Who Answers for AI? Epistemic Visibility, Responsibility, and Error Repair in Collaborative Learning

In the modern classroom and the modern workplace, groups often solve problems together by pooling their knowledge. This process relies on a delicate social agreement: when someone offers an idea, the group must know where it came from, how it was formed, and who is responsible for checking its accuracy. This is the foundation of collaborative learning, where the goal is not just to share information, but to build a shared understanding that no single person could create alone. However, a new layer has been added to this dynamic: artificial intelligence. When a student or employee uses an AI tool to generate an answer, they face a critical choice. They can bring the idea back to the group as a simple suggestion, hiding the fact that a machine produced it. Or, they can reveal the entire history of that idea, showing the questions asked, the raw answers generated, and the edits made. This distinction creates a gap between what the group sees and what actually happened. The question is not whether AI helps, but how the visibility of its contribution changes the way people trust, argue, and fix mistakes when working together.

A researcher at University College London set out to explore this gap by observing how graduate students worked in groups over an eight-week period. They did not simply ask students what they thought about AI; they watched them work in real time. The researcher organized forty-eight students into twelve groups and gave them a complex task: to design an educational intervention to fix inequities in science and math education. To see how visibility mattered, the students rotated through three different ways of using AI. In the first scenario, a student could use AI privately. They could type a question, get an answer, edit it, and paste the final result into the group document without anyone else seeing the process. In the second scenario, the student still used their own private AI tool, but a special panel on the screen showed the rest of the group exactly what was typed, what the machine said, and how the student changed it. In the third scenario, the whole group used a single AI tool together, where everyone could see the questions and answers as they happened.

The researcher recorded every word the students spoke, every keystroke they made, and every change they made to their documents. They also introduced specific moments where the AI was programmed to give a wrong or misleading answer, just to see how the group would react. They wanted to know if the group would catch the mistake, who would feel responsible for fixing it, and whether the final product would be better if everyone could see the AI's work.

The study found that when the AI's work was hidden, the group dynamic shifted in a way that made learning less effective. When a student used the private, opaque AI, they often became the sole gatekeeper of that information. They would bring a polished idea to the group, and because the others could not see the machine's original output or the questions that led to it, they treated the idea as if it came directly from the student. This created a situation where the student who controlled the AI held all the power to decide what the group knew. If the AI made a mistake, the other students often did not realize they needed to check the work. They assumed the student had already verified it, or they simply accepted the fluent-sounding answer without question. Responsibility for the idea quietly shifted from the group to the individual student, and then often vanished entirely, leaving no one to truly own the claim.

When the researcher made the AI's work visible, the behavior of the groups changed dramatically. In the condition where students used their own AI but showed the full history to the group, the dynamic became much more balanced. The other members could see exactly what the machine had said and how the student had edited it. This transparency allowed the group to act as a team of inspectors. They could ask, "Why did you ask the AI that?" or "The machine said this, but you changed that part; why?" This visibility turned the AI from a black box into a shared object that the group could examine together. The students were more likely to challenge the AI's suggestions, verify facts, and repair errors when they happened. The group did not just accept the output; they actively regulated the process, deciding together what was trustworthy and what needed more work.

Interestingly, the study found that having a shared AI workspace, where everyone used the same tool at the same time, did not always produce better results than the traceable personal AI. While the shared workspace made the AI's output public, it did not necessarily encourage the same level of critical inspection as the traceable personal condition. The key factor was not just that everyone could see the screen, but that they could see the specific path of how a single student interacted with the machine. The ability to trace the history of a claim—from the initial question to the final edited sentence—seemed to be the most powerful tool for keeping the group honest.

The researcher also looked at how the groups handled the specific moments when the AI gave a wrong answer. In the hidden conditions, these errors often slipped into the final group project unnoticed, or they were accepted without a real debate about their validity. In the visible conditions, the errors became trigger events for collaboration. The group would pause, discuss the discrepancy, and work together to find the correct information. This process of detecting and fixing errors was linked to higher-quality final products and better individual learning outcomes. The students who worked in the visible conditions not only produced better work, but they also demonstrated a stronger ability to apply what they learned to new tasks later on.

The study suggests that the value of AI in group work depends less on how much access everyone has to the technology and more on whether the group can inspect the work. When the path of knowledge production is invisible, responsibility becomes blurred, and the group may lose its ability to regulate its own learning. When that path is visible, the group can maintain its agency, ensuring that the AI serves as a tool for inquiry rather than a source of unexamined authority. The research indicates that for AI to be a true partner in learning, the social conditions must allow the group to see, question, and repair the machine's contributions. Without that visibility, the group risks outsourcing its thinking without realizing it. With it, the group can harness the power of AI while keeping the responsibility for knowledge firmly in human hands.