There are two ways to work with AI. You can paste in the task and accept what comes back. Or you can decide what you actually want, break the job into parts, steer each part, and judge the result before it goes anywhere.
The first way feels easier. A new experiment from Microsoft Research suggests the second way produces better work, and it found that people who were nudged into it rated the AI as less helpful while doing better. If you want to use AI without losing critical thinking, that gap between how it feels and how it performs is the most useful thing to understand.
Below: what four research papers actually found (numbers checked against the papers), what they don't show, and a practical workflow for being an active partner with AI rather than a passive prompter. At the end, we show how we think a canvas like Slashspace fits, and we mark clearly where that's our interpretation rather than a finding.
The paper is "Metacognitive scaffolds create productive friction for AI-assisted knowledge work" by Viktor Kewenig, Lev Tankelevitch, Sean Rintel, Advait Sarkar and Abigail Sellen of Microsoft Research. It was posted to arXiv on October 22, 2026, so it's a preprint that hasn't been through peer review yet.
The researchers recruited US-based adults through Prolific to write two professional briefs each, such as a city-planning proposal on a four-day work week. The final sample was 297 people and 594 essays.
Everyone used the same setup: a chat panel connected to GPT-5.3 and an editable canvas panel for the document.
The difference was in the first task. Half the participants got a metacognitive scaffold. ("Metacognition" just means thinking about your own thinking: knowing what you're aiming for and checking whether you're getting there.) After they typed a request, the interface asked them to fill in the message they wanted to convey, the target reader, the outcome they wanted from that reader, their main objective and the scope. Then it generated task-specific controls for breaking the work down: analytical approaches, structure and critical points to cover.
In the second task, nobody had the scaffold. That let the researchers see whether anything stuck.
Better writing, and it carried over. An AI judge (GPT-5.4) scored essays on a seven-part rubric covering things like argument, evidence and clarity of the recommendation. Scaffolded writers scored higher on the first task (Hedges' g = 0.29, a small but real effect). On the second task, with the scaffold gone, the previously scaffolded group still scored higher (g = 0.21).
It wasn't just extra time. The scaffolded group took longer on the first task (about 28 minutes versus 16). On the second task both groups took about 11 minutes, and accounting for time didn't change the result.
People learned to prompt differently. On the second task, prompts were coded for goal clarity and task decomposition. Previously scaffolded participants scored 14.4 out of 39, versus 11.1 for the others. After one exposure, people stated goals and broke work down on their own.
They remembered more. About three days later, participants were asked what each piece had been about. For the first task, 64.6% of scaffolded participants recalled the topic, versus 45.6% of controls. For the second task it was 54.4% versus 38.0%.
And they liked it less. Scaffolded participants rated the AI as less helpful, reported less sense of control, and reported more mental load during the scaffolded task (4.71 versus 3.96 on the NASA-TLX workload scale). On the next task, the load difference was small and not statistically reliable (4.08 versus 3.93).
The authors put it this way: scaffolded participants "rate the AI as less helpful and more cognitively demanding even as they perform better, suggesting that subjective ease can understate the value of support." They call this "a productive friction worth designing for."
One more detail: in the second task, 37% of previously scaffolded participants and 44% of the others pasted the task wording in as their prompt, despite a bonus on offer for the best briefs. Full delegation is tempting.
The authors are careful about limits:
So this is promising evidence, not a settled law. It fits with earlier work, though.
Confidence in AI goes with less critical thinking. Hao-Ping Lee and colleagues surveyed 319 knowledge workers about 936 real examples of using generative AI at work (CHI 2025). They found that "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking." They also found the work shifting toward "information verification, response integration, and task stewardship": turning intentions into prompts, steering responses and checking them against your own standards. It's a self-report survey, so it shows associations, not cause and effect.
Ready-made summaries can make learning shallower. Shiri Melumad and Jin Ho Yun ran seven experiments with 10,462 participants (PNAS Nexus, 2025). People who learned a topic from AI summaries, rather than from regular web search, felt less invested in the advice they then wrote, and their advice was sparser, less original and less likely to be adopted by the people reading it. The authors compare it to "being shown the solution to a math problem rather than trying to solve it oneself." Their tasks were everyday topics like planting a vegetable garden, so it may not hold for specialist subjects or simple fact lookups.
Working with AI is like managing a team. Lev Tankelevitch and colleagues (CHI 2024) argued, from psychology research and earlier user studies, that generative AI puts heavy "metacognitive demands" on users. Their analogy: "A manager needs to clearly understand and formulate their goals, break down those goals into communicable tasks, confidently assess the quality of the team's output, and adjust plans accordingly along the way." It's a perspective paper, not an experiment.
Put together: the useful human work around AI is setting goals, decomposing and judging. Tools that hand you a finished answer can skip those steps. Tools that bring them forward can feel harder and still give you better work.
Here is the difference in practice.
| Passive prompter | Active partner |
|---|---|
| Pastes the task in as the prompt | Writes down the goal, reader and outcome first |
| Asks for the whole thing at once | Breaks the work into parts and asks for each |
| Takes the first answer | Tries alternatives and compares them |
| Lets the AI decide what matters | Chooses what the AI reads |
| Judges by "does this look done?" | Judges against the goal they wrote down |
| Ends with an output | Ends with an output and an understanding of it |
None of this means doing everything by hand. It means spending the effort AI saves you on the parts only you can do.
These steps come straight from what the research measured. They work in ChatGPT, Claude, Gemini, TypingMind or anything else.
This is the part where we talk about our own product, so to be plain about it: the studies above did not test Slashspace, and Slashspace doesn't have the study's goal-and-decomposition scaffold. What follows is our reading of how a canvas lines up with the habits the research rewards. You still have to do the thinking. A canvas just gives it somewhere to live.
Slashspace is an AI infinite canvas: instead of one scrolling chat, your conversations, notes and sources sit side by side as nodes.
Your intent can sit on the canvas. Write your goal, reader and outcome in a text node, which works like a page. Connect it to a chat and the chat reads it with every message, so the AI works against what you said you wanted, and so do you when you review.
Decomposition becomes visible. Each part of the problem can be its own chat node, so you can see the whole problem at once instead of scrolling back through one thread.
You choose what each chat reads. Connections decide what a chat sees: connect two sources to one chat and your notes to another, and neither reads the other's material. A chat in Isolated context reads only its own conversation, which the docs suggest for a critique you don't want anchored by your notes.
Alternatives sit side by side. Branching forks any answer into a new chat beside the original, which stays as it was. Sibling branches can't see each other, so three branches give you three clean angles. Select a sentence you doubt and ask about it, and the question opens in its own branch.
Models can disagree in front of you. Each chat node has its own model, so you can put the same question to Claude, GPT, Gemini and a local model and read the answers next to each other. Where they disagree is where to think harder.
Related material stays together. A group wraps related nodes in a labeled box; connect it and a chat reads everything inside. Stacks, added in Slashspace 4.0, put finished pieces away without deleting them and pass their members to a chat the same way.
Your reasoning stays on the record. Weeks later, the canvas still shows which questions led where and which branches you dropped. Replays, in beta since 4.0, let others watch a canvas build itself in the order you made it: your thinking, not just your conclusion.
Our post on why your work needs a repository is about giving AI the right material to read. This one is the other half: keeping yourself in the loop while it does.
@ in a chat to mention and connect a node quickly.One honest note: Slashspace also has an Orchestrator mode where the AI builds the canvas from a goal you describe. That's useful for speed. To stay an active partner, build the structure yourself first and let the AI fill in the pieces you've defined.
A canvas is for people who want to stay the author of the work: researchers weighing sources, writers shaping an argument, students who need to understand rather than just submit, and anyone making a decision they'll have to explain later. For quick one-off answers, a plain chat app is fine, and often faster. Our post on linear vs non-linear AI chat apps goes into when each makes sense.
It can, depending on how you use it. A survey of 319 knowledge workers found that people with more confidence in AI reported less critical thinking. An experiment with 10,462 participants found that learning from AI summaries led to shallower advice than learning through web search. But the newest study shows the reverse is possible: a tool that made people plan and decompose led to better work and better recall.
It means handing mental work to a tool, like using a calculator or a to-do list. Offloading routine work is often sensible. The research above is about offloading the parts that build understanding and judgment, such as deciding what you want and checking the result.
Not yet. It was posted to arXiv as a preprint on October 22, 2026. The two CHI papers and the PNAS Nexus paper are peer-reviewed publications.
No. The experiment used its own interface with a chat panel and an editable canvas panel. Our view that a visual canvas supports the same habits is an interpretation, not a finding, and a canvas won't make you think harder on its own.
$89/year (3-day card-required trial) or $249 one-time lifetime. See pricing.
If you'd rather think with AI than hand your thinking to it, download Slashspace and start with one text node: write down what you actually want before you ask for anything.
Kewenig, V., Tankelevitch, L., Rintel, S., Sarkar, A., & Sellen, A. (2026). Metacognitive scaffolds create productive friction for AI-assisted knowledge work (arXiv:2610.30945) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2610.30945
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 1121, pp. 1-22). ACM. https://doi.org/10.1145/3706598.3713778
Melumad, S., & Yun, J. H. (2025). Experimental evidence of the effects of large language models versus web search on depth of learning. PNAS Nexus, 4(10), pgaf316. https://doi.org/10.1093/pnasnexus/pgaf316
Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A. E., Sarkar, A., Sellen, A., & Rintel, S. (2024). The metacognitive demands and opportunities of generative AI. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (pp. 1-24). ACM. https://doi.org/10.1145/3613904.3642902