Researchers Make AI Their Confidant, But Can’t Afford To Let Their Guard Down

Imagine spending a decade trying to solve one of mathematics’ hardest problems. You finally make progress only to see an AI model produce a possible solution in a matter of days.

This is what happened to Tristan Buckmaster, a 41-year-old Australian mathematician at New York University who has spent years studying the mathematics of how fluids move and behave.

At the core of the issue are the Navier–Stokes equations, a set of mathematical equations developed in the 19th century to describe how fluids move. Engineers and scientists use them to model everything from airflow around aircraft to weather systems.

But mathematicians still aren’t sure whether these equations always hold, or could break down and produce an infinite or otherwise impossible result. Solving this is one of seven Millennium Prize Problems.

According to OpenAI, it asked an unreleased AI model to work on the problem on September 1. Four days later, the model produced a possible proof.

This is exactly where the controversy starts.   

Buckmaster and his collaborator Levent Alpöge had put drafts of their unpublished research into OpenAI’s coding tool, Codex. Buckmaster was concerned that OpenAI’s AI may have drawn on his unpublished work, but he had no way to prove that it had.  

While OpenAI denied that the mathematician’s work influenced the system, it acknowledged that data from users’ interactions with its products may, depending on the settings, be used to improve its models.

The entire episode is especially concerning for researchers, engineers and academicians, who routinely put unpublished papers, code, designs and ideas into AI tools. And if AI models are using the research work, how can researchers safely use AI without exposing work that is not yet public? Does India have rules to answer this question? Let’s find out…

In Indian Labs Draft AI Rulebook 

As AI seeps into all walks of life, all kinds of workplaces are utilising it, and research and academia are no exception. One of the significant examples is Google Deepmind’s AlphaFold, an AI system that predicts the 3D structure of proteins from their amino acid sequences with high accuracy. This also helped Demis Hassabis and John Jumper of Google DeepMind win the 2024 Nobel Prize. This system has been utilised by over 3.3 Mn researchers worldwide to accelerate breakthroughs in biology, medicine, and disease research.

In fact, research intelligence firm Wiley found that the overall usage of AI tools surged from 57% in 2024 to 84% in 2025, including specific use for research and publication tasks, which grew significantly to 62% from 45%. 

Indian academicians and researchers are also relying on AI at varying degrees. For instance, professor Arindama Singh, head of the department of mathematics at IIT Madras, uses AI for literature review and idea generation. However, he added that his approach to AI usage is cautionary, a practice he has been following even before the OpenAI-Buckmaster case blew up. 

“We do not give our complete solutions to AI for checking; we rather suggest an idea to check. We do not ask AI to check a new method in its entirety, or any patentable idea,” said Singh. 

Singh is not alone in drawing that line, and institutions are now trying to formalise what individual researchers have been doing by instinct.

At IIIT Hyderabad, the caution extends to which models get used. The institute has no institutional AI subscription. It has compute of its own — smaller open-weight models running on campus servers, with individual projects maintaining their own GPU clusters. Those are what researchers are encouraged to use when the data is sensitive.

The reasoning is jurisdictional as much as technical. According to Sandeep Kumar Shukla, professor and institution director, the problem with a frontier model is that the user controls nothing behind it.

That thinking is now being written down. IIIT Hyderabad constituted a committee in May to frame an institutional AI policy, covering how much AI is permissible for students in homework and lab work, and for research scholars and faculty. 

Deemed university MAHE, Bengaluru Campus, is at a similar stage. Mangalampalli Surya Sudheer, assistant director of research told Inc42 it is developing institutional guidelines for the responsible and ethical use of AI, including the secure handling of unpublished research. 

As of now, researchers are expected to rely on broader principles of research integrity and confidentiality.

Where Indian Law Leaves Researchers 

India’s one serious attempt at a rulebook for AI training was not written with researchers in mind. For instance, in 2025, the Department for Promotion of Industry and Internal Trade formed a committee to examine where generative AI meets copyright. Its working paper, released that December, proposed a hybrid licensing framework: AI developers would get an automatic right to train on any lawfully accessed copyrighted work, rights holders could not refuse, and creators would be paid a share of revenue instead. A second part on AI outputs is still expected. None of it is law yet.

The framework treats the harm as financial. Creators lose money, so creators get paid. That works for a publisher or a musician. It works poorly for a scientist.

“DPIIT’s proposed compulsory royalty model doesn’t work here because money is not the only main harm for researchers,” said Mishi Choudhary of SFLC.in, a legal services organisation that brings together lawyers, policy analysts, students, and technologists to protect freedom in the digital world. Lose priority on a result and you lose the credit, the citation, and sometimes the career step that follows. 

Choudhary’s rule is a list of what should never go in: “unpublished manuscripts, proofs, source code, datasets, grant proposals, theses under embargo, patent sensitive material or confidential peer review material,” unless the terms clearly prohibit training use.

That last condition is where the legal exposure sits. “The real issue isn’t whether you typed your idea into an AI tool, it is what the tool does with it,” says Gunjan Paharia, managing partner at Zeus IP Advocates LLP.

Her advice to researchers and founders comes down to five steps:

• Protect Your Idea First: File a provisional patent application before sharing a patentable idea with an AI tool. 
• Read AI Tool’s Terms: Don’t rely only on what the product’s marketing says. Check whether your data can be used for training, whether humans can review it, and how long the company retains it. Where possible, use enterprise or API plans that explicitly prohibit training and data retention.
• Ask For Stronger Safeguards: If you are working with an AI provider, seek clear commitments on no training, no human review, data deletion timelines, breach notifications and applicable jurisdiction.
• Keep Sensitive Work Out Of AI Tools: Use AI for lower-risk tasks such as literature searches, translation, formatting and early drafts. Keep unpublished findings, core ideas, patent-sensitive material and confidential data offline.
• Keep A Clear Record Of Your Work: Maintain dated notebooks, version histories and other records that can establish when you developed an idea. Institutions should also have clear AI-use policies that researchers formally acknowledge.

“Work that never should have gone outside your four walls cannot be fixed with careful process later,” Paharia said.

According to Shukla of IIIT Hyderabad, “There is no technological method for stopping them from looking at your uploaded information.”

For researchers, the lesson from the Buckmaster episode is that AI comes with questions around confidentiality, ownership and credit. Indian institutions are beginning to address these concerns, but formal policies and legal protections are still evolving. Until clearer rules emerge, researchers are largely left to protect their own work by limiting what they share. 

At a time when AI is being touted as man’s closest companion for everything, including research, the real question is: do we really understand the stakes here?

The post Researchers Make AI Their Confidant, But Can’t Afford To Let Their Guard Down appeared first on Inc42 Media.