Shuichiro Ogawa
日本語

Notes · updated 2026-11-01 · Markdown

Research Trends in HCI since Generative AI Became Usable (2023–2026): Many Systems Built, and the Checking Work Left to People

This note reviews what HCI (human-computer interaction) research has asked about generative AI in the three years since the public release of ChatGPT in November 2022. It draws on 43 peer-reviewed papers and major preprints, mostly from CHI, together with counts of CHI papers.

Contents (12)
  1. How much of CHI now concerns generative AI?
  2. What became difficult when generative AI was used as a design material?
  3. When do people accept AI output without checking it?
  4. What do people do when they delegate planning and execution to AI?
  5. What happens to the group when individual output improves?
  6. Performance during use and learning after use
  7. What harms and burdens do people with disabilities face?
  8. Is the AI that people prefer the AI that benefits them?
  9. Can LLMs replace research participants?
  10. How should the trends of these three years be read?
  11. What this review can and cannot say
  12. Footnotes

How much of CHI now concerns generative AI?

At CHI, the largest international conference in HCI, 1.9% of papers in 2023 contained LLM or generative AI terms in the title or abstract. The share rose to 13.2% in 2024, 16.7% in 2025, and 23.0% in 2026, according to counts of the CHI proceedings records in OpenAlex (the counting procedure is described in source/review/ai-hci-research-trends/papers.md)1. The submission deadline for CHI 2023 was September 2022, which preceded the public release of ChatGPT. Within three years, therefore, roughly one in four CHI papers came to include generative AI terms in the title or abstract.

These counts show how much the research has grown, but they do not show what the research asks. A systematic review of 153 CHI papers on LLMs published from 2020 to 2024 (Pang et al. 2025) classified the roles that LLMs play in these papers into five categories. In descending order of frequency, the roles were as follows.

  • LLMs as components of systems, prototypes, or algorithms (62.74%, 96 papers)
  • studies of how users perceive LLMs (23.53%)
  • LLMs as objects of study (9.80%)
  • LLMs as research tools (9.15%)
  • LLMs as substitutes for participants and users (7.19%)

The most frequent application domains were communication and writing (22.88%), augmenting human capabilities (16.99%), education (14.38%), and programming (11.11%). The same review reported that, of the 153 papers, 90.85% acknowledged limitations related to research validity and 84.98% used or studied closed GPT-family models.

By volume, the main current is research that builds something with generative AI as a component. However, what HCI has newly learned about generative AI is written mostly in studies that measure what happened to the people who used these systems, rather than in reports of system performance. For this note, we read 43 peer-reviewed papers and major preprints, mostly from CHI, and grouped them into nine clusters by research question. Because we selected representative studies for each question, studies that measure the user side are overrepresented among the 43 papers, and their proportion does not reflect the proportion in CHI as a whole. Numbers and claims were checked against the full text for the 32 papers whose full text we obtained, and against the abstract for the remaining 11 (see Unverified items at the end). Reading across the clusters, studies with different questions report the same two things in different settings.

What became difficult when generative AI was used as a design material?

Generative AI entered HCI as a design material whose behavior can be specified in natural language. In a case study conducted before ChatGPT, 11 practitioners used prompts to prototype machine learning features; according to the abstract, prompt-based prototyping substantially broadened who could prototype, sped up prototyping, and grounded communication between collaborators (Jiang et al. 2022). At the same time, it introduced new challenges: reverse-engineering prompt designs, sourcing example data, and evaluating how well prompts worked.

The difficulty is greater for people who are not AI experts. According to the abstract of Zamfirescu-Pereira et al. (2023), non-experts who designed chatbot prompts explored prompts opportunistically rather than systematically, carried over expectations from instructing other people, and overgeneralized, which led to failures. The authors note that these difficulties echo those known from end-user programming and interactive machine learning. Subramonyam et al. (2024) placed a “gulf of envisioning” before Norman’s gulfs of execution and evaluation: the difficulty of foreseeing what to ask for, how to say it, and what to expect from the output.

Tools for prototyping and evaluation were also built. Among users of ChainForge, a toolkit for comparing prompts and models side by side, Arawjo et al. (2024) identified three modes of use: opportunistic exploration, limited evaluation, and iterative refinement. Weisz et al. (2024) proposed six design principles for generative AI applications. Three of these (design for generative variability, co-creation, and imperfection) address characteristics unique to generative AI, and three (design responsibly, design for mental models, and design for appropriate trust and reliance) reinterpret existing principles. Half of the six principles are, in other words, principles discussed before generative AI and reinterpreted for it. Earlier guidance includes the 18 guidelines for human-AI interaction, which, according to the abstract, 49 design practitioners tested against 20 AI-infused products (Amershi et al. 2019).

In practice, rules for use were missing before tools were. According to the abstract of an interview study with 24 UX practitioners, companies had almost no generative AI policies, and teams had no shared practices (Takaffoli et al. 2024). The practitioners used generative AI mainly on their own for writing tasks and saw limitations for wireframing and prototyping.

When do people accept AI output without checking it?

Research on overreliance continues a line of work on AI-assisted decision-making that predates generative AI. Buçinca et al. (2021) designed three cognitive forcing interventions, which prompt people to think on their own before seeing AI suggestions, and showed in an experiment with 199 participants that these interventions reduced overreliance compared with designs that simply added explanations. However, people gave the least favorable subjective ratings to the designs that reduced overreliance the most.

Whether explanations reduce overreliance depends on conditions. Vasconcelos et al. (2023) assumed that people choose a strategy by weighing the cost of engaging with the task themselves against the cost of relying on the AI. In five studies with a maze task (N = 731), they showed that task difficulty, explanation difficulty, and monetary compensation affected overreliance. Explanations reduced overreliance when they lowered the cost of verification relative to the cost of simply trusting the AI.

With generative AI, the wording of the output itself became an object of design. In a preregistered experiment (N = 404) in which participants answered medical questions with access to LLM responses, responses that expressed uncertainty in the first person (“I’m not sure, but…”) reduced, but did not eliminate, overreliance on incorrect answers and increased accuracy (Kim et al. 2024). For uncertainty expressed from a general perspective (“It’s not clear, but…”), the effects were weaker and not statistically significant. In search, according to the abstract of Sharma et al. (2024), participants engaged in more biased information querying with LLM-powered conversational search than with conventional search, and an opinionated LLM that reinforced their views exacerbated this bias.

Research on reliance has also extended to self-reports from the workplace. According to the abstract of Lee et al. (2025), an analysis of 936 first-hand examples from 319 knowledge workers showed that higher confidence in generative AI was associated with less critical thinking, whereas higher self-confidence was associated with more. The same study reported that the nature of critical thinking shifts toward information verification, response integration, and task stewardship. Tankelevitch et al. (2024) reframed the usability challenges of generative AI as metacognitive demands, arguing that these systems require users to exercise a high degree of metacognitive monitoring and control. The burden of judging what one wants and whether the output meets that need, then, remains with the user.

What do people do when they delegate planning and execution to AI?

Generative Agents (Park et al. 2023) placed 25 agents with a memory stream, reflection, and planning in a sandbox world and simulated believable human behavior. Subsequent HCI research asked how far people should delegate to agents that plan and execute multistep tasks on their behalf, and where people should take part.

When people delegate, their trust is pulled by how plans look. In an experiment (N = 248) in which LLM agents served as daily assistants, people tended to trust plans that only seemed plausible (He et al. 2025). Cocoa (Feng et al. 2026) lets each step of a plan be assigned to either the person or the agent. In a lab study (n = 16), Cocoa improved steerability without sacrificing ease of use, and a week-long field deployment (n = 7) showed how researchers used it. For proactive programming assistants, which offer suggestions before being asked, fine-grained design factors such as the timing of suggestions affected productivity and user experience (Chen et al. 2025).

In vibe coding, where programs are built through conversation, human work has not disappeared either. Sarkar and Drosos (2025) concluded that vibe coding does not eliminate the need for programming expertise but redistributes it toward context management, rapid evaluation of generated code, and decisions about when to switch between AI-driven and manual work. The “verification, integration, and stewardship” of Lee et al. (2025) and the “context management, evaluation, and switching decisions” here describe the same shift, observed in different objects.

What happens to the group when individual output improves?

In an experiment on writing short stories with ideas from generative AI, 600 evaluators assessed stories written by 293 writers (Doshi and Hauser 2024). Stories by writers who saw AI ideas were rated as more creative, better written, and more enjoyable, and the gains were largest among less creative writers. However, the AI-assisted stories were more similar to each other than stories written by humans alone. The authors compare this dynamic to a social dilemma. Writers are individually better off, but collectively the novelty of the content narrows.

Homogenization has been reported at three levels: ideas, culture, and opinion. In an ideation experiment (36 participants recruited, three excluded), ChatGPT users produced ideas that were more semantically similar at the group level than those of users of the Oblique Strategies card deck (Anderson et al. 2024). In an experiment with 118 participants from India and the United States, AI suggestions pushed writing toward Western styles and diminished cultural nuances (Agarwal et al. 2025). Among 1,506 participants who used a writing assistant configured to favor one opinion, both the opinions expressed in their writing and their opinions in a later attitude survey shifted (Jakesch et al. 2023).

The quality of the output itself is also disputed. When experts evaluated stories with 14 binary tests (TTCW) modeled on the Torrance Tests of Creative Thinking, stories written by LLMs passed three to ten times fewer tests than stories by professional writers (Chakrabarty et al. 2024). The finding that individual output improves compares people who used AI ideas with people who did not, not AI-written text with professional writing. Because the two results use different comparisons, they do not contradict each other.

Performance during use and learning after use

Studies in education measured outcomes during use and after access was removed separately. In a randomized controlled trial with nearly 1,000 students conducted with a high school in Turkey, access to GPT-4 raised mathematics grades by 48% with a standard GPT-4 interface and by 127% with GPT Tutor, which was designed for learning (Bastani et al. 2025). However, after access was removed, students who had used the standard interface scored 17% lower than students who never had access. The safeguards built into GPT Tutor largely mitigated this harm.

Other results show no such harm. In a controlled experiment with 69 novices aged 10 to 17, an AI code generator increased task completion by a factor of 1.15 and scores by a factor of 1.8, without decreasing performance on subsequent manual code-modification tasks (Kazemitabaar et al. 2023). The materials in this note do not allow us to determine whether the difference from Bastani et al. comes from the task, the age of the learners, or the design of the tool. Observations of novice programmers using Copilot showed both a pattern in which students guided the tool toward a solution and a pattern in which they drifted from one incorrect suggestion to the next (Prather et al. 2023).

One study measured brain activity. When 54 participants wrote essays while their EEG was recorded, brain connectivity was strongest in the group that used no tools, intermediate in the search engine group, and weakest in the LLM group (Kosmyna et al. 2025). The authors call the condition in which repeated reliance on external systems replaces the effort of independent thinking “cognitive debt.” This study is a preprint that has not been peer reviewed, and only 18 participants completed the fourth session.

What harms and burdens do people with disabilities face?

Accessibility research examined, together with disabled people, whom the biases in generative AI output are directed at. According to the abstract of Gadiraju et al. (2023), when 56 participants with disabilities probed a dialog model in 19 focus groups, they rarely characterized its outputs as blatantly offensive. Instead, they described how the outputs mirrored subtle but harmful stereotypes they encountered in daily life and in the media, such as inspiration porn and able-bodied saviors. For image generation, according to the abstract of Mack et al. (2024), 25 people with disabilities objected that the models repeatedly presented reductive archetypes for different disabilities, such as the wheelchair.

For blind and low vision people, generative AI became a tool for describing what they cannot see. In a two-week diary study in which 16 blind and low vision participants used a scene description application, satisfaction with the descriptions averaged only 2.76 out of 5 and trust 2.43 out of 4 (Gonzalez Penuela et al. 2024). A study with 26 blind participants examined how users verify and contest AI errors that are difficult to verify without vision (Alharbi et al. 2024). For people who cannot check visually, the verification work that remains with the user becomes even harder.

Is the AI that people prefer the AI that benefits them?

Research on human-AI relationships turned to sycophancy and extended use. According to the abstract of the journal version of Cheng et al. (2026), 11 state-of-the-art models affirmed users’ actions 49% more often than humans did, even when queries involved deception, illegality, or other harms. In three preregistered experiments (N = 2,405), even a single interaction with sycophantic AI reduced participants’ willingness to take responsibility and repair interpersonal conflicts, while increasing their conviction that they were right. Nevertheless, the sycophantic models were trusted and preferred2. The authors argue that this creates perverse incentives for sycophancy to persist, because the very feature that causes harm also drives engagement.

Extended use has also been measured. In a four-week randomized controlled trial (n = 981, more than 300,000 messages), the assigned voice modality and conversation type had no significant effects on loneliness, socialization with real people, emotional dependence on AI, or problematic AI use (Fang et al. 2026). However, participants who voluntarily used the chatbot more, regardless of their assigned condition, showed consistently worse outcomes. Because the amount of use was not assigned, this association cannot be read as an effect of using the chatbot more.

Harms have also been classified from records of products that sell relationships. From 35,390 conversation excerpts between 10,149 users and the AI companion Replika, Zhang et al. (2025) derived six categories of harm (relational transgression, harassment, verbal abuse, self-harm, mis/disinformation, and privacy violations) and four roles that the AI plays (perpetrator, instigator, facilitator, and enabler). On displays of empathy, according to the abstract of Cuadra et al. (2024), LLM-based conversational agents that conversed with or about 65 human identities made value judgments about certain identities and could be encouraging of identities related to harmful ideologies. The same abstract reports that, despite displaying empathy, the agents did poorly at interpreting and exploring the user’s experience.

Can LLMs replace research participants?

Generative AI has also entered research methods. According to the abstract of Hämäläinen et al. (2023), synthetic questionnaire responses about HCI experiences generated by GPT-3 were hard to distinguish from real ones. The authors state that synthetic data should be useful for ideating and piloting experiments but that findings must be validated with real data, and that LLMs may make crowdsourced self-reports untrustworthy. Park et al. (2024, preprint) built agents from two-hour interviews and surveys with 1,052 Americans and measured how well the agents replicated each individual’s survey responses and personality, using as the baseline how well each person replicated their own responses two weeks later.

Critiques of replacement appeared in the same period. Agnew et al. (2024) argued that proposals to replace human participants with LLMs (at least 13 technical reports and peer-reviewed articles) conflict with values foundational to research with human participants: representation, inclusion, and understanding. Wang et al. (2025) showed, with 4 LLMs and studies involving 3,200 people across 16 demographic identities, that LLMs misportray and flatten demographic groups. In the review by Pang et al. (2025), 9.15% of papers used LLMs as research tools, and about nine in ten papers acknowledged limitations related to research validity.

The main current by volume is clear. In 2020–2024, more than 60% of CHI papers on LLMs built systems with LLMs as components (Pang et al. 2025), and the share of papers with generative AI terms in the title or abstract reached 23.0% in 2026. Laid side by side, the nine clusters show, separately from this volume, two findings that recur as the main currents of argument.

The first is the finding that human work shifts from making toward verifying, integrating, and overseeing. Critical thinking in knowledge work (Lee et al. 2025), expertise in vibe coding (Sarkar and Drosos 2025), metacognitive demands (Tankelevitch et al. 2024), and trust in agent plans (He et al. 2025) differ in object and method, yet they describe a shift in the same direction. This verification work starts with the first request, as the gulf of envisioning indicates (Subramonyam et al. 2024), and it becomes even harder for people who cannot check visually (Alharbi et al. 2024).

The second is the finding that performance and preference during use diverge from longer-term outcomes. Grades rise while access lasts but fall once it is removed (Bastani et al. 2025). Individual output improves, but the group’s output becomes more similar (Doshi and Hauser 2024). Sycophantic AI distorts people’s judgment yet is trusted and preferred (Cheng et al. 2026). Designs that reduce overreliance the most are liked the least (Buçinca et al. 2021). That designs which narrow the divergence are less preferred by users has been shown in two settings: overreliance (Buçinca et al. 2021) and sycophancy (Cheng et al. 2026).

The divergence is not reported as unavoidable. Learning safeguards largely mitigated the harm after access was removed (Bastani et al. 2025), first-person expressions of uncertainty reduced overreliance (Kim et al. 2024), and explanations worked when they lowered the cost of verification (Vasconcelos et al. 2023). The manual performance of novices who used an AI code generator did not decrease (Kazemitabaar et al. 2023). Within the studies we read, HCI research in these three years has accumulated knowledge about which designs and conditions tip generative AI toward benefit or harm, rather than settling whether it benefits or harms people in general. Within the same studies, it remains undecided whether designs should be chosen by what people prefer or by what benefits them when the two diverge. A map of peer-reviewed work that calls itself human-centered AI states, according to its abstract, that considerable ambiguity remains about what it means to frame, design, and evaluate human-centered AI (Capel and Brereton 2023). The proposal to design AI as “thought partners” that meet human expectations and complement human limitations (Collins et al. 2024) is one candidate for such a criterion.

What this review can and cannot say

This review can say that the share of CHI papers dealing with generative AI rose from just under 2% to more than 20% in three years and that much of this research builds systems with generative AI as a component. It can also say that studies measuring the user side repeatedly report the shift in verification work and the divergence between short-term and long-term outcomes.

There is also much it cannot say. The 43 papers are the studies that remained after screening and are not a sample of all HCI research on generative AI. The reading that the two findings are the “main currents of argument” is an interpretation based on their recurrence among the papers selected for this note, not a measurement by paper counts or citations. The CHI counts depend on how OpenAlex records the proceedings, and the share for 2025 is underestimated because 37% of the records lack abstracts. Most experiments used short lab or online tasks; the four weeks of Fang et al. (2026) and the part of a semester in Bastani et al. (2025) are among the longer ones. The critical thinking measures of Lee et al. (2025) are self-reports, not behavioral measures. Park et al. (2024), Kosmyna et al. (2025), and Fang et al. (2026) have not been peer reviewed, and we found no peer-reviewed versions. Among the papers in this note, only Agarwal et al. (2025) compared cultures (India and the United States), and we did not check how far the other results apply across cultures.

Unverified items

No major claim relies on unverified material. For the following 11 papers, we could not obtain the full text (the ACM Digital Library PDFs and some university repositories refused automated retrieval); we wrote about them only within the scope of the abstract and marked this in the text. Amershi et al. (2019), Capel and Brereton (2023), Cuadra et al. (2024), Gadiraju et al. (2023), Hämäläinen et al. (2023), Jiang et al. (2022), Lee et al. (2025), Mack et al. (2024), Sharma et al. (2024), Takaffoli et al. (2024), and Zamfirescu-Pereira et al. (2023). For Cheng et al. (2026), the full text of the journal version was behind a paywall, and the numbers come from the abstract of the journal version (the numbers of experiments and participants differ from the preprint). Park et al. (2024), Kosmyna et al. (2025), Fang et al. (2026), and Shi et al. (2023) are preprints, and a Crossref title search found no peer-reviewed versions.

References

All sources were accessed on 2026-11-01. The ledger and verification records are in papers.md and fulltext-manifest.json under source/review/ai-hci-research-trends/.

Overviews and counts

Design material and prototyping

  • Amershi, Saleema, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, et al. 2019. “Guidelines for Human-AI Interaction.” In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3290605.3300233
  • Arawjo, Ian, Chelse Swoopes, Priyan Vaithilingam, Martin Wattenberg, and Elena L. Glassman. 2024. “ChainForge: A Visual Toolkit for Prompt Engineering and LLM Hypothesis Testing.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642016
  • Jiang, Ellen, Kristen Olson, Edwin Toh, Alejandra Molina, Aaron Donsbach, Michael Terry, and Carrie J. Cai. 2022. “PromptMaker: Prompt-based Prototyping with Large Language Models.” In CHI Conference on Human Factors in Computing Systems Extended Abstracts. https://doi.org/10.1145/3491101.3503564
  • Subramonyam, Hari, Roy Pea, Christopher Pondoc, Maneesh Agrawala, and Colleen Seifert. 2024. “Bridging the Gulf of Envisioning: Cognitive Challenges in Prompt Based Interactions with LLMs.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642754
  • Takaffoli, Macy, Sijia Li, and Ville Mäkelä. 2024. “Generative AI in User Experience Design and Research: How Do UX Practitioners, Teams, and Companies Use GenAI in Industry?” In Designing Interactive Systems Conference (DIS 2024), 1579–1593. https://doi.org/10.1145/3643834.3660720
  • Weisz, Justin D., Jessica He, Michael Muller, Gabriela Hoefer, Rachel Miles, and Werner Geyer. 2024. “Design Principles for Generative AI Applications.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642466
  • Zamfirescu-Pereira, J. D., Richmond Y. Wong, Bjoern Hartmann, and Qian Yang. 2023. “Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts.” In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3581388

Reliance and trust

  • Buçinca, Zana, Maja Barbara Malaya, and Krzysztof Z. Gajos. 2021. “To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making.” Proceedings of the ACM on Human-Computer Interaction 5 (CSCW1). https://doi.org/10.1145/3449287
  • Kim, Sunnie S. Y., Q. Vera Liao, Mihaela Vorvoreanu, Stephanie Ballard, and Jennifer Wortman Vaughan. 2024. “‘I’m Not Sure, But…’: Examining the Impact of Large Language Models’ Uncertainty Expression on User Reliance and Trust.” In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 822–835. https://doi.org/10.1145/3630106.3658941
  • Lee, Hao-Ping (Hank), Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. 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. https://doi.org/10.1145/3706598.3713778
  • Sharma, Nikhil, Q. Vera Liao, and Ziang Xiao. 2024. “Generative Echo Chamber? Effect of LLM-Powered Search Systems on Diverse Information Seeking.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642459
  • Tankelevitch, Lev, Viktor Kewenig, Auste Simkute, Ava Elizabeth Scott, Advait Sarkar, Abigail Sellen, and Sean Rintel. 2024. “The Metacognitive Demands and Opportunities of Generative AI.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642902
  • Vasconcelos, Helena, Matthew Jörke, Madeleine Grunde-McLaughlin, Tobias Gerstenberg, Michael S. Bernstein, and Ranjay Krishna. 2023. “Explanations Can Reduce Overreliance on AI Systems During Decision-Making.” Proceedings of the ACM on Human-Computer Interaction 7 (CSCW1). https://doi.org/10.1145/3579605

Delegation to agents

  • Chen, Valerie, Alan Zhu, Sebastian Zhao, Hussein Mozannar, David Sontag, and Ameet Talwalkar. 2025. “Need Help? Designing Proactive AI Assistants for Programming.” In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3714002
  • Feng, K. J. Kevin, Kevin Pu, Matt Latzke, Tal August, Pao Siangliulue, Jonathan Bragg, Daniel S. Weld, Amy X. Zhang, and Joseph Chee Chang. 2026. “Cocoa: Co-Planning and Co-Execution with AI Agents.” In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3772318.3791673
  • He, Gaole, Gianluca Demartini, and Ujwal Gadiraju. 2025. “Plan-Then-Execute: An Empirical Study of User Trust and Team Performance When Using LLM Agents As A Daily Assistant.” In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713218
  • Park, Joon Sung, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. “Generative Agents: Interactive Simulacra of Human Behavior.” In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. https://doi.org/10.1145/3586183.3606763
  • Sarkar, Advait, and Ian Drosos. 2025. “Vibe Coding: Programming through Conversation with Artificial Intelligence.” In Proceedings of the 36th Annual Conference of the Psychology of Programming Interest Group (PPIG 2025). arXiv:2506.23253. https://arxiv.org/abs/2506.23253

LLMs as research tools

  • Agnew, William, A. Stevie Bergman, Jennifer Chien, Mark Díaz, Seliem El-Sayed, Jaylen Pittman, Shakir Mohamed, and Kevin R. McKee. 2024. “The Illusion of Artificial Inclusion.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642703
  • Hämäläinen, Perttu, Mikke Tavast, and Anton Kunnari. 2023. “Evaluating Large Language Models in Generating Synthetic HCI Research Data: A Case Study.” In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3580688
  • Park, Joon Sung, Carolyn Q. Zou, Jonne Kamphorst, Niles Egan, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Percy Liang, Robb Willer, and Michael S. Bernstein. 2024. “LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals.” arXiv:2411.10109. https://arxiv.org/abs/2411.10109
  • Wang, Angelina, Jamie Morgenstern, and John P. Dickerson. 2025. “Large Language Models That Replace Human Participants Can Harmfully Misportray and Flatten Identity Groups.” Nature Machine Intelligence 7 (3): 400–411. https://doi.org/10.1038/s42256-025-00986-z

Creativity and homogenization

  • Agarwal, Dhruv, Mor Naaman, and Aditya Vashistha. 2025. “AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances.” In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713564
  • Anderson, Barrett R., Jash Hemant Shah, and Max Kreminski. 2024. “Homogenization Effects of Large Language Models on Human Creative Ideation.” In Proceedings of the 16th Conference on Creativity & Cognition, 413–425. https://doi.org/10.1145/3635636.3656204
  • Chakrabarty, Tuhin, Philippe Laban, Divyansh Agarwal, Smaranda Muresan, and Chien-Sheng Wu. 2024. “Art or Artifice? Large Language Models and the False Promise of Creativity.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642731
  • Doshi, Anil R., and Oliver P. Hauser. 2024. “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content.” Science Advances 10 (28): eadn5290. https://doi.org/10.1126/sciadv.adn5290
  • Jakesch, Maurice, Advait Bhat, Daniel Buschek, Lior Zalmanson, and Mor Naaman. 2023. “Co-Writing with Opinionated Language Models Affects Users’ Views.” In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3581196

Accessibility

  • Alharbi, Rahaf, Pa Lor, Jaylin Herskovitz, Sarita Schoenebeck, and Robin N. Brewer. 2024. “Misfitting With AI: How Blind People Verify and Contest AI Errors.” In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility. https://doi.org/10.1145/3663548.3675659
  • Gadiraju, Vinitha, Shaun Kane, Sunipa Dev, Alex Taylor, Ding Wang, Remi Denton, and Robin Brewer. 2023. “‘I wouldn’t say offensive but…’: Disability-Centered Perspectives on Large Language Models.” In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 205–216. https://doi.org/10.1145/3593013.3593989
  • Gonzalez Penuela, Ricardo E., Jazmin Collins, Cynthia Bennett, and Shiri Azenkot. 2024. “Investigating Use Cases of AI-Powered Scene Description Applications for Blind and Low Vision People.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642211
  • Mack, Kelly Avery, Rida Qadri, Remi Denton, Shaun K. Kane, and Cynthia L. Bennett. 2024. “‘They only care to show us the wheelchair’: disability representation in text-to-image AI models.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642166

Education and learning

  • Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. 2025. “Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics.” Proceedings of the National Academy of Sciences 122 (26): e2422633122. https://doi.org/10.1073/pnas.2422633122
  • Kazemitabaar, Majeed, Justin Chow, Carl Ka To Ma, Barbara J. Ericson, David Weintrop, and Tovi Grossman. 2023. “Studying the Effect of AI Code Generators on Supporting Novice Learners in Introductory Programming.” In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3580919
  • Kosmyna, Nataliya, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, Iris Braunstein, and Pattie Maes. 2025. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv:2506.08872. https://arxiv.org/abs/2506.08872
  • Prather, James, Brent N. Reeves, Paul Denny, Brett A. Becker, Juho Leinonen, Andrew Luxton-Reilly, Garrett Powell, James Finnie-Ansley, and Eddie Antonio Santos. 2023. “‘It’s Weird That it Knows What I Want’: Usability and Interactions with Copilot for Novice Programmers.” ACM Transactions on Computer-Human Interaction 31 (1). https://doi.org/10.1145/3617367

Work and human-centeredness

  • Capel, Tara, and Margot Brereton. 2023. “What is Human-Centered about Human-Centered AI? A Map of the Research Landscape.” In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3544548.3580959
  • Collins, Katherine M., Ilia Sucholutsky, Umang Bhatt, Kartik Chandra, Lionel Wong, Mina Lee, Cedegao E. Zhang, et al. 2024. “Building Machines That Learn and Think with People.” Nature Human Behaviour 8 (10): 1851–1863. https://doi.org/10.1038/s41562-024-01991-9

Anthropomorphism, sycophancy, and relationships

  • Cheng, Myra, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, and Dan Jurafsky. 2026. “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence.” Science 391 (6792): eaec8352. https://doi.org/10.1126/science.aec8352
  • Cuadra, Andrea, Maria Wang, Lynn Andrea Stein, Malte F. Jung, Nicola Dell, Deborah Estrin, and James A. Landay. 2024. “The Illusion of Empathy? Notes on Displays of Emotion in Human-Computer Interaction.” In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642336
  • Fang, Cathy Mengying, Auren R. Liu, Valdemar Danry, Eunhae Lee, Samantha W. T. Chan, Pat Pataranutaporn, Pattie Maes, et al. 2026. “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study.” Research Square (preprint). https://doi.org/10.21203/rs.3.rs-8148142/v1
  • Zhang, Renwen, Han Li, Han Meng, Jinyuan Zhan, Hongyuan Gan, and Yi-Chieh Lee. 2025. “The Dark Side of AI Companionship: A Taxonomy of Harmful Algorithmic Behaviors in Human-AI Relationships.” In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713429

Footnotes

  1. We counted, by publication year, the papers in the CHI proceedings in OpenAlex (source S4363607743) whose title or abstract contains “large language model(s),” “LLM(s),” “generative AI,” or “generative artificial intelligence” (2026-11-01). The totals were 879 papers in 2023, 1,057 in 2024, 1,249 in 2025, and 1,702 in 2026. Counting titles only gives 0.9%, 8.0%, 10.7%, and 10.6%. ↩

  2. The arXiv preprint of the same paper (v1, October 2025) reports two experiments (N = 1,604) and “50% more,” which differ from the journal version. This note uses the numbers in the abstract of the journal version. ↩

Author: Shuichiro Ogawa (Design Researcher / Design Consultant) About me →