Each row is a claim that several studies support together, and each column is one study. A filled dot marks a study that is a main source of evidence for the claim. An open dot marks one that adds supporting evidence, and a diamond one that qualifies or complicates it. Pick a claim to read what the studies show, how strong the evidence is and what it does not show. Pick a study to see every claim it feeds.
AI moves effort rather than removing it, from doing work to stewarding it
Across a review, a survey and user studies, generative AI took over production while people's effort shifted to checking, integrating and adapting output they did not create. That checking needs understanding the AI may erode.
Evidence: A narrative review, a correlational survey of 319 knowledge workers and small qualitative studies.
Does not show: How much productivity is lost or gained. The survey measured how effort felt, not actual effort.
● Core evidence
- Ironies Decades of automation research predict a shift from producing work to checking output, which demands understanding people no longer have.
- Critical Thinking Effort moved from gathering information to verifying it, and from solving problems to fitting AI responses into the work.
○ Supports
- Rubber Duck Checking answers was laborious for spreadsheet users, and made-up content was hard to spot outside their own expertise.
- Future of Skill Skill is done together with colleagues and tools, so designers should ask whether AI extends that skill or stands in for it.
The new work is metacognitive, saying what you want and judging what you get
Using generative AI well means stating goals that usually stay implicit, breaking tasks down and judging output with well-calibrated confidence. People found stating intent the hardest part, and many skipped it.
Evidence: A conceptual framework drawing on user studies, small qualitative studies and a randomised online experiment.
Does not show: How common or how large these demands are across tasks. The framework is argued, and the studies were small.
● Core evidence
- Metacognitive Demands Prompting requires awareness of one's own goals and how to break them down, and judging output needs confidence that matches ability.
- Rubber Duck Spreadsheet users found it hardest to say what they wanted and to work out what context the AI needed.
○ Supports
- Scaffolds Even with a bonus for the best work, 40% of people pasted the task description in as their prompt.
- Participatory Prompting Used to search engines, people started with short queries that could not carry their context.
- Role Prompts Role-based prompts draw on shared ways of sorting people into roles, and AI replies reflect that understanding back.
Confidence governs scrutiny
People checked AI output less when they trusted the AI more than themselves, or when a task seemed trivial or time was short. Habits of trying first and checking went with better learning.
Evidence: A correlational survey, an association within a randomised trial, a review and interviews. None shows that trust causes less thinking.
Does not show: That trusting AI causes less critical thinking. The survey is correlational and self-reported.
● Core evidence
- Critical Thinking More confidence in the AI went with less reported critical thinking, and more self-confidence went with more.
- Learning Behaviour Students who tried problems first and checked the AI's answers scored higher on a later exam without AI.
○ Supports
- Metacognitive Demands Fast, fluent output may inflate people's confidence in it.
- Ironies The shift to checking, with reduced awareness, can breed complacency and over-reliance, as automation research predicts.
- New Calculator Students trusted AI less where they lacked expertise and double-checked its output against other sources.
What feels helpful is not always what helps
In three experiments, people preferred or enjoyed the easier AI condition while their learning or work gained more from the harder one. Judging tools by satisfaction or speed can pick the wrong design.
Evidence: A preregistered school experiment, a randomised online experiment and a lab study, plus a workshop argument.
Does not show: That easier tools always help less. The experiments were single sessions with specific tasks.
● Core evidence
- Notes vs Chatbot Students enjoyed the chatbot more and found it easier, yet understood and remembered less than when they took notes.
- Scaffolds People rated the AI less helpful and felt less in control while the scaffold was improving their work.
○ Supports
- Prompt Middleware People preferred controls generated for each request and rated them better for steering, though they were not more successful with them.
- Tools for Thought Designs that add friction to improve thinking meet resistance, so tools should be judged by how they change thinking.
◆ Complicates
- Ironies Generative AI helps most with simple tasks and can make demanding ones harder, so ease depends on the task.
How people use AI shapes what they learn, and that skill is unevenly spread
Access alone is not the lever. Learning depended on whether people synthesised, tried first and checked, and better-prepared learners did more of this, so equal access may widen gaps.
Evidence: Two randomised experiments in education and interviews with students and educators.
Does not show: That teaching AI habits closes learning gaps. That has not yet been tested.
● Core evidence
- Learning Behaviour Better-prepared students more often tried first and checked the AI, and these habits explained part of their exam advantage.
- Notes vs Chatbot Taking notes kept students doing the synthesis that the chatbot alone may have done for them.
○ Supports
- New Calculator With guidance focused on plagiarism, students governed their own AI use by unspoken rules, some anxiously.
- Scaffolds A brief scaffold may be a cheap way to build the skill of directing AI, which is unevenly distributed.
- Tools for Thought Using AI for direct solutions can impair learning, while using it to seek explanations can help.
Scaffold intent, ask before answering, and keep people in charge
The designs that helped asked people to state goals before AI drafted, offered questions and options rather than verdicts, and left judgement to people. The cost is effort, which has to be seen to pay off.
Evidence: One randomised experiment, a lab study and design prototypes. Most of these designs have not been tested at work.
Does not show: That these designs help in real workplaces over time. Most were tested once, online or in a lab.
● Core evidence
- Scaffolds Prompting people to set goals and break down a task before AI drafted improved their work and their later prompting.
- Metacognitive Demands Tools could prompt planning and self-evaluation, then fade the prompts as people take on the strategies themselves.
○ Supports
- Prompt Middleware Options generated with each answer worked as guidance, showing people what they could ask for and prompting reflection.
- YES AND AI personas that spoke only when confident of adding something new asked clarifying questions instead of jumping to solutions.
- Meeting in the Middle Blind colleagues wanted raw cues about a meeting to interpret themselves, not machine verdicts.
- Ironies Feedback, explanations and well-timed suggestions can keep people in the loop without breaking their concentration.
Goals become the shared object between people and AI
As AI takes over production, people's work shifts to setting, tracking and judging goals. Goals that persist across tools, and work histories AI can query, would let people and agents work from the same purpose. Meeting intentionality is the same practice for a group.
Evidence: A position paper, a systems demonstration and commentaries. These are proposals, not tested results.
Does not show: That shared goal objects improve work with AI. These are proposals and demonstrations, not tested results.
● Core evidence
- Goals First-Class Goals should be persistent objects that people and AI agents share, carried across applications and shown during the work.
○ Supports
- Stund When every voice, video and edit leaves a trace AI can query, a team's history becomes working material.
- Role Prompts AI could coach people to state the stances that get them relevant replies, and help teams understand each other.
- Future of Skill Designing for future work means attending to how people thread colleagues and tools together, not only to adding intelligence.