When AI Stops Answering Questions and Starts Making Discoveries

Preamble

In my previous Kerbside Consult article, “How Good Is Medical AI, Really?” (Article 47/2026, 10 October 2026), I examined how well AI performs on clinical tasks and why a benchmark score cannot tell us whether it can safely manage a patient. [1] Here, I want to explore a different question: can AI contribute genuinely new scientific knowledge?

AI is increasingly moving beyond answering questions to producing mathematical results, proposing scientific hypotheses and designing potential medicines. But generating a promising result and establishing a discovery are not the same thing.

In October 2026, OpenAI released a collection of mathematical results from an internal model. The announcement attracted attention, but also questions about originality, verification and scientific significance. [2]

Yet computational discovery is not entirely new. Pharmaceutical researchers have used molecular modelling and machine learning for decades. What is changing is the speed, sophistication and scale of AI—and the involvement of major technology companies. Are we witnessing a new era of discovery, or a faster way of generating possibilities that scientists must still prove?

1. AI discovery did not begin yesterday

It is tempting to associate AI with ChatGPT’s arrival in 2022. Its scientific applications, however, go back much further. Pharmaceutical researchers were already using molecular modelling, structure-based drug design and compound screening in the 1980s. Not all these computational methods were AI.

An early example came from Vertex Pharmaceuticals, founded in Cambridge, Massachusetts, in 1989 with an emphasis on structure-based drug design. In 1998, researchers published work using neural-network methods to distinguish drug-like from non-drug-like molecules. This illustrated how machine learning could assist medicinal chemistry long before modern generative AI. [3]

Over time, pharmaceutical and biotechnology companies expanded AI applications to target identification, toxicity prediction, biomarker discovery and compound design. Their objective was not simply to produce more molecules, but to identify promising candidates earlier and reduce unproductive experiments.

2. What happened in mathematics?

On 6 October 2026, OpenAI announced a broad release of mathematical results from an internal frontier model, with Lean formalisations of many of the proofs; Lean is a programming language that allows mathematical proofs to be checked by computer. [2]

A mathematical proof establishes a conclusion from specified assumptions and logical steps. Software such as Lean can check formalised steps, but correctness does not establish that a result is original or important. A large collection of results should not be equated with an equal number of independently established major discoveries.

Nevertheless, AI is moving beyond solving familiar problems and contributing results at the research frontier. If results can be generated faster than experts can examine them, verification may become a bottleneck.

3. From predicting proteins to designing medicines

AlphaFold demonstrated major advances in predicting protein structures. AlphaFold 3 extended this work to interactions involving proteins, DNA, RNA and other molecules. Such predictions can guide research, but are not themselves proof of a useful treatment. [4]

Rentosertib, developed by Insilico Medicine for idiopathic pulmonary fibrosis, provides an example further along the development pathway. AI approaches were used in target identification and candidate design. A randomised phase 2a trial reported in Nature Medicine in 2025 enrolled 71 patients; its primary focus was safety, while efficacy signals were preliminary and require further study. [5]

The details matter. The trial lasted 12 weeks, was conducted at sites in China, and had about 17–18 patients per arm; it was not powered to test efficacy. Lung function (forced vital capacity, FVC) was a secondary endpoint. At the highest dose (60 mg once daily), mean FVC rose by 98.4 mL (95% CI 10.9 to 185.9), compared with a fall of 20.3 mL (95% CI −116.1 to 75.6) on placebo. Treatment-related adverse events were more common with rentosertib than placebo, and liver injury was the most frequent reason for stopping treatment. Sixteen of the 71 patients withdrew before 12 weeks. The authors themselves called for larger, longer trials. [5]

AI can help identify promising candidates earlier, but it does not eliminate medicinal chemistry, toxicology, manufacturing development, clinical trials or regulatory review. The greater challenge is not producing more molecules: it is finding treatments that are safe, effective and clinically meaningful.

4. When technology companies join the laboratory

Pharmaceutical companies, biotechnology firms and academic researchers have worked with AI for years. Now major technology companies are also investing in the data and technologies needed for biological research.

On 7 October 2026, Biohub announced an expansion of its Virtual Biology Initiative with partners including Google DeepMind, Isomorphic Labs, Meta, the US Department of Energy and the National Institutes of Health. The approximately US$1.8 billion commitment encompasses funding, existing data resources, computation and measurement technology—not simply new cash investment. Biohub's announcement attributes US$300 million collectively to Google DeepMind, Isomorphic Labs and Meta; the remainder comes from Biohub itself, a US Department of Energy investment of more than US$500 million over five years, and existing NIH-coordinated datasets built through more than US$500 million of prior federal funding. [6]

The ambition is to create better predictive models of how cells respond to interventions, so researchers can prioritise experiments before entering the laboratory. But these models need extensive, reliable biological data; a virtual cell is not a human patient. This remains a research ambition, not a clinically validated substitute for experiments.

5. Can AI really make discoveries?

AI can recognise patterns, generate hypotheses, design molecules and sometimes propose research results. These are different levels of contribution. A new hypothesis is not necessarily a new scientific fact.

In mathematics, a proof may establish a conclusion within its assumptions. In medicine, predictions must survive laboratory experiments, biological complexity and ultimately clinical evaluation. A candidate that succeeds in a simulation may fail in patients.

Data quality, bias, reproducibility, authorship and accountability also matter. The real measure of progress is not how many hypotheses or manuscripts AI produces, but how many findings survive independent scrutiny and contribute to knowledge or patient care.

6. What does this mean for Malaysia?

Malaysia has spent decades developing clinical research infrastructure, ethics review, Good Clinical Practice training and industry partnerships. These provide a useful foundation for AI-enabled research.

We need not build the largest AI models ourselves. Opportunities include responsibly analysing Malaysian disease registries, identifying research questions relevant to diabetes and cardiovascular disease, improving trial feasibility and investigating infectious diseases. Such work should be grounded in Malaysian epidemiology, not merely imported datasets.

This will require reliable and representative local data, clear governance, patient confidentiality, research expertise and collaboration across universities, hospitals and industry. An AI-generated signal is a starting point for research—not evidence sufficient to change clinical practice.

7. The scientist is still in the laboratory

Will AI replace scientists? I doubt that is the most useful question. It may increasingly undertake literature review, complex analysis and some aspects of molecular design, freeing researchers to focus on important questions and experimental interpretation.

But scientists must still judge which questions matter, whether assumptions are reasonable and whether a technically impressive result is biologically plausible or clinically relevant. Training must preserve the expertise needed to challenge AI, not merely operate it.

8. Conclusion

AI is increasingly contributing mathematical results, biological hypotheses and candidate medicines. Its roots in pharmaceutical research are decades old; today’s systems bring unprecedented speed and scale.

For medicine, success will be measured by independently validated discoveries that improve prevention, diagnosis, treatment and patient outcomes.

AI may help us discover what we did not know. Science must still establish whether what we have discovered is true.

Sources and further reading

1. Vadivale M. How Good Is Medical AI, Really? Kerbside Consult, Article 47/2026 (10 October 2026).
2. OpenAI. Sharing AI Progress in Mathematics (6 October 2026).
3. Ajay, Walters WP, Murcko MA. Can We Learn to Distinguish Between Drug-Like and Nondrug-Like Molecules? Journal of Medicinal Chemistry (1998).
4. Abramson et al. Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3. Nature (2024).
5. A Generative AI-Discovered TNIK Inhibitor for Idiopathic Pulmonary Fibrosis: A Randomized Phase 2a Trial. Nature Medicine (2025).
6. Biohub. International Cross-Sector Collaboration for AI Models of Biology (7 October 2026).