Tools for Thought: Understanding, Protecting, and Augmenting Human Cognition with Generative AI — From Vision to Implementation
Authors: Zelun Tony Zhang, Nick von Felten, Leon Reicherts, Lev Tankelevitch, Zhitong Guan, Sean Rintel, Yue Fu, Jessica He, Kenneth Holstein, Advait Sarkar, Gonzalo Ramos, Anuschka Schmitt, Anjali Singh, Haotian Li, Srishti Palani, Peter Dalsgaard
A proposal for a CHI 2026 workshop on generative AI as a tool for thought, meaning technology judged by its effect on thinking, not only productivity. It builds on a first workshop at CHI 2025.
Draws onPrior studies of generative AI and learning, and the first workshop, which had 56 registered participants and 34 accepted submissions.
ExaminesThree themes: strategies for designing and using such tools, the outcomes to measure, and adoption.
PlansA 180-minute workshop for a target of 45 participants, in small groups that each make and pitch a tangible output.
BWhat the paper argues
What is knownprior studies the proposal cites
•Studies cited in the proposal suggest how generative AI is used matters for learning.Using it for direct solutions can impair learning, while seeking explanations can help. Students may rate AI as more helpful even when learning is worse.
•The paper notes that many tools from the first workshop did not easily transfer.These tools were specific designs for particular domains. The paper asks how to turn such examples into reusable strategies.
What the field needsthe proposal's arguments
•The paper argues tools for thought should be judged by how they shape thinking.Judging should cover activities such as decision-making, reflection and exploration, not only whether the task gets done.
•The paper argues evaluation of tools for thought must combine several methods.Some effects on thinking are hard to quantify. Evaluation could combine process tracing, analysis of what people produce, cognitive assessments and long-term studies.
•The paper argues designs that add friction to improve thinking can meet resistance.Users may dislike the extra work. A cited global study found 96% of senior executives expect AI to boost productivity.
CKey takeaway
Generative AI tools for thought can protect and extend people's thinking as well as speed up work. Judging tools by their effect on thinking and learning can guide better design.
Made for this site with AI assistance from the verified research brief. It is not the paper's own figure.
Abstract
Building on the first Tools for Thought (TfT) workshop at CHI 2025, we invite researchers, designers, and practitioners to further operationalise approaches for the design, usage, and evaluation of Generative Artificial Intelligence (GenAI) as a TfT. The first goal of the workshop is to put into focus which outcomes a TfT should help people achieve to effectively augment their cognition while avoiding its erosion. Secondly, we will explore how to achieve these outcomes through design and usage strategies. Third, the workshop will also address what a TfT needs for its successful adoption and integration into people’s flow, so that they can benefit from the tools’ potential in their own terms. By focussing on these three research goals, the workshop aims to further develop and advance the multidisciplinary TfT community interested in exploring research frameworks, theories, methods, and approaches to conceptualising, designing, and researching GenAI as a TfT.
Research brief
By Zelun Tony Zhang, Nick von Felten, Leon Reicherts, Lev Tankelevitch, Zhitong Guan, Sean Rintel, Yue Fu, Jessica He, Kenneth Holstein, Advait Sarkar, Gonzalo Ramos, Anuschka Schmitt, Anjali Singh, Haotian Li, Srishti Palani, and Peter Dalsgaard.
Takeaway: Generative AI can be a tool for thought (protecting and augmenting reasoning rather than only speeding work), but that means judging tools by how they change thinking, not just output.
Why this matters
What was known
Generative AI use has documented costs to thinking and learning, but well-designed use can augment cognition. The first Tools for Thought workshop at CHI 2025 began mapping these effects. Prior work has found worsened learning outcomes and decreased metacognitive effort resulting from GenAI use, both in knowledge work and in education(§1, p. 2)
What this paper adds
This second workshop moves from mapping the field to operationalising it. It focuses on concrete strategies, outcomes and measures, and on adoption of generative AI as a tool for thought. our aim in this year's iteration is to transition from the foundational to the operational(§1, p. 2)
Who it matters to
Researchers, designers, practitioners and provocateurs working on generative AI and human cognition. (§8, p. 6)
Research questions
What makes a TfT effective?
What are the concrete mechanisms and appropriate outcomes of TfTs?
How can we establish and validate sensible metrics to measure their impact on human cognition?
Method and data
Design
Workshop proposal for CHI 2026, building on the first Tools for Thought workshop at CHI 2025.
Setting
A 180-minute workshop in two 90-minute sessions, organised around small groups of four to five participants clustered by theme from their submissions and Miro boards.
Participants / data
Submissions of up to four pages on at least one of three themes (strategies, outcomes, adoption). A target of 45 participants.
Analysis
Groups work on a tangible output of their choice (for example a prototype, framework, design principles or theory) and pitch it to the plenary. Afterwards, a gap-finding exercise and a public synthesis report.
Findings
How AI is used matters for learning. Prior studies we cite show that generative AI's effect on learning is mixed. Using it for direct solutions can impair learning, while using it to seek explanations can help. Perceptions can also mislead: #
“students may perceive LLMs as more helpful than traditional study methods, such as note-taking, even when their learning outcomes are worse”
Good examples, hard to transfer. Many tools for thought presented at the first workshop were specific designs that did not easily carry over to other domains or problems, so we ask how to turn them into reusable strategies. #
“How can we systematise these examples into tangible strategies that can reliably lead to effective and positive TfTs, given a new problem domain or context?”
Judge thinking, not just output. We argue that a tool for thought should be judged by how it shapes, supports or extends thinking in activities such as decision-making, reflection or exploration, not only by whether the task gets done. #
“for example, when supporting a sensemaking activity by taking the role of a socratic tutor and facilitating users' critical engagement with a topic”
Evaluation must be pluralistic. Some effects on thinking are hard to quantify or appear only in users' own accounts of what they achieved, so evaluation could combine process tracing, analysis of what people produce, cognitive assessments and long-term studies. #
“This calls for evaluation methods that are flexible and pluralistic, capturing not just what users produce, but also how they think and change in the process.”
Friction meets resistance. Citing earlier studies, we argue that designs adding friction to improve thinking can meet resistance, both from users who dislike the extra work and from stakeholders who expect AI to speed work up. #
“A global study of 2500 executives and workers found that 96% of C-suite executives expect AI to boost productivity”
Questions this paper answers
Does using AI like ChatGPT make people worse at thinking and learning?
It depends on how it is used. Studies have found worse learning and less metacognitive effort with some generative AI use. But using it to seek explanations, rather than direct solutions, can help learning. (See §1, p. 2.)
How should we judge whether an AI tool helps people think?
We argue for looking beyond task performance to how a tool shapes thinking. That means the artefacts that scaffold thought, changes in understanding and skills, and task outcomes. Flexible methods such as process tracing, artefact analysis and longitudinal studies can capture these. (See §1.2, p. 3; §1.2, p. 3.)
Why might people resist AI tools designed to make them think harder?
Because added friction means extra work. Designs that add friction can meet resistance even when they aim to improve thinking, from users who dislike the effort and from stakeholders who expect AI to accelerate work. (See §1.3, p. 3.)
What is a 'tool for thought'?
A system designed and judged mainly by its short- and long-term effect on people's thinking (such as reasoning, memory, ideation or critical thinking), rather than by productivity alone. Even a general-purpose chatbot can serve this way if people use it in ways that protect their cognition. (See §1.1, p. 2; §1.2, p. 3.)
Quantitative results
Measure
Result
Where
CHI 2025 Tools for Thought workshop: registered participants, accepted submissions, submissions received
56 officially registered participants, with 34 accepted submissions and 72 submissions received
§7, p. 5
Contributions and implications
The field needs concrete pathways toward generative AI as an effective and positive tool for thought. There is a need for concrete pathways toward GenAI as an effective and positive TfT.(§1, p. 2)
Strategies should be found that let people see value in, and attain, both productivity and cognitive outcomes. How might we manage potential trade-offs between productivity and cognitive outcomes to maximise the likelihood of their successful integration and sustained use?(§1.3, p. 3)
Limitations
Tools for thought are an emerging area, and gaps in understanding are expected. As TfTs are an emerging research area, we expect to find gaps in our understanding of them.(§1, p. 2)
›
Concepts
Tools for Thought (TfT)
What we mean by the term: we refer to the characterisation of (sociotechnical) systems that put special emphasis on their impact on cognition - in both the short and long term - over mere productivity(§1, p. 2)
Process-oriented support
A design strategy that emerged from the first workshop: This strategy focusses on supporting users to reason forward to their own solution, rather than backward from an AI-generated one.(§1.1, p. 2)
Usage strategies
The second type of strategy, which applies even to general-purpose chatbots: Users themselves can also employ workflows or activities involving GenAI in a way that augments or protects their cognition(§1.1, p. 2)
Intermediary, cognitive and task outcomes
Three types of outcome of using a tool for thought, drawn from Yvonne Rogers and colleagues: artefacts that scaffold thinking, changes in understanding and skills, and task performance. Cognitive Outcomes: Changes in understanding, learning, or cognitive skills, such as critical thinking and metacognitive engagement.(§1.2, p. 3)
How to cite
Zelun Tony Zhang, Nick von Felten, Leon Reicherts, Lev Tankelevitch, Zhitong Guan, Sean Rintel, Yue Fu, Jessica He, Kenneth Holstein, Advait Sarkar, Gonzalo Ramos, Anuschka Schmitt, Anjali Singh, Haotian Li, Srishti Palani, and Peter Dalsgaard. 2026. Tools for Thought: Understanding, Protecting, and Augmenting Human Cognition with Generative AI — From Vision to Implementation. CHI 2026. https://doi.org/10.1145/3772363.3778733
This brief was prepared with AI assistance and published by Sean Rintel. It is not the authors' published text. Page numbers refer to this PDF.
BibTeX
@inproceedings{zhang2026tools,
author = {Zhang, Zelun Tony and von Felten, Nick and Reicherts, Leon and Tankelevitch, Lev and Guan, Zhitong and Rintel, Sean and Fu, Yue and He, Jessica and Holstein, Kenneth and Sarkar, Advait and Ramos, Gonzalo and Schmitt, Anuschka and Singh, Anjali and Li, Haotian and Palani, Srishti and Dalsgaard, Peter},
title = {Tools for Thought: Understanding, Protecting, and Augmenting Human Cognition with Generative AI — From Vision to Implementation},
booktitle = {CHI 2026},
year = {2026},
month = {April},
publisher = {ACM},
url = {https://doi.org/10.1145/3772363.3778733},
doi = {10.1145/3772363.3778733},
}