About the Scholars’ Council for AI
A faculty-centered academic research organization that studies the conditions of judgment and responsibility needed when AI takes part in people’s lives and in the work of organizations.
Why this research
Five questions where the research began
When work is handed to AI, what else is handed over with it?
Anyone would welcome being able to hand an AI the sorting of a backlog of email and the replies, and so lighten the day’s work. But if that AI, at someone else’s request, hands over my email as well, then the work I entrusted to it turns out to have included a judgment about what may be shown to whom. The more we use technology that does our work for us, the more we also hand over authority that affects our information, our relationships and our choices.
Read more — A case found in an experiment
“Agents of Chaos,” released in February 2026, is an experimental study that observed the behavior of AI agents that carry out work using tools. The researchers tested six agents that handle email, files and the like for about two weeks. Among the cases was one in which an agent provided 124 email records at the request of a researcher who was not its owner. A case was also reported in which an agent refused a request to give out personal information directly but complied with a request to forward an email containing the same content. These are failures found in a research environment and do not represent the actual incident rate of all AI. They do, however, reveal the problem that the ability to carry out work well and the ability to distinguish whose requests to accept, and how far, should be verified together.
Is it enough for a person to check at the end?
If the employee who approves work processed by AI has neither the time to look at the evidence nor the authority to stop a wrong action, that check may amount to no more than a signature. Holding someone responsible after a problem arises is not enough; it should be possible to check what information a judgment was based on, and there should be people who re-examine unusual results and measures that are actually carried out. Only when these conditions are in place does saying that AI can be supervised mean something to people.
Read more — A case showing that monitoring itself needs verification
METR (Model Evaluation and Threat Research), a U.S. nonprofit research organization, studies what AI can do on its own and whether people can detect and stop dangerous behavior. A report it released on 22 October 2026 stated that, after examining the internal monitoring system of the AI developer OpenAI in July–August of that year, it found gaps in monitoring and possibilities of circumvention. What was evaluated was the system as it stood at the time; OpenAI explained that it later fixed most of the vulnerabilities, and METR did not verify the fixes. This is not a report that AI actually exploited all of them, but a case showing that monitoring itself needs verification.
Debate also continues over whether AI might one day slip out of human control. The International AI Safety Report 2026, in which experts from many countries brought together research on the capabilities and risks of AI, summarized, when it was released in February of that year, that experts differ widely on the likelihood and severity of loss of control. Recent failures alone cannot foretell the end of humanity, but neither can the research needed now be put off until there is agreement on the long-term outlook. Situations in which personal information is passed on wrongly, the reasons for important decisions cannot be known, and there is no way to fix an error even once it is found are problems to be solved in their own right.
Why should several disciplines work together?
Several disciplines contribute together to the research problems of the Scholars’ Council for AI, and which disciplines take part is decided by what explanations and methods a particular research problem needs. This need becomes clearer when several AIs work together. Which roles to divide, which information to use in common, and how to have differing judgments reviewed need to be studied in their own right.
Read more — Many disciplines and many AIs
The Scholars’ Council for AI seeks to connect the knowledge that academic disciplines have built up to the work of designing and using AI, and to verify whether that knowledge actually helps people in real-world settings. Its starting point is the idea that people should be able to understand, choose and correct their own work even while receiving help from AI. This requires the questions of several disciplines together. Scholarship that has studied knowledge and learning, decision-making and organizations, and human behavior provides ways to explain and test the conditions under which wrong judgments arise.
In a study published in August 2026 in the Proceedings of the National Academy of Sciences (PNAS), a multidisciplinary journal published by the U.S. National Academy of Sciences, even in a simple cooperative task of AIs settling on a name, individual AIs’ preferences could be amplified or reversed depending on their interaction and the size of the group, and collective biases absent at the individual level could appear. It is a simple experiment, but it is evidence that checking the performance of a single AI alone makes it hard to fully explain the results of the whole that works together.
What remains with people when they receive help from AI?
What the Scholars’ Council for AI seeks to protect is people’s safety and, along with it, their capacity to learn and judge for themselves. Even when AI helps with difficult work, whether understanding and experience remain with the person who used it should be examined separately.
Read more — Learning, and the conditions for reaching help
In a mathematics experiment at a high school in Türkiye, published in PNAS in 2025, students who received ordinary AI help did better on practice problems but scored lower than students who had received no help on a test taken without AI. By contrast, with an AI built to give teacher-designed hints instead of answers, this disadvantage was greatly reduced. This means that what remains with people can differ depending on how AI is designed and used. Uses should be found that make today’s work easier while also building the capacity to understand tomorrow’s problems.
The conditions for reaching that help also differ from person to person. Simply opening a service does not give everyone the same opportunity, so the environment of use and the means of support should also be studied, so that people who are unfamiliar with it can understand explanations and ask for the help they need.
In the 2025 Digital Divide Survey [translated title, provisional] released in March 2026 by the National Information Society Agency of Korea, the level of digital access of four vulnerable groups — people with disabilities, low-income households, farmers and fishers, and older adults — was 96.6, with the general public set at 100, but their level of competence was 65.9. The figures do not measure the ability to handle AI alone, but they reveal the gap between having devices and actually being able to use them.
What should be checked in the field?
Research like this should be checked in the places where people live their lives. If AI is used at an airport, besides how much faster guidance became, it should be examined whether travelers who speak an unfamiliar language understood the content correctly, whether people who find devices hard to use received help, and whether staff could check the grounds and correct wrong guidance when it occurred. This is because even when the same technology is brought in, the changes that remain with people can differ.
Read more — What is sought as the outcomes of the research
The Scholars’ Council for AI seeks to make such differences the outcomes of its research: reducing the losses and inconvenience caused by wrong judgments, enabling people to demand explanations for and correct decisions that affected their own work, and letting the experience of receiving help remain as learning. Whether people and organizations lacking expertise and resources actually benefited is also checked. To this end, the aim is research that reviews scholarly evidence, applies it to the conditions of the field, and can openly examine even the reasons why results differed from expectations.
How the research proceeds
In what order does the research proceed?
From questions to change.
Difficulties people face
wrong guidance · judgments without explanation
weakened learning · unequal access to help
into questions
Scholarly questions and design
why does this happen?
which conditions to change and test?
conditions to test
Comparative experiments and field observation
what changed from the existing way?
effects · costs · failures · limits of application
checked in the field
Changes that remain in people’s lives
reducing harm · the right to judgment and accountability
the capacity to learn independently · opportunities to use
From changes that remain in people’s lives back to the difficulties people face — remaining problems and new questions
The change sought
The change the Scholars’ Council for AI seeks
Knowledge should reach those who need help, people who are unfamiliar with these tools should also gain chances to learn, and people should be able to demand an explanation and a correction when faced with a wrong decision. The change sought is for progress in AI to lead to these everyday possibilities.
How each goal is examined
Reducing harm
examined through the kinds of errors and their effects.
The right to judgment and accountability
examined through whether decisions were actually re-examined and corrected.
Human capability
examined through changes in learning and in independent performance.
Gaps in access and use
examined through the outcomes and burdens for each group.
Read more — When a goal is hard to apply directly
For a goal that is hard to apply directly to a project, the reason is stated. If it has not been confirmed what social change a model’s accuracy or cost savings led to, technical results and social outcomes are reported separately.
What the research rests on
Three things the research rests on
People. Evidence. Practice.
Where should AI research start?
AI research should start from the changes that actually occur for the people who entrust work to AI and for the people affected by it.
Why is scholarly evidence needed?
Quoting a sentence from a paper does not by itself make its application valid. Some studies dealt with people in particular jobs, and some results were observed only with particular methods of measurement. When the target of application changes, the way the results are interpreted should also be re-examined.
What should research change in the field?
The outcomes aimed for are changes in which people suffer less harm, actually take part in important decisions, learn in the course of receiving help, and reach the knowledge they need more easily.
Read more — The three foundations
Where should AI research start?
When a company hands customer inquiries to AI, the person who decides to adopt it, the staff who operate the system and the customers who receive the answers stand in different positions. Even if processing has become faster, if customers cannot find where to correct wrong guidance or staff spend more time checking errors, the result cannot be seen as the same success. The scope covered extends to what knowledge AI uses, how it judges and acts with that knowledge, and what results it produces in its relationships with people and with other AIs.
Why is scholarly evidence needed?
Scholarly intelligence refers to the combination of the content of scholarly knowledge and the judgment that handles that knowledge.
Suppose an AI reviewing purchase requests proposed cutting review steps, citing past research on work improvement. The redundant procedure in that research may have been a step of entering the same information repeatedly, while the review steps in the present work may each check a different kind of error. Applying the same improvement because the number of steps looks similar changes the meaning of the original research. What should be examined is not wording that looks similar on the surface but the relationship between information and function.
Iyad Rahwan, who leads the Center for Humans and Machines at the Max Planck Institute for Human Development in Germany, studies the social impact of AI and human–AI cooperation. “Machine behaviour,” published by Rahwan and colleagues in 2019, proposed that understanding the behavior AI shows in real environments requires using the perspectives of computer science and other sciences together. This approach brings together explanations of how a system was built and observations of how it behaves under real conditions. Building on this concern, the Scholars’ Council for AI seeks to connect scholarly concepts to observation and experiment.
When scholars take part, the original research can be interpreted more concretely. A researcher can explain the process that led to a conclusion, the difficulties of measurement, and follow-up questions not fully visible in the paper. At the same time, applying research to AI is a new claim and so should be verified separately. The original researcher’s agreement is an important record of review, and effects in a new environment are confirmed by experiments in that environment.
When scholarly review takes place, the meaning of failure also changes. When a system does not produce the expected effect, it can be examined separately whether this was an error of implementation, a mismatch in the conditions of application, or a limit of the original hypothesis. What the Scholars’ Council for AI seeks to build up is not only good results but a record of the evidence and judgments that make such distinctions possible.
What should research change in the field?
Suppose a traveler guided by an airport’s AI went to the wrong place. The operator may assess that it cut guidance time, but the traveler is left with lost time and anxiety. Staff may have to check the situation from the beginning without knowing what guidance was given. In such a case, the results research should check include, along with the accuracy of the answer, the process by which errors are found, the time taken to fix them, and the understanding and burden of the person who received the help. The human dignity and the right to judgment that the Scholars’ Council for AI pursues should be connected to concrete experiences like these.
To reduce wrong judgments, it should be possible to find out and check what is not known. Even when AI has found material, it should be checked whether the target of application and the effective date are right, and when evidence is lacking, it should be possible to request more information or to defer judgment. Cases in which deferring judgment makes someone wait needlessly or causes a legitimate request to be refused are also examined. The task of research is to check which responses help people under which conditions.
The process after an error is found is given the same weight. It is checked whether an understandable answer comes back when an explanation is requested, whether the person in charge of re-examination has the records and authority needed, and whether corrected results are reflected in the actual service. Whether people who received help from AI can do better the next time, and whether that help reaches those who need it most, are also subjects of research.
Having an inquiry desk is not enough in itself. People who have run into a problem should be able to know where to ask for help, and the operator should be able to follow through with measures to the end.
If staff are told to review all of AI’s results without being given the time and information to review them, the burden of work can instead grow. Research does not only count how many times a person intervened, but examines whether the needed judgment could be made at the important moment. Roles and authority are designed so that unnecessary checks are reduced and attention can be focused on decisions with large effects, and whether that design is effective is judged by comparing the experience of members in the field with actual results.
Experienced people may need support in comparing complex alternatives, and beginners may need support in learning the reasons that lead to a conclusion. Along with the speed at which work was finished, the ability to explain the evidence, the ability to find errors, and the ability to apply what was learned to new situations are examined. By observing the difference between working with AI and working without help, the conditions under which convenience leads to learning can be studied concretely.
Users who find unfamiliar terms hard to understand, users who find it hard to read a screen or handle a device, and small organizations that find it hard to keep specialist staff each need different support. The time and cost it takes them to finish a task, how well they understood explanations, and the process of asking a person for help are examined. Even if the overall average has improved, if a particular group faces greater difficulty, that difference should be brought out and its reasons examined. Studying these problems needs, along with computer science, disciplines that deal with human learning and judgment and with how organizations and society work.
In empirical validation, the expected change is stated first and comparison conditions to check that change are prepared. What difficulties the existing way had, what is to be changed, and whose experience and which results are to be observed are decided. Along with improvements, increased burdens, conditions under which there was no effect, and effects not yet confirmed are also kept on record.
The direction of the initial joint research is to choose field projects where such comparison is possible, and to build up scholarly review and experimental plans together with comparison results and ways of using them. Results obtained in the field are organized with their evidence and conditions so that other researchers can review them, while personal information and the conditions for using data are respected. Whether a study contributed to changes that remain with people should be explainable with evidence from the field, and the aim of improving people’s lives serves as the criterion of judgment throughout the whole process of choosing, designing and evaluating projects.