https://www.cas.cn/cm/202610/t20261012_5122081.shtml
As AI becomes more and more deeply involved in scientific research, and as scientific researchers become more and more inseparable from AI, a discussion of whether human scientific research capabilities will be degraded as a result has become lively.
Some people are worried that if AI is asked about everything, its independent thinking and value judgment capabilities will decline; some believe that not only will it not degrade, but human-machine collaboration will promote the evolution of scientific research capabilities; some point out that what is changing is the ability structure and scientific research paradigm, not the scientific research ability itself.
This issue invites front-line scientific researchers in different fields to express their own opinions. Their views may be different, but they all point to the same question: Which problems should be left to AI, and which abilities must be left to humans?
- Ye Tao (Professor of School of Mechanical and Electrical Engineering, China University of Mining and Technology [Beijing]):
If all the power of thinking is given to AI, its abilities will inevitably degrade. As the capabilities of large models continue to increase, AI is gradually turning from an auxiliary tool into scientific research infrastructure, improving scientific research efficiency, but it also brings a question: Will scientific researchers gradually weaken their ability to think independently, continuously and deeply due to their long-term dependence on AI, leading to inertia in thinking and degradation of scientific research capabilities?
This risk deserves vigilance. Taking the multi-sensor fusion perception research our research group is engaged in as an example, when it is discovered that there is a time error between the camera and the IMU (Inertial Measurement Unit), we should continue to ask: Where does the error come from? What impact will synchronization errors of 1 millisecond, 3 milliseconds and tens of microseconds have respectively? Should the error be resolved through hardware synchronization, or can it be compensated for through backend time offset estimation? These continuous “whys” constitute the real chain of scientific research thinking. With AI, it is easy to form another working mode: ask AI directly, and AI will give answers or suggestions. Researchers may have skipped the most critical layer: why this method can solve the problem, and under what circumstances it cannot solve the problem.
Whether AI leads to inertia in thinking, the dividing line lies not in whether to use AI, but in what researchers give to AI. Preliminary literature screening, code assistance, data sorting and other procedural and repetitive tasks can fully utilize AI under the premise of manual review and standardized use; candidate algorithms and even candidate research plans can also be generated with the help of AI. But there are some things that cannot be completely handed over: why this problem is studied, what is the real scientific problem; why this technical route is chosen; what can the experiment prove; why the results are credible; and what should be believed in the end. These are “cognitive responsibilities” that scientific researchers must bear.
From this point of view, AI in the future will not necessarily make scientific researchers stronger or weaker as a whole, but may lead to further differentiation of scientific research capabilities. Some scientific researchers leave problems, plans, codes, and result interpretations to AI. The short-term scientific research efficiency may be very high, but the independent research ability may gradually decline. Other researchers take different approaches: they define the problem themselves and let AI provide candidate solutions; they design their own verification standards and let AI help complete the experiment; and finally they ask AI to attack their own conclusions. Using AI in this way will not necessarily reduce the thinking of scientific researchers, but may actually deepen it.
What also needs to be vigilant is to actively hand over the power of thinking to AI. The most important ability of scientific researchers is to judge what questions are worth asking, what answers are questionable, and what results can withstand testing. If researchers outsource problem definition, causal reasoning, and academic judgment to AI, there will inevitably be a risk of degradation in their ability to think independently in the long run.
Let AI replace mechanical labor rather than replace cognitive responsibilities; let AI expand thinking rather than end thinking – if we adhere to this principle, AI will not only not weaken scientific research capabilities, but may liberate scientific researchers from a large amount of repetitive work, giving them more time to complete the work that truly belongs to scientists: asking questions, understanding mechanisms, questioning conclusions, and discovering unknowns.
2. Chen Xiaole (Vice Dean of School of Pharmacy, Fujian Medical University):
AI is evolving, and the ability to control AI will also evolve.
Will AI lead to the degradation of scientific research capabilities? In my opinion, the answer is no.
When I first took charge of the project independently, the discovery of natural peptides mainly relied on the traditional path of “activity tracking + step-by-step separation”. The separation and purification process must be planned at the beginning of every research, from crude extract to obtaining an active peptide with a purity standard and a clear structure, which can take a few weeks or a few months. When it comes to structural identification, the peptide fragment ions generated by high-resolution tandem mass spectrometry need to be analyzed one by one and compared with the database one by one. A complex natural peptide component often corresponds to hundreds of candidate sequences, and the splicing logic of each sequence must be repeatedly deduced and confirmed.
What really made me feel the profound changes in the research paradigm was the integration of AI-assisted virtual screening into the entire process of discovering natural active peptides. I was deeply impressed by the Chinese research team’s research on protease hydrolysates of marine products: there are thousands of peptide sequences that can be identified in the samples. If we follow the traditional idea of “one by one isolation and purification – high-resolution tandem mass spectrometry identification – activity verification”, the sample consumption, time cost and manpower investment will be unbearable. The team established a new workflow of “activity tracking separation – high-resolution tandem mass spectrometry targeted identification – AI-assisted virtual screening”, and finally selected 20 peptides from 96 high-confidence candidate structures for experimental verification. Such efficiency is almost unimaginable under traditional research models. For my own research, the value of such tools goes far beyond speeding up: it allows us to predict which components and sequences are more worthy of investing resources in the early stages of separation and purification, avoiding spending a lot of energy on components that are ultimately proven to be inactive, and achieving precise allocation of research resources.
In literature research and peptide sequence optimization, the value of AI is also outstanding. In the past, optimizing the structure and activity of a lead peptide required systematically combing through a large number of literature reports on similar peptides, sorting out activity data corresponding to different site mutations and different chemical modifications one by one, and then relying on research experience to determine the direction of modification. This is time-consuming and prone to information omissions. Today, I will use the peptide AI design tool to conduct preliminary research: as a closed-loop peptide design framework, it modularizes functions such as sequence generation, physicochemical property prediction, and single-residue mutation generation. Each prediction step has a clear algorithm basis, and the results can be reproduced and the process can be traced back. For scientific researchers, this traceability tool is crucial – our scientific research judgments must not be based on “black box” conclusions, and every decision must be supported by clear scientific logic.
AI does not replace human scientific research capabilities, nor will it cause degradation of capabilities; it promotes the iterative upgrade of scientific research capabilities, freeing researchers from repetitive and ergodic labor and investing more energy in core links that truly require scientific wisdom. Tools are evolving, and so are the definitions of scientific research capabilities. This is not a regression, but an inevitable adaptation and active evolution for scientific researchers in the new era.
3. Luo Tao (Associate Professor, School of Water Conservancy and Construction Engineering, Northwest A&F University):
AI wants to choose topics, but people are less and less able to “ask questions”
Topic selection is the starting point and premise of scientific research. Without scientific questions, scientific research will have no goals. All experiments and analyzes are purposeless data collection and cannot be counted as research. Good questions also determine the value, direction and upper limit of final results of research. Continuously raising new questions and constantly challenging existing cognitions are the fundamental driving force for subject iteration and theoretical updating.
With the development of artificial intelligence, I have noticed an increasingly common phenomenon in the process of supervising graduate students: before starting a topic, many students are accustomed to asking large models to sort out the current research status, recommend topic selection directions, and generate “innovation points.” The framework provided by AI has a complete structure and thoughtful wording, which is impressive at first glance. However, after careful reading, it is found that most of these “innovative points” are permutations and combinations of existing work, lacking in-depth understanding of actual engineering problems. Taking the rock-fill concrete that our team has studied in recent years as an example, the truly valuable problems often come from the project site, such as an abnormal crack, a set of unreasonable monitoring data, etc. These “something wrong” signals can only be caught by repeated observation and thinking on the dam and in front of the test bench.
Questions never emerge automatically from data. I once undertook a scientific research mission to investigate freeze-thaw disasters in the China-Nepal transportation corridor. During field surveys in cold plateau areas, I found that the damage to concrete caused by the freeze-thaw cycle is much more complex than what was shown in indoor tests. The temperature gradient, moisture migration, and physical phase changes in the real environment cannot be completely reproduced in any single model. This perception of “on-site complexity” is precisely the prerequisite for asking good questions. AI cannot replace the time when you see a crack in a dam in the cold wind and ask yourself: Why here?
This does not deny the value of AI empowering scientific research. AI can become a powerful hypothesis generator and efficiency accelerator, but it cannot replace the core role of researchers in problem condensation, mechanism elucidation, value judgment, and the formation of final scientific conclusions. In other words, AI is good at finding correlations in the known, but it does not have the ability to think creatively by combining domain experience and theoretical frameworks. If researchers leave the most critical cognitive link of “asking questions” to algorithms, scientific research may be reduced to “data-driven knowledge transfer.”
The deeper concern is that the degradation of problem awareness is gradual and hidden. When reviews can be generated with one click, and when innovation points can be automatically extracted, do students still need to rummage through piles of old papers, ask questions when experiments fail, or run around in engineering sites with harsh environments? Over time, scientific researchers may become better and better at “answering questions” but less and less good at “asking questions”.
In the AI era, what is the “source of living water” for scientific research? It’s not computing power, it’s not data, it’s people’s curiosity and questioning about the unknown world. What truly defines a scientist is not what tools he is good at using, but what questions he asks.
4. Chen Xi (doctoral supervisor at the University of Chinese Academy of Sciences):
Human-machine collision, “asking more” becomes a catalyst for innovation
In the AI era, knowledge is at your fingertips. But the most precious thing in scientific research is often not a ready-made answer, but a voice that keeps asking questions and forcing you to think outside the box. In recent years, large models have allowed me to find such a “scientific research helper” that I can talk to at any time. It does not solve the problem for me, but helps me see the blind spots in my thinking through repeated questioning and counter-questioning, and sometimes new scientific research points emerge.
I study microscopic mapping of brain connections. The brain of a larval zebrafish, which is less than 1 cubic millimeter, has 100,000 neurons. The whole brain needs to be cut into 22,887 30-nanometer-thick slices, photographed layer by layer, and then reconstructed into three-dimensional data using AI splicing. The original data volume is as high as 200 trillion bytes. In the past, we were stuck in the three-dimensional registration of millions of images. Errors accumulated layer by layer and the reconstruction results were seriously distorted. There were no big models at that time, and the turning point came from repeated discussions between the team and peers. A collaborator asked rhetorically: “Have you considered changing the layer-by-layer serial registration to global optimization?” This rhetorical question forced me to think outside the box, and ultimately significantly improved the reconstruction accuracy. Looking back today, this experience of “being asked back” is what I valued most when I later used large models.
Now, I often think of the large model as a “questioner” that I can talk to at any time. For example, if the registration accuracy cannot be improved, I asked it what ideas it had. It lists optical flow, feature matching, non-rigid registration, graph optimization and other solutions. If I am only satisfied with the first solution, it is easy to be led into ideas that seem lively but are actually unproven, so I will continue to ask. Once I asked it: What if the error is not random, but a systematic bias? Following this hypothesis, it helped me list several possible sources – sectioning order, imaging distortion, staining differences, mechanical drift. These must be checked one by one, but this rhetorical question really reminded me that I cannot just focus on the registration algorithm itself. As I keep asking, new ideas sometimes pop up: Can local registration be put into the global consistency framework? Can cross-layer geometric constraints be used to suppress error accumulation? Can we use simulated degradation data to verify the algorithm first, and then use real data? These ideas may not all be correct, but they may indeed become new scientific research points.
Of course, large models may fabricate documents, omit conditions, and give unverified inferences. Any idea inspired by AI must eventually be tested back to experiments and data. I encourage students to “ask more” when talking to AI: don’t be satisfied with the first solution, but ask about assumptions, counterexamples, and boundary conditions. The real difficulty is never “knowing an answer”, but constantly confirming: Have we asked the right questions? Have you fallen into a local optimum? Do you regard accidental success as a universal rule?
AI can expand knowledge access and improve exploration efficiency, but logical thinking must be rooted in human training and practice. AI is not the answer, it is the “scientific assistant” that keeps asking me questions. It forced me to change “I think” into “how to prove” and “almost” into “test again.” Only by allowing the collision of inspiration in human-machine dialogue to become a catalyst for innovation can technological progress be truly transformed into scientific creativity.
5. Miao Dongjing (researcher at the Institute of Geometric Quantity, Chinese Academy of Metrology):
Being accustomed to trusting AI, the ability to review and judge is eroded
When the output of AI becomes more and more like the correct answer, will people still review and judge it, or have they become accustomed to accepting it without thinking? At the end of 2025, there were media reports that a doctoral student from a university in Hong Kong published a paper in an academic journal, and a considerable part of the dozens of references were AI fictitious. The paper was published after passing two academic reviews, but no one found any abnormalities. The Lancet study shows that the rate of false citations of biomedical papers has increased significantly in the past two years. When AI fictional document formats and author signatures are difficult to identify even through institutionalized academic review, the degradation of review judgment is evolving from individual negligence to systemic risk.
The failure of review and judgment is often due to the fact that AI’s ability to imitate is so high that conventional review is difficult to penetrate. In September 2026, 25 Fields Medal winners including Tao Zhexuan and Deng Yu jointly issued a statement “The Serious Misalignment of Artificial Intelligence in Mathematics”, which pointed to this dilemma: AI is mass-producing true and false conclusions at an increasing speed, and the burden of identifying authenticity, correcting errors, and integrating insights falls on humans. There is a metaphor in the statement: famous mathematical problems are like landmarks and beacons. Solving the problem is the sign of discovering new insights, not the end. AI directly gives the answer, eliminating the need to accumulate opportunities along the way. The answers are coming in like a tidal wave. If the embankment of review and analysis is not increased accordingly, false conclusions will flow unimpeded into the knowledge production chain and mislead subsequent judgments and decisions.
The foundation of critical thinking is also being eroded. Artificial intelligence expert Andrew Ng has been promoting AI, but he has not avoided this issue: the so-called judgment is essentially that humans have context that AI does not have – accumulated through long-term hands-on experience, so they can identify things that seem correct but actually do not work. Once you get used to directly accepting the results output by AI and no longer go through the complete path of data verification and logical deduction in person, the context will shrink, and the parties may not be aware of it.
The first thing that is affected by the shrinkage of context is not the ability to correct errors, but the perception of where there may be errors. There are standard procedures to follow for error correction, and perception relies on intuition accumulated over a long period of time. In the industrial field, the process of judging whether the results are reliable after getting them is called calibration – it is up to humans to determine which link the error lies in. If you are used to trusting AI, deficiencies will easily occur in the calibration process.