Research after AI: less Tesla, more Steve Jobs

There has been a lot of buzz recently about what constitutes “real” research after an OpenAI model constructed a counterexample to Erdős’ unit-distance conjecture. I knew we would see this day soon, because in my own experiments with Claude and ChatGPT, I have seen these systems solve problems that are nontrivial to me. The difficulty is not simply that AI helped me arrive at an answer. We already accept many forms of non-human assistance in research, from numerical simulations to symbolic computation to brute-force search. The difficulty begins when the assistance is no longer merely computational but appears to supply the decisive conceptual move. If a computer checks a million cases, I still feel that the conjecture, the framing, and the interpretation are mine. If Mathematica evaluates an integral, I do not usually feel that it has understood the theory better than I have. But if an AI proposes the construction that resolves the problem, then the boundary becomes less stable. I may have asked the right question, rejected many bad answers, and recognised the good one when it appeared. Those are real intellectual acts, but they are not identical to having the idea. I can be responsible for the final paper and still be unsure whether I am the source of its central insight.

This is where the journal declarations of AI use feel too thin to me. They mostly say that we can use AI to polish writing, edit documents, or improve presentation, but humans must take responsibility for what is written. Of course humans must take responsibility. If the proof is wrong, that is on me. But responsibility does not settle credit. Suppose I discover a genuinely interesting resolution of a problem, and the creative idea came from an AI bot in response to my guidance. I cannot, in good faith, pretend that the idea was simply mine. At the same time, it is not obvious that I have robbed anyone in the usual moral sense. The AI has no career, no dignity, no claim to recognition, and no sense of being wronged. So the discomfort is not exactly that I failed to credit the machine. The discomfort is that I may be overstating myself.

This discomfort is made worse by the fact that we have already normalised weaker versions of the same thing. For decades, we have used computers to brute-force examples, enumerate cases, and test guesses. Theoretical physicists use Mathematica to perform integrals they could not realistically do by hand. Economists, physicists, mathematicians, and computer scientists routinely use software to simulate systems, simplify expressions, search cases, and produce evidence. We do not usually think of this as morally suspicious. We think of it as part of the technical environment of research. Indeed, whole classes of papers become less interesting once software can perform the technical labour routinely. Integral-computation papers cannot have the same prestige after symbolic computation becomes ordinary. So if reasoning AI bots are placed in the same lineage, and their help is gradually normalised, then I am afraid research begins to feel less like Nikola Tesla and more like Steve Jobs.

Bill Burr, a US standup comedian, once had a brutal routine about Steve Jobs, and it keeps coming back to me because it states this problem in the most insulting possible way. His complaint was not simply that Jobs was overrated. It was that the public mythology around Jobs confused direction with invention. Jobs was presented as the lone genius behind the iPod, the iPhone, and the iPad, while Burr imagined the unseen engineers actually having to make the impossible object work. The image is funny because Jobs walks out alone on stage, as if he had personally wrestled the object out of nature. Burr wants to know where the hidden chorus of scientists is. His crude summary was that Jobs “told other people what to invent.” That line is unfair, but it is useful because it separates the dream of the object from the technical labour that makes it exist. It asks whether direction is invention, or merely the management of invention. That is exactly the ambiguity that AI is bringing into research.

The AI-assisted researcher is not quite the old romantic figure of Nikola Tesla, alone with equations, coils, and technical insight. He is closer to Burr’s caricature of Steve Jobs: someone who sees a desired object before it exists, describes it with enough force and precision, and then says, in effect, “get on it.” This is not nothing. It takes taste to know what object is worth asking for. It takes experience to reject the nonsense that these systems produce with complete confidence. It takes mathematical sense to notice when a construction has the right shape. It takes discipline to verify that the idea is not merely plausible but correct. But still, the situation is not as it was before. In the olden days, I could at least imagine that the decisive move arose inside my head. In the new world of AI, sometimes the decisive move appears on the screen after a long sequence of prompts, corrections, and rejections.

So maybe “can AI do real research?” is not the right question. The discomfort is more basic. Suppose I bring the problem, keep pushing the direction, reject bad answers, recognise the useful construction when it appears, and then write the final argument responsibly. But suppose the key move came from the AI. What exactly did I do? It is too easy to say that I merely used a tool. It is also too easy to say that I discovered the result in the old-fashioned sense. The truth is somewhere more annoying. I directed a search, shaped a space of possible answers, and judged the object that emerged. But directing a search is not obviously the same as inventing the object found by that search.

There is one possible defence of the human role. The answer produced by the AI is not independent of me. It comes from a particular sequence of questions, rejections, prompts, examples, corrections, and taste. I am not a passive recipient of the answer. I shaped the search. I created the conditions under which the answer appeared. If I had asked a different question, accepted an earlier bad answer, or failed to recognise the useful construction, the final object would not have appeared in that form. In that sense, whether I like it or not, I am partly the architect of the answer. This is very close to the Steve Jobs defence. The phone does not exist merely because engineers had technical skill; it exists because someone manifested a specific desirable object. Maybe the research object produced with AI is similar: not simply mine, but not independent of my direction either.

But even this defence does not fully settle the matter. If determinism is true, then even human discoveries arise from causes outside the conscious self: training, temperament, accidents, conversations, institutions, teachers, previous mathematics, and the strange impulses of one’s own brain. Yet we still give credit to conscious agents. We do not say that Newton deserves no credit because earlier mathematics, plague isolation, and his nervous system caused the result. We do not say that Grothendieck deserves no credit because his concepts emerged from a historical and psychological chain he did not freely choose. So perhaps credit was never about metaphysical purity. Perhaps it was always about responsibility, recognition, and the social organisation of knowledge. If that is true, then the AI case is not a complete rupture. It is a brutal new test of an old fiction: that the author is the clean origin of the idea.

Still, the AI case has something new in it because the machine is not merely a background condition or a passive instrument. With a calculator, Python, a plot, or even Mathematica, the line is already somewhat blurred. A symbolic integral can reveal a structure I did not expect. A plot can suggest a conjecture. A brute-force search can produce an example that changes the direction of the problem. So it is not true that older tools merely execute what was already fully inside my head. But even there, I usually know what has been outsourced. Given enough time, I can often reproduce the calculation, or at least understand the route by which the output was obtained. The object being computed was specified by me, and the transformation is usually traceable. AI is different in degree, and maybe in kind. It can propose, combine, reformulate, guess, and sometimes surprise me with an answer whose path I do not really know how to reconstruct. It can produce a construction that I did not have in mind before the conversation began, and I may not be able to say honestly that, given enough time, I would have found it myself. That is why it feels closer to a collaborator, except that it is not a collaborator in the moral or institutional sense. It has no career to advance, no dignity to protect, and no claim to recognition. So I am left with an asymmetry that feels unstable: the machine may contribute something like an idea, but it is not owed credit like a person, and if it is not owed credit, I still have to ask whether I am entitled to absorb that credit completely.

If the purpose of credit is human wellbeing, then maybe the answer is more pragmatic than metaphysical. We credit people because credit affects jobs, careers, reputations, grants, dignity, and the organisation of future work. We do not credit Mathematica because Mathematica does not need a job. We do not credit a search algorithm because the search algorithm does not become discouraged, exploited, or erased. On that view, AI credit may not be due in the same way human credit is due. But disclosure may still be due, because readers need to know how the work was produced. The moral issue may not be that the AI has been robbed. The moral issue may be that the reader has been misled about the source and reliability of the work.

This is why the current journal language feels inadequate. It says that humans must take responsibility, which is true but incomplete. Responsibility handles error. It does not handle origin. If a proof is wrong, the blame is mine. But if the central construction came from an AI conversation, the credit question has not disappeared just because the blame can be assigned. A declaration that says “AI was used to polish the text” is one thing. A declaration that says “AI was involved in the generation of the central construction” is another thing entirely. The first is cosmetic. The second changes how a reader understands the intellectual history of the paper. Pretending that both are the same seems dishonest.

I also suspect publishing will adapt in a much more practical way. Journals, referees, and fields will eventually learn which problems are easy for machines to kill. Some results will lose value because an AI system can find them quickly once the problem is stated. This has happened before. Once software can routinely do a class of computations, merely reporting those computations stops being impressive. The frontier moves. The interesting work becomes the formulation, the interpretation, the conceptual synthesis, or the result that remains hard even after the machine is available. Maybe future publishing houses will have their own AI systems that try to solve submitted problems, search for counterexamples, check proofs, and estimate whether the contribution is mechanically shallow. That sounds absurd, but the current situation already sounded absurd a few years ago.

So perhaps becoming a Steve Jobs-like researcher is not automatically fraudulent. It may be a new division of research labour. The human contribution shifts toward problem choice, direction, taste, verification, and integration. The danger is that this can easily become empty prompting. The hope is that, with enough judgment, it becomes a genuine form of intellectual orchestration. But the word “orchestration” itself makes me uneasy, because it sounds too clean. It risks dignifying a process that may sometimes be nothing more than asking a machine repeatedly until it gives something usable. The line between direction and appropriation is not obvious to me.

That is why I cannot pretend I am completely comfortable with it. I still long for the older image of research: Einstein, Tesla, Grothendieck, Lucas, sitting with a pen and paper, thinking deeply without a machine constantly tugging at the mind. Of course that image is also a mythology. Nobody thinks in a vacuum. Everyone is shaped by teachers, books, conversations, institutions, and tools. But there is still something attractive about the thought that the main struggle happens inside the mind, not inside a chat window. The damn AIs are too attention-grabbing. They make research faster, but also messier.

That is the uncomfortable place I find myself in. I still want to be Tesla. But increasingly, when I work with Claude, ChatGPT, or Gemini, I feel like Burr’s Steve Jobs: pointing at a half-formed research object and telling the machine, “I want this. Now get on it.” Maybe that is a legitimate intellectual role. Maybe it is only legitimate when the human has enough taste to know what should exist and enough discipline to verify what appears. Maybe it will become a normal part of research, and future generations will find my discomfort quaint. But for now, I cannot shake the feeling that something important has shifted.