Can We Still Tell What We Know?

Working with AI can make a result clearer. Two recent studies suggest it can also blur how much of that clarity is ours.

When I work with AI on a difficult text, the result often becomes clearer. An argument finds its shape. A missing connection appears. A sentence that I struggled with for twenty minutes suddenly says what I was trying to say.

That is useful. But the improvement leaves me with another question: how much of that clarity is now mine?

I do not mean who owns the sentence. I mean something more practical.

If the conversation disappeared, could I explain the argument? Would I recognize where it is weak? And, a week later, would I remember which part of the idea I brought to the conversation in the first place?

In the first essay in this series, I asked which difficulties are worth keeping when AI makes good answers cheap. The question underneath was whether better work with AI also develops the person doing the work.

There is another layer to that problem. We may need to ask whether we can tell when development has happened.

The feeling of understanding

Consider a student reading a difficult passage. They can make notes, ask an AI questions, or do both. At the end of the session, one approach may feel much more helpful than another. That feeling matters: an accessible explanation can open a subject that otherwise stays closed. But it may not predict what the student will remember.

In a randomized study in English secondary schools, 405 students aged 14 and 15 read history passages using notes, an LLM, or a combination. Three days later, the students who had taken notes, with or without the LLM, performed better on comprehension and retention than those who had used the LLM alone. Yet students generally found the LLM more helpful than note-taking on its own. When they could combine the two, many preferred the combination. [1]

The interesting point is not that the students were wrong to value the AI. It could answer questions and make material accessible. The combined approach did well. What the study exposes is a gap between how helpful an activity feels and what remains after it.

Those are different measurements. The researchers did not show that every student was overconfident about their test score, nor did they isolate precisely why note-taking helped. But a student can leave a session with a genuine sense of progress that does not map neatly onto later understanding.

That possibility is familiar outside the classroom. A good explanation can make an idea feel obvious. A fluent summary can make a paper feel read. The moment of recognition (yes, that makes sense) is real. It is not always the same as being able to reconstruct the idea oneself.

The hidden contribution

A second study asks a different question. Instead of testing what people learned, it tests whether they remember how a piece of work came into being.

In an experiment with 184 adults, participants developed short ideas for problems and then wrote one-sentence elaborations. Sometimes they worked alone; sometimes they used an AI chatbot at the idea stage, the elaboration stage, or both. About a week later, they saw the results again and were asked whether the original idea and its elaboration had been made with or without AI assistance. [2]

They were especially likely to lose track of the idea’s source when AI had helped generate the idea but they had written its elaboration themselves. In that condition, the researchers estimated correct idea-source attribution at about 38%, compared with about 92% when both stages had been completed without AI. Confidence was also a weak guide to correctness in the mixed workflows. [2]

This is a specific memory experiment, not evidence that AI users generally claim credit for work they did not do. The study asked people to remember whether they had used AI, a simpler distinction than the messy question of intellectual ownership. But the pattern is striking: doing the later work ourselves may make the earlier contribution of the tool harder to see in retrospect.

The reverse matters too. If I supplied the original question and used AI to develop it, I might forget that my own choice of question was the decisive step. Collaboration can blur credit in either direction. The experiment does not resolve who deserves credit. It shows how unreliable memory of the process can become even before we reach that judgment.

Two kinds of self-knowledge

These studies are easy to fold into one warning about “AI overconfidence.” That would be too simple.

The school study concerns our judgment of what helps us understand and remember. The writing study concerns our memory for the source of an idea and our confidence in that memory. Someone could be well calibrated about one and mistaken about the other.

Together, however, they suggest a broader question. AI can change not only what we produce, but also the evidence we use to judge our own part in producing it. The polished answer is in front of us. The sequence of attempts, prompts, revisions and abandoned ideas is much less visible.

This problem did not begin with generative AI. In eight experiments published in 2021, Matthew Fisher and Daniel Oppenheimer studied how people judge their abilities when using outside assistance across memory and problem-solving tasks. People were better calibrated when help arrived after a delay or required an active choice. That work was not a test of LLMs. It does, however, suggest that the timing and manner of assistance can affect whether we notice how much we relied on it. [3]

This is where the speed of AI becomes interesting to me. A response can arrive before I have fully formed my own answer. If I never try to retrieve a fact, articulate a position or sketch a solution first, I lose a useful point of comparison. I know the joint result. I have less direct evidence about what I would have done alone.

That is an inference from several lines of research, not an established finding that instant AI responses damage self-knowledge. The difference matters. The question worth investigating is whether immediacy makes certain kinds of self-assessment harder, and whether a pause before assistance would sometimes help.

What would make the difference visible?

None of this argues for a permanent return to working alone. AI can make difficult material accessible, and the notes-plus-LLM group in the school study is a reminder that assistance can be combined with an active learning task. The question is how to preserve enough of the process to understand our own contribution.

There are simple possibilities. Before asking AI, try writing down what you think the answer might be. After using it to work through a paper, explain the central argument without looking at the chat. When developing an idea together, keep a trace of the initial question, the alternatives offered and the choices you made. These are proposals to examine, not interventions that the studies above have established as a universal remedy.

Their purpose would not be to make AI use slower for its own sake. It would be to retain some evidence of the thinking that happened on each side of the exchange.

Perhaps that is one of the less obvious skills we will need in AI-assisted work: the ability to distinguish a good joint result from our own understanding of it, and to remember how that result came about.

I can imagine a future in which those distinctions become easier to make. Tools could help us test what we understand, show how an idea developed, and make the points where our judgment changed more visible. I can also imagine becoming so accustomed to fluent collaboration that I stop noticing the boundary at all.

The research does not yet tell us which future is more likely. For now, it leaves me with a question I cannot answer by looking at the finished page:

When a thought feels like mine, what would tell me whether I really understand it, and how it became mine?

Sources

1. Pia Kreijkes et al. Effects of LLM Use and Note-Taking on Reading Comprehension and Memory: A Randomised Experiment in Secondary Schools. Computers & Education 243 (2026); available online 2025.
2. Zindulka et al. The AI Memory Gap: Users Misremember What They Created With AI or Without. CHI 2026. Full text.
3. Matthew Fisher and Daniel M. Oppenheimer. Harder Than You Think: How Outside Assistance Leads to Overconfidence. Psychological Science 32, no. 4 (2021): 598–610.