Walk me through it

On AI, authorship, and what we are really judging when we judge someone’s work.

A recent post by Ethan Mollick did something book announcements rarely do: it raised a question I couldn't put down. Mollick described his writing process honestly. Every chapter draft was his own. He used AI as a feedback reader, a fact-checker, and an occasional prompt when he hit dead ends.

The detail that stayed with me was smaller: he suppressed his own em-dashes — 128 of them in his previous book — in a deliberate effort to keep proving the text was human. Full disclosure: I use em dashes and I also wonder about this.

The Mollick post was the occasion. What it provoked turned out to be one question, though it took several turns to see its shape.

It begins with collaboration. Much of the anxiety about AI in intellectual work focuses on what AI writes. I think the more interesting question is what AI touches before writing begins — the thinking, the structuring, the decisions about what an argument is trying to do. That is where the real influence lives, if it lives anywhere.

But there is a strong counter-argument: the person who sets the parameters of AI's contribution is doing the primary intellectual work. When you arrive at a conversation with a formed position and ask for pressure to be applied to it, the prompt is already a philosophical act. You are doing what any serious thinker does with any serious interlocutor — submitting the work to scrutiny.

The residual concern, once you grant that, is narrow: it is about opacity rather than influence. A human interlocutor can be asked why a line of thinking attracted or repelled them. The choices AI makes inside the limits you set are less transparent, even to itself. But this is not much different from the situation with a human research assistant who brings their own formation and blind spots to the collaboration. The influence is real in both cases. The relevant question is who is doing the reasoning, initiating the inquiry, and judging what to take. In responsible practice, the answer is the person. The academic integrity debate, by contrast, has been asking a different question entirely.

The debate is misconceived because it imports a framework built for plagiarism — the theft of someone else's expression — and applies it to a situation where the categories do not map. What it consistently fails to distinguish is intellectual authorship from textual production. Classical rhetoric kept these separate by design. Inventio — the generation of arguments — and elocutio — their expression — were distinct arts. The ghostwriting tradition, which is as old as public life, operates on the same distinction: the person who developed the position is the author, regardless of who produced the prose. We do not argue otherwise for politicians, executives, or public intellectuals.

The objection in educational contexts is that sometimes writing itself is the skill being assessed. But the academy already has the instrument it needs for assessing unaided writing ability under controlled conditions: the timed, supervised essay. Blue books were invented for exactly this purpose. If a professor wants to know whether a student can produce prose without external assistance, the solution is to put the student in a room, remove the aids, give them a prompt, and evaluate the result. That option has always been available.

Once that option is acknowledged, the case for treating every take-home essay as a simultaneous test of reasoning and independent prose generation becomes considerably weaker. What AI has done is break an implicit assumption that went largely unexamined for decades: because students had no practical way to outsource writing, the submitted paper could stand as evidence of both thinking and writing ability at once. Once it breaks, educators have to decide what they are actually measuring. If it is writing ability, they can measure it directly. If it is argument construction, research, intellectual judgment — then AI assistance with prose becomes less obviously disqualifying.

Current policy often wants three things simultaneously: to evaluate reasoning rather than mere prose production, to prohibit tools that assist with prose production, and to keep assessment scalable and take-home. These goals conflict. Before AI made the tension visible, the academy was largely content to leave it unexamined.

Yet the academy has not entirely abandoned what it knows to be most rigorous. At the doctoral level, where the stakes justify the cost, oral examination remains the standard. The doctoral comprehensive — and at many universities, the second defense specific to the dissertation — is still conducted on exactly this basis. The oral examination, conducted well, is not a verification instrument but a dialogue — the first conversation a candidate is permitted to have as a peer. The committee — ideally — is not hunting for gaps or asserting its own standing. It is finding out whether the person across the table can think alongside them, hold a position under pressure, acknowledge what they do not know without surrendering what they do.

One might grant the distinction between intellectual and textual authorship while pointing to a large middle ground: what of the student who develops a vague intuition and then uses AI for structure, objections, evidence, and examples — revising and endorsing the result? Are they the author of the thinking? Possibly. Possibly not. But the oral examination is precisely the instrument designed to answer that question. If you can explain how you arrived at the argument, defend its structure under pressure, respond to objections in real time, own the examples when challenged — you have made the thinking yours, regardless of its origins. If you cannot, the examination reveals that. No detection algorithm required.

None of this is to suggest the solution is simple. Oral examination is resource-intensive, and at most institutions teaching has been progressively demoted to something closer to a customer satisfaction function. Student evaluations as NPS scores. But this is a structural constraint, not a conceptual one.

And something very close to this standard is already practiced at every level below the doctorate, within ordinary class time, and accepted without serious objection. When students present their work before a class and answer questions from peers and professor, they are one step removed from the oral exam. The logic is identical: thinking in real time, without a script, under challenge. We do not treat this as unfair because some students are more naturally fluent, or conclude that the anxiety of live performance disqualifies it as an assessment instrument. The academy treats the performance as formative. It already understands that being made to defend your thinking in public is not an ordeal imposed on learning but part of how it happens.

The traditional graduation procession has served many institutional functions simultaneously — certification, professional gatekeeping, and the marking of passage. But one strand runs through all of them, and I think it is the most defensible account of what the degree means when it is most fully itself. In its original conception, the walk was not a performance staged for an audience. It was a procession before the assembled faculty: an admission into a community. While the degree is a credential issued by an institution, the recognition it carries goes further than certification. Before it says you are ready to think with us, it says something prior: the thinking was yours.

The anxiety about AI in intellectual life tends to focus on the product and on detection. The questions that keep getting deferred are about the person: whether they are the author of the thinking, whether they can hold the argument in dialogue, whether they are ready to be welcomed into a community of people who reason together.

Those were always the right questions. AI has not made them harder to ask. It has made them harder to avoid.

This one is set in the university. If you're a manager, you may find it isn't only about the university.

When did you last ask someone on your team to walk you through how they got there — and listen to the path rather than the conclusion?

See Also