hallucinations. Hallucinations are easy to understand: The AI says a book exists, you look for the book, and the book does not exist. The tool was wrong.
Those mistakes matter, but they are not the only mistakes a student has to learn to catch in order to think for themselves. A much harder problem appears when every source exists, every date is correct, and the argument is still skipping steps.
At Lexandria, we use AI tools regularly, but we don't let a model make the final judgment. A person still has to check the source, evaluate the argument, and decide whether the answer makes sense. AI can perform some of the intellectual labor involved in producing an answer. We never let it relieve the person using it of responsibility for judging the answer. That distinction is central to how we think about AI in education.