AI Ushers in a New Era of Discovery: The Paradigm Shift from Tool to Scientific Partner

AI is transforming from an auxiliary tool into a co-participant in scientific discovery. "We are entering a new era of discovery." This brief tweet points not merely to another model upgrade, but to a wholesale shift in the scientific research paradigm—AI is beginning to participate in proposing hypotheses, designing experiments, interpreting data, and even actively seeking research paths that humans have never envisioned.

From Accelerating Computation to Participating in Discovery

Over the past decade, AI's role in scientific research has primarily been that of an "accelerator": faster classification, regression, simulation, and search. AlphaFold predicts protein structures, AI assists in materials screening, and machine learning accelerates particle simulations—all of these place AI within problem frameworks defined by humans.

The key to the new paradigm lies in the shift of roles. AI is no longer just answering "how can this problem be computed faster," but is beginning to participate in "what questions are worth asking." When models can integrate literature, experimental data, and simulation results across modalities, they expose correlations within human research blind spots. For example, in biomedicine, AI can already propose new regulatory hypotheses from seemingly unrelated gene expression patterns; in materials science, generative models are beginning to output structures that human chemists had not prioritized. The center of gravity of discovery is shifting from verifying the known to exploring the unknown.

AlphaFold is a protein structure prediction system developed by DeepMind. AlphaFold2, released in 2020, raised the accuracy of protein three-dimensional structure prediction to near-experimental levels in the CASP14 competition, and is regarded as a landmark event in AI-driven scientific discovery. Its core innovation lies in modeling the mapping from protein sequence to structure as an end-to-end deep learning problem, and introducing attention mechanisms to handle long-range interactions between residues. The significance of AlphaFold lies not only in solving a problem that had remained unresolved for fifty years, but also in demonstrating that AI can achieve substantive progress on fundamental scientific problems that humans have long failed to crack, thereby providing a real-world anchor for subsequent discussions of "AI participating in discovery."

Three Technological Paths Currently Unfolding

Autonomous experimental loops. Robotic laboratories combined with AI planners form a closed loop of "hypothesis—experiment—feedback—correction." Systems can complete hundreds of condition screenings overnight and hand the most promising results back to humans the next day. This increases experimental throughput and also enables AI to obtain correction signals directly from the physical world.

Multimodal scientific foundation models. Single-modality models are being replaced by foundation models capable of simultaneously processing sequences, structures, images, text, and numerical tables. Such models can transfer representations across proteins, molecules, and materials, migrating the logic of discovery from one field to another.

Reasoning enhancement and hypothesis generation. Chain-of-thought reasoning, tool calling, and retrieval augmentation enable models to construct testable scientific arguments rather than merely outputting plausible-sounding text. This makes AI-generated content easier to verify or falsify through experiments.

The common thread among these three paths is that AI's outputs are increasingly approaching "actionable scientific claims" rather than just "useful information."

Scientific Foundation Models refer to general-purpose models pre-trained on large-scale, multimodal scientific data that can be transferred to a variety of downstream scientific tasks. Unlike general-purpose large language models, they typically need to handle heterogeneous data such as protein sequences, molecular graphs, crystal structures, microscopic images, and numerical tables, while maintaining cross-modal semantic alignment. For example, a model pre-trained on protein structure data can be used, after minimal fine-tuning, to predict molecular binding affinity or material stability. The value of such models lies in breaking the traditional "one task, one model" paradigm, enabling knowledge to transfer across different scientific domains, thereby accelerating hypothesis generation and experimental design.

What This Means for Researchers

The core skills of researchers are being redefined. In the past, asking good questions, designing rigorous experiments, and interpreting complex data were a trinity of capabilities; now, AI can take on a considerable portion of the execution-level work, but the demands on researchers' judgment are actually higher.

First, question taste becomes more critical than technical operation. When AI can rapidly generate large numbers of hypotheses, screening out the questions truly worth investing experiments in becomes a scarce capability. Second, the ability to verify and doubt becomes more important. Models can confidently output erroneous conclusions, and researchers need to design experiments that can distinguish "correlation" from "causality." Third, cross-domain translation ability appreciates in value. Those who can translate biological problems into AI-processable representations and then translate model outputs back into domain language will become key nodes in the new paradigm.

This does not mean researchers are being replaced. More accurately, repetitive labor is further compressed, and creative judgment is pushed to the forefront.

Retrieval-Augmented Generation (RAG) is a technical architecture that enables language models to first retrieve relevant documents from external knowledge bases before generating answers, then reason and output based on the retrieval results. In scientific research scenarios, RAG enables models to cite the latest papers, experimental records, or database entries, rather than relying solely on parametric knowledge fixed during training. This significantly reduces the risk of model "hallucination" and makes generated hypotheses or arguments easier to trace and verify. For scientific writing and hypothesis generation that require rigorous citations, RAG has become a key bridge connecting model capabilities with domain knowledge.

The Real Boundaries of the New Era

The phrase "a new era of discovery" easily conjures images of fully automated science. Reality is far more complex. Currently, AI-generated hypotheses still depend heavily on the distribution of training data, and reliability drops markedly when stepping beyond known ranges. The degree of automation in experimental loops is limited by hardware, and most laboratories still rely primarily on manual operations. Insufficient interpretability also makes many fields cautious about AI-led discovery.

A more likely picture is this: AI becomes a "high-bandwidth collaborator" for scientists. It handles search, combination, preliminary verification, and literature integration, while humans set direction, judge value, and bear ethical and responsible burdens. The pace of discovery will accelerate, but the meaning of discovery still needs to be defined by humans.

The value of this tweet may lie precisely in shifting attention from "what tasks can AI do" to "what kinds of discovery can AI participate in." When tools begin to ask questions, the scientific methodology also needs to be updated in tandem.

Out-of-Distribution Reliability refers to the fact that a model's predictive performance often drops significantly when encountering inputs that did not appear or rarely appeared in the training data. In scientific discovery scenarios, this means AI may give reliable predictions for proteins, molecules, or materials within known ranges, but once it explores entirely new chemical spaces or biological mechanisms, its outputs may be unreliable or even misleading. This limitation stems from the fact that models are essentially statistical inductions based on historical data, rather than a genuine understanding of physical laws. Therefore, current AI is better suited as a "hypothesis generator" than a "conclusion arbiter," and its outputs must undergo experimental or theoretical verification.

Frequently Asked Questions

Can AI truly make scientific discoveries independently? Currently, a more accurate statement is that AI can participate in the discovery process. It can generate hypotheses, design experimental plans, and analyze results, but independently completing the full chain from problem formulation to theoretical explanation remains unrealistic.

What practical impact does this have on ordinary researchers? The most direct impact is improved efficiency in literature review, hypothesis screening, and experimental design. Researchers need to learn to collaborate with AI, rather than using it as a search engine.

Which fields will feel the change first? Data-intensive fields where experiments can be automated, such as protein engineering, materials screening, and drug discovery, will progress faster. Fields that rely on long-term field observation or highly intuitive judgment will change more slowly.

One-line Summary

AI is transforming from a research accelerator into a participant in discovery, but human-AI collaboration remains the real boundary.

Paragraph Summary

The article points out that AI's role in scientific research is shifting from accelerating computation to participating in discovery, beginning to propose hypotheses, design experiments, and interpret data. Three key technological paths—autonomous experimental loops, multimodal scientific foundation models, and reasoning enhancement—are making AI outputs increasingly approach actionable scientific claims. This requires researchers to shift their core competitiveness from technical operation to question taste, verification ability, and cross-domain translation ability. However, current AI discovery is still limited by declining out-of-distribution reliability, insufficient hardware automation, and lack of interpretability. A more realistic picture is AI becoming a high-bandwidth collaborator for scientists. Scientific methodology needs to be updated in tandem as AI participates in discovery.