For years, AI in research has mostly been framed as an assistant: helping scientists search literature, write code, analyze data, or test ideas faster.

That boundary is beginning to move.

OpenAI recently reported that its researchers are increasingly using coding agents throughout their daily work. Researchers are contributing code faster, running more experiments, and delegating increasingly complex tasks to agents. The direction is clear: AI is moving deeper into the research process itself.

This raises a more interesting question than whether AI can make scientists more productive:

If AI can run the experiment, write the code, and analyze the result, what is the scientist actually responsible for?

We may already be seeing an early version of that future in mathematics. An internal OpenAI model recently disproved a long-standing conjecture related to Erdős' unit-distance problem, producing a new construction that was subsequently checked by external mathematicians.

The significance is not simply that an AI system solved a difficult problem.

It is that the boundary between using AI to do research and AI participating in discovery is becoming less obvious.

That does not make the human researcher irrelevant. It may instead change where human value is concentrated.

Running experiments, generating implementations, and exploring large solution spaces can increasingly be delegated. But deciding which questions are worth asking, recognizing why a result matters, challenging assumptions, connecting discoveries to the wider world, and deciding what should happen next remain fundamentally different responsibilities.

There is also a risk in that transition. Research on human oversight of AI agents warns that extensive automation can weaken the very expertise and critical judgement humans need to evaluate automated systems effectively.

So perhaps the future of research is not about keeping humans responsible for every step.

It is about making sure humans remain responsible for direction, meaning, and judgement.

AI may increasingly help us discover answers.

But choosing which questions deserve an answer should remain a deeply human task.

References

  1. OpenAI. (2026). Research acceleration: The view inside OpenAI. https://openai.com/index/research-acceleration-view-inside-openai/
  2. OpenAI. (2026). An OpenAI model has disproved a central conjecture in discrete geometry. https://openai.com/index/model-disproves-discrete-geometry-conjecture/
  3. Mitchell, M., Ghosh, A., & Passi, S. (2026). AI Agents Push Humans Out of the Loop. arXiv:2608.23642. https://arxiv.org/abs/2608.23642