# Northeastern students find AI isn’t a cure all for drug discovery > Source: > Published: 2026-08-26 17:52:29+00:00 # Northeastern students find that AI isn’t the cure-all in drug discovery Northeastern University capstone students find AI still isn’t as capable as human researchers for drug discovery and development. There’s been a lot of talk about artificial intelligence’s potential to transform drug research and discovery — and companies are spending billions in pursuit of that dream. In 2025, the AI drug discovery market was valued at 2.3 billion and is projected to grow to $13.8 billion by 2033, according to market analysis firm [Grand View Research](https://www.grandviewresearch.com/industry-analysis/artificial-intelligence-drug-discovery-market). Demis Hassabis, chief scientist at the technology company Alphabet, has even made claims [that AI could help scientists discover](https://www.cbsnews.com/news/artificial-intelligence-google-deepmind-ceo-demis-hassabis-60-minutes-transcript/) a cure for most if not all human diseases within the 10 years. But to what extent is this all backed by real science and technological advancement – and not just marketing hype? Northeastern graduate students in Anton Sinitskiy’s Applied AI capstone class are setting out to find the answer. Over the past academic year, three cohorts of students in the capstone course have tested a few of the most popular open source AI frameworks used for complex medical and biomedical research – including GPT Researcher and Agent Laboratory – to see if they could get to the bottom of the AI industry’s claims. And time and time again, the students have come to a similar conclusion — while AI can certainly be useful in some applications, there’s still a ways to it taking center stage in the drug discovery process, said Anton Sinitskiy, a teaching professor in the Applied AI program in the College of Professional Studies. Today’s models continue to make dangerous factual mistakes, are incapable of drafting well-sourced scientific reports on their own, and need constant hand holding. That’s the takeaway from three scientific reports produced by students in the capstone class. “If you just listen to what is written in the media, on scientific papers, and marketing materials, it sounds like AI is capable of solving all problems, including drug design,” Sinitskiy said. “But unfortunately the truth is very far from these overhyped statements.” The capstone class is one of the final courses students take in their graduate studies and is designed to challenge them to put into practice everything they’ve learned throughout their degree in a real-world context. In the students’ [first report](https://www.biorxiv.org/content/10.64898/2026.01.05.697809v1.abstract), which was shared in January 2026, they analyzed eight open source models used for complex research to see if they were capable of reproducing human made algorithms designed for drug discovery. One of the algorithms the students had the AI agents try to replicate was developed at Northeastern by a student working under Lei Xie, a Northeastern professor of pharmaceutical and biomedical sciences. The human-made Northeastern model is useful in predicting properties of molecules when they interact during the drug development process, Sinitskiy said. ## Editor’s Picks All the open source models failed to even draft scientific reports and were incapable of reproducing the algorithm developed by Lei and his students, Sinitskiy explained. Xie said the findings put into perspective “the scope” of these technologies. While AI can be useful in some contexts, it cannot “replace human intelligence,” he said. For [the second report](https://www.biorxiv.org/content/10.64898/2026.06.24.734302v1.abstract), shared on June 27, the students performed similar tasks again using “five more advanced AI frameworks.” But this time, the models were asked to replicate the findings of a drug discovery report developed by the Swiss pharmaceutical company Novartis. The Novartis study was chosen for the second report because it addressed the same topic that was tackled by Lei’s algorithm and was published by “a big pharma company known for its high quality research,” Sinitskiy said. While the students found that these more advanced models were slightly more capable of drafting original hypotheses and drafting full reports, they lacked depth and were rife with errors. For Niventhini Senthilselvan, who completed Sinitskiy’s capstone course in the winter and is one of the student co-authors of the report, the experience highlighted the importance of not taking AI chatbot’s answers at face value. “It taught me to look beyond simply getting an answer from AI and ask deeper questions: How reliable is it? Why did the model produce it? And how can we improve it?,” she said. It’s a particular valuable skill for the recent Northeastern graduate who works as an AI engineer in the telecommunications industry. “As AI becomes increasingly integrated into real-world decision-making, understanding not just an answer, but its reliability and uncertainty, is equally important,” she said. For t[he most recent report](https://www.biorxiv.org/content/10.64898/2026.06.23.734132v1.abstract) published on June 29, students tested the models’ capacity to independently review scientific papers, large data sets, and formulas for analysis. Once again, they found that the models largely failed, producing inaccurate and misleading results. Notably, as part of this study, the researchers also conducted separate experiments where they worked collaboratively with the models to produce much more accurate and usable data, explained Jude Dcunha, a student co-author who completed Sinitskiy’s capstone course this spring. In simple cases, AI can get you 95% of the way on its own, Dcunha noted. But in more complex cases “it can fail miserably, and you need to hold its hand.” Sinitskiy noted that while students analyzed a handful of specific AI models, the goal of the research was to serve as an overview of “the current stage of affairs” rather than a deep breakdown of how specific models are totally inadequate. “It’s not an issue with one specific AI framework, but with the state of the field,” he said. But it wasn’t all bad news. In some respects, AI can be a useful tool in completing repetitive tasks, statistical support and stress testing ideas, Sinitskiy said the research revealed. “The point is AI can do really important things, but you need to know what kind of things it can do, and you need to run it properly,” he said. “That’s what I’m teaching students in my courses to do.”