A student asks a chatbot to explain an idea, spots an inconsistency and asks it to try again. A programmer challenges a suggested fix. Someone preparing a difficult email changes the tone, rejects a sentence and requests a clearer explanation. These ordinary exchanges are the subject of new Stanford research into AI collaboration, and they complicate the image of people simply pressing a button and surrendering their thinking.
Researchers examined approximately 250,000 anonymised conversations with Anthropic's Claude assistant and found frequent evidence of people guiding, questioning and negotiating with the software. In Stanford's October 2026 account of the study, 56% of the collaborations were classed as consequential or more significant. That finding reflects the researchers' classification of observed conversations, not an objective measure of whether the chatbot improved the outcome or whether the decisions made were correct.
AI collaboration is more than delivering an answer
The familiar picture of AI at work is a command followed by an output: write an email, summarise a report or produce a piece of code. That sequence does occur, but it misses much of what people actually do. A conversation can begin with a vague intention, pass through several corrections and end with the user understanding the problem differently from when they began.
Stanford's researchers were interested in the nature of that exchange. They examined how important the tasks appeared to be, whether people retained initiative and how disagreements or misunderstandings developed. Their conclusions come from the patterns visible in the conversations, not from following participants home or into their offices to see how decisions turned out.
That distinction allows for a more useful question. Instead of asking whether a person used AI, ask how they used it. Someone who immediately accepts a polished answer is engaged in a very different activity from someone who checks the assumptions, compares alternatives and insists on an explanation. Both may appear in the same headline statistic about adoption.
Disagreement can be a sign of useful work
The Stanford team described friction in roughly half of the examined conversations. By friction, it meant moments when the interaction did not simply proceed from request to agreeable answer. The person or the model might question a claim, identify confusion or require clarification. That could be annoying, but it could also be where actual learning takes place.
Imagine preparing a letter to a landlord. The assistant proposes a forceful response; the tenant worries it will escalate the dispute. A second draft becomes more restrained, and the user asks what rights they should verify independently. That is not a triumph of one perfectly worded prompt. It is a sequence of choices in which the person remains responsible for the outcome.
According to Stanford's account, in about two-thirds of interactions the researchers identified a pattern in which users asked the AI to explain or teach rather than merely produce material. When conversations involved negative friction, users attempted to recover in almost four-fifths of cases. Those findings are descriptive observations, however, and do not establish that the resulting advice was safe or educationally effective.
Friction is not automatically beneficial. A chatbot can be stubborn, confused or falsely confident. A user can keep arguing with a wrong answer and emerge more confused than before. The valuable part is not disagreement for its own sake, but the willingness to inspect what is being said and change course when evidence warrants it.
The serious conversations demand extra caution
The researchers observed people bringing consequential matters to the assistant, including professional, legal, financial and personal questions. When the stakes rose, conversations tended to become more involved. This makes intuitive sense: someone choosing an adjective in a birthday message has less reason to scrutinise each reply than someone asking about an employment dispute or complex software deployment.
Yet more conversation does not automatically mean more reliable advice. A person might ask thoughtful questions and still lack the expertise needed to identify a subtle mistake. An assistant can produce detailed explanations that sound convincing while resting on an incorrect premise. The danger increases when fluent language makes uncertainty disappear from view.
That is why any lesson from these findings must preserve human responsibility. A chatbot can help organise questions for a solicitor, explain a financial term or suggest what evidence to collect, but it should not be treated as a substitute for the professional examination of a high-stakes case. In some situations, the best use of the conversation is to prepare for a decision that happens elsewhere.
The problem also has an inequality dimension. People who already understand a subject may be better at challenging the model. Someone encountering the same subject for the first time may not know which question to ask. A useful tool should help less experienced users detect uncertainty rather than reward only those who arrive with substantial expertise.
What a quarter-million chats can and cannot tell us
The sample is unusually large for a close study of human-machine interactions. It provides a view of repeated conversational behaviours at a scale that would be hard to obtain through interviews alone. The Stanford account describes the analytical focus on task importance, agency and friction, while related Stanford workplace research explores how managerial support affects the way staff engage with AI. The original data were anonymised and analysed through an academic-industry collaboration, rather than assembled by individually interviewing every participant.
Still, the sample consists of conversations from one named assistant under the conditions of the research. People who choose to use Claude are not a random sample of the population. Messages also show what users typed and what the system returned, not everything they thought. A person might independently verify an answer after leaving the chat, or might ignore a warning. Neither behaviour is necessarily visible in the transcript.
The analysis also relies on definitions. Deciding whether a task is consequential involves judgement, and labelling a conversation as collaborative does not prove that participants learned something lasting. The study has been reported by Stanford as accepted for publication in Nature Human Behaviour; that status should not be confused with evidence that every finding has been independently replicated across other assistants.
These limitations do not erase the value of the work. They show why a behavioural map of conversations should be treated as one layer of evidence. The next layer would be whether different interaction styles lead to better factual accuracy, more durable skills or wiser decisions outside the chat window.
The comparison with human teams is revealing
Earlier LiveAIWire reporting on human-AI teamwork considered whether people and systems perform better together than either does alone. That is a question about results across experimental tasks. Stanford's new analysis asks something more intimate: what the exchange itself looks like when people are working through problems in the wild.
A capable human colleague does not always agree with an initial request. They may point out that a deadline is impossible, ask for missing information or suggest a safer plan. A useful AI assistant ought to make room for similar correction, without pretending that a statistical model carries the responsibilities or experience of a human professional.
The distinction between help and hidden extra labour is also important. Our report on AI making work faster but increasing responsibilities captured another side of the productivity discussion. A conversation can be intellectually engaging while taking longer than simply doing a straightforward task. Employees need to judge whether the improvement in quality, understanding or confidence justifies the time.
And people may interact differently when they know their work is being studied or when an assistant changes its behaviour. LiveAIWire's examination of Anthropic's user interviews provides related context for why listening to people's experience matters alongside numerical measures. Neither interviews nor transcript analysis alone tells the whole story.
Better assistants might invite the right kind of question
The most practical implication is a design one. If people benefit from being able to ask for reasons and challenge suggestions, interfaces should make those activities easier. An assistant could distinguish what it knows from what it is inferring, suggest where a calculation needs checking and offer an explanation appropriate to the user's experience. It could also make it less awkward to say that the previous answer was wrong.
Users can practise that habit immediately. When the stakes are meaningful, ask what assumptions an answer depends on, where the evidence comes from and what a credible alternative would look like. Check the important claims outside the conversation. Notice whether the assistant is actually responding to a challenge or simply rewriting the same unsupported conclusion in more reassuring language.
The Stanford findings resist both easy optimism and easy despair. People are not always helpless passengers, and a chatbot is not automatically a wise collaborator because it keeps talking. The promising development is more modest: many users are already treating AI as something to question and direct. The next challenge is making that behaviour lead to demonstrably better judgement.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.