Five Tentative Theses on Collaborating with AI

There is a great deal of anxiety among academics and intellectuals about what it means to think and write now that we can do significant portions of these activities with the help of machines. I have recently had conversations with very thoughtful people who have adopted what I would call a strong defensive position: AI erodes what makes the very activity of thinking, reasoning, and writing human. To use it is to surrender something essential.

There is a lot of truth in this. As a heavy user of large language models, I have already developed a powerful allergy to the bland, conciliatory, timid way AI writes. Every single day I look forward to my antidotes: I just finished teaching a class on Cicero’s speeches, I’m reading Edith Wharton and Henry James, I have a book club with a couple of friends on Thomas Aquinas. These days, to keep company with these authors is even more refreshing than in the past.

But the strong defensive position against the use of AI is at the same time quite superficial. It treats a genuinely difficult problem as though it had a simple answer. Here are my five tentative reasons for thinking it does not.

The categorical framing—Human activity versus AI activity—invites some extreme and likely impractical solution: ban AI from serious intellectual work, or embrace it without reservation. But the actual landscape is not a clean binary. It is a messy mixture of the two, and the messy-mixture framing, while less dramatic, reveals a number of narrower questions that are very difficult but potentially tractable.

Consider legal reasoning. Here are five things a judge might do with AI. They do not all seem the same to me. My intuitions about them are different and, I think, appropriately nuanced.

• Judge A receives from AI some facts she was unaware of. They are not in the record but can be legitimately used for adjudication. These facts change her mind on the outcome.

• Judge B brainstorms a logical argument with AI, and AI identifies a fatal weakness. Judge B abandons the argument, which changes her reasons for the decision.

• Judge C proposes a legal argument to AI and asks it to assess its soundness against existing legal materials.

• Judge D asks AI to provide the best arguments for and against the parties’ theories, then responds with objections, and asks AI to reply adversarially. This iterates several times. Eventually Judge D reviews the entire conversation and makes a final decision.

• Judge E has already decided which outcome is just. She tells AI several sets of plausible reasons in support of that outcome and asks AI to attack each of them. She chooses the reasons that fare best in this adversarial stress test.

If you think all five of these cases are equally troubling—or equally untroubling—you are not paying attention. It is going to be messy.

The strong defensive position tends to suppose that there is some pure, unmediated form of human thought which AI now threatens. But what would that form look like?

Did the word processor not change the way we write, and therefore the way we think about what we are writing? Did the PDF not change the way we read, and therefore the way we think about what we are reading? What about searchable databases of academic journals and books? Or the printing press. Or modern pens as opposed to quills. Or clay tablets that had to be incised with metal styli. Or, for that matter, the invention of writing itself.

There is no Platonic idea of human writing and thinking with no technological content. Whatever form of “human activity” the defensive position is defending is simply one of the latest versions of an interaction between humans and technology. The question is not whether to interact with technology—we have no choice in the matter—but which forms of interaction preserve or enhance what we value about the activity, and which forms degrade it.

The 2026 MLB season is the first to use an automated ball-strike system. I personally think this is bad for baseball. But I do not think that using technology to interpret what happened in a game is always bad. In a paper I recently published in the Harvard Journal of Sports and Entertainment Law, I argue that calling balls and strikes, despite its apparent simplicity, is an inherently interpretive practice, shaped not only by clear-cut rules but also by shared practices and unwritten principles about how the game ought to be played.

Technology can make this activity worse, but it can make it better. Consider instant replay:

Instant replay review allows managers to challenge certain calls, which are then reviewed from different camera angles at the Replay Command Center in New York. For many fans, the function of instant replays is to make the game fairer, which is to say, to make umpires’ calls more faithful to the independent reality of the play. This is, undoubtedly, a mechanical view of the game. In practice, however, instant replay extends rather than replaces human judgment in baseball. When thousands of spectators turn toward the jumbotron to scrutinize slow-motion footage, revising or confirming their impressions of what just occurred—often in spirited debate with friends and rivals—the interpretive activity at the heart of baseball is enhanced, not diminished.

An automated ball-strike system is, I think, different. Unlike replay, an electronic prism leaves no open texture for the exercise of human judgment. The centrality, fallibility, and contingency of human judgment in baseball are not imperfections of the game but essential features of its character. Were technology to become so accurate as to replace human judgment and render the interpretive conversation unnecessary, baseball would not become fairer it would become something else entirely.

Philosophers of sport are used to thinking hard about this problem. (Alva Noe has a similar take on replays and has written extensively on other difficult problems.) Replay review, high-tech simulators, elbow surgery for pitchers, performance-enhancing drugs: what kinds of interaction with technology make a sport less of a sport? The answer is nuanced and complicated. Technology is successful in baseball, I argued, when it expands and enhances the role of judgment and interpretation, not when it diminishes it. The same will prove true of AI in intellectual work.

Pshychologists and management scholars have been thinking about collaborative creation for some time, and their findings are suggestive. My friend Bess Rouse, in a careful study of creative dyads, argues that when collaboration works well the product becomes genuinely irreducible to individual contributions. We can’t really tell whether partner A contributed insight X and partner B contributed insight Y. Rather, the creative output is emergent: it belongs to the collaboration itself, and any attempt to decompose it into individual shares distorts what actually happened. Rouse notes that when one partner begins claiming disproportionate credit—when the collaboration is reframed as a star-and-assistant relationship—the collaborative dynamic itself is destroyed. The willingness to challenge and push, which is essential to developing better ideas, depends on a shared sense of ownership.

This is interesting because much of the anxiety about AI collaboration seems to rest on the assumption that contributions are always decomposable: the human had the ideas, the machine merely executed them, or (the nightmare scenario) the machine had the ideas and the human merely signed off. But if the best collaborative work is emergent and non-decomposable—if the best ideas arise precisely between the collaborators, in the friction and iteration—then the right question is whether human-AI interaction can produce that kind of friction.

There is a form of naive consequentialism, especially common among economists (the crowd I frequent most after legal academics), which cares only about the product. If the output is good, why should we care about how exactly we got there?

There is truth in this. We want to get inflation right, and vaccines, and citizenship rights, whether we do so with or without AI. And there are serious consequentialist concerns about the process even on its own terms. Accuracy is one: AI may get things wrong in ways that are difficult to detect. The labor market is another: what if we get inflation, vaccines, and citizenship law right but everyone who used to work on these problems is unemployed?

But the bigger truth is that we care about process too. The philosopher C. Thi Nguyen has a concept he calls “striving play” that captures a great deal of what it means to engage in intellectual activity:

In striving play, you try to win not because winning is important, but because the act of trying to win gives you a delicious struggle. In striving play, you don’t really care about winning in a lasting way. You temporarily induce in yourself a desire to win, so you can enjoy the process of trying. Striving play involves a motivational inversion of ordinary, practical life. In normal life, we struggle in order to attain some goal that we really want. In striving play, we adopt a goal in order to have the struggle that we really want.

This is a crucial feature of intellectual life. The researcher or writer who spends weeks working through a problem she could have answered in minutes with AI has not wasted her time. She has done something that is, in a precise sense, more valuable than the answer: she has had the experience of struggling toward it. The struggle is where the understanding lives.

Being skeptical of the strong defensive position does not mean denying this essential truth about what makes researching, thinking, reasoning, and writing truly meaningful: the playful, painful, frustrating, exhilarating struggle. But taking this seriously enough means to ask the genuinely hard question: under what conditions does collaboration with AI preserve or enhance striving play, and under what conditions does it destroy it? The answer is unlikely to be simple.