PaperlayerView exact revision on arXiv

Source-linked AI summary

Who Thinks First? Designing Productive Friction with Engage-to-Unlock GenAI

Xiaotian Su, Laura Rimell, Jiazheng Li, Amal Rannen-Triki, Ulrich Paquet, Lisa Anne Hendricks, Rida Qadri, Daphne Ippolito, Piotr Mirowski

arXiv:2610.01518v1cs.HC

TL;DR

Frictionless GenAI access may arrive before writers develop their own ideas, raising questions about when to make generative assistance available. The paper tests Engage-to-Unlock, which releases capabilities in response to task engagement, and finds effort shifted toward writing without longer overall task time, alongside more AI prompts.

  • Problem

    Frictionless generative access may lead to cognitive offloading before writers develop their own ideas, while the value of engagement-responsive access remains to be examined.

    Source: §1 · pp. 1–2

  • Method

    A controlled experiment compared Human-Only, Standard Chatbot, engagement-based Engage-to-Unlock, and time-matched unlocking, followed by unaided evaluation of AI-generated passages.

    Source: §1 · p. 2

  • Results

    Engage-to-Unlock participants spent more time writing and less time evaluating without significantly greater overall task duration, and submitted more prompts than users in other AI-assisted conditions.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The findings support engagement-responsive AI access as a form of productive friction that can retain active AI use while concentrating effort earlier in the workflow.

    Source: §1 · p. 3

    A timer constraint affected unlock delivery for 14 of 94 Time-Matched Unlock participants, and that group was recruited in a separate phase, limiting how precisely the comparison isolates contribution-based access.

    Source: §5.3 · p. 20

Abstract

from arXiv · show

Generative AI can support writing, but frictionless access may cause cognitive offloading before users develop their own ideas. We introduce Engage-to-Unlock, a productive-friction mechanism that unlocks generative capabilities after users meaningfully engage with the task. In a controlled experiment (N = 398), participants completed a writing task under one of four conditions: Human-Only, Standard Chatbot, Engage-to-Unlock, or Time-Matched Unlock, which matched unlock timing to Engage-to-Unlock participants but independent of users' engagement, then evaluated passages for evidence and inferential errors. Results show that Engage-to-Unlock redistributed effort across tasks: participants spent more time writing and less time evaluating, without increasing overall task duration. They also submitted more prompts than in other AI-assisted conditions and showed the highest accuracy-per-time evaluation efficiency across conditions. These findings suggest that designing GenAI access to encourage early human engagement may provide a productive form of friction, while retaining active AI use and efficient downstream evaluation.

1. Introduction

The paper proposes engagement-responsive access to generative AI to preserve users’ early independent formulation, then tests whether it adds value beyond immediate or time-matched access.

1. Introduction

  • The motivation is that generative assistance arriving before writers develop their own ideas may shape their positions and reduce opportunities for independent engagement.Prior work cited in the introduction describes possible effects on expressed views, perspectives, and voice.Source: §1 · pp. 1–2
  • Engage-to-Unlock participants spent more time writing and less time evaluating without significantly increasing overall task duration, while submitting more prompts than users in other AI-assisted conditions.The comparison with Time-Matched Unlock also found faster evaluation, comparable error-identification accuracy, and stronger error-classification performance.Source: §1 · pp. 2–3
  • Compared with Time-Matched Unlock, Engage-to-Unlock participants submitted more prompts and completed evaluation faster, with comparable error-identification accuracy and stronger error-classification performance.Time-Matched Unlock matched release timing independently of the current participant’s engagement.Source: §1 · pp. 2–3
  • Engage-to-Unlock initially restricts generative features and releases them in response to users’ task engagement.In argumentative writing, the implementation responds to evidence that writers have developed a position and supporting rationale.Source: §1 · pp. 2–3
  • A time-matched control distinguishes engagement-responsive friction from delay alone, while Standard Chatbot provides a comparison with immediate AI access.The study also compares unaided Human-Only writing with the AI-assisted conditions.Source: §1 · pp. 2–3

2. Related Work

Related work frames GenAI as both a means of delegating cognitive work and a source of questions about effort, judgment, and output diversity. Prior friction designs motivate testing access that responds to users’ actual task engagement.

2.1. Generative AI, Cognitive Delegation, and Knowledge-Work Outcomes

  • GenAI can support productivity while shifting effort toward verification or integration; the concern is offloading formative work before users develop their own understanding or position.The cited literature also reports that incorrect AI outputs can be over-relied upon.Source: §2.1 · p. 3
  • Research on AI-assisted creativity reports varied effects across individual quality and cross-user diversity, motivating evaluation of both dimensions.Findings include reduced lexical or semantic distinctiveness in some settings and increased collective semantic diversity in another.Source: §2.1 · p. 4

2.2. Frictions in Human–AI Interaction

  • Designed friction can require reflection or reasoning before AI assistance, while delayed assistance may preserve independent thought but also impose usability costs or unnecessary delay.The cited timing findings indicate that earlier access can help under time pressure but may impair critical thinking when time is sufficient.Source: §2.2 · p. 4
  • Unlike interventions ending at a preset time, task stage, or prescribed action, Engage-to-Unlock progressively releases capabilities based on observable engagement.Its comparison with time-matched unlocking tests whether engagement-responsive friction adds value beyond delay alone.Source: §2.2 · p. 4

3. Experiment

The experiment compares four GenAI access conditions in argumentative writing and subsequent unaided passage evaluation. Its methods define the intervention, participant groups, tasks, and planned measures and comparisons.

3. Experiment

  • A four-condition, between-participants study examined AI access during argumentative writing and subsequent evaluation of AI-generated content.Human-Only was the unaided baseline, Standard Chatbot offered immediate generation, and the two unlock conditions delayed generation using contribution-based or matched-time access.Source: §3 · p. 5

3.1. Experimental Design and Conditions

  • Teach-me used a pedagogical prompt and Tell-me was the standard chatbot; Engage-to-Unlock released Tell-me after a position and linked argumentative development were complete.Time-Matched Unlock replayed a Phase 1 participant’s unlock time regardless of the current writer’s essay progress.Source: §3.1 · pp. 5–6

3.2. Participants

  • The final analysis included 398 participants recruited in two phases; the three Phase 1 conditions were randomly assigned, while all Phase 2 participants received Time-Matched Unlock.Condition sample sizes were Human-Only 99, Standard Chatbot 98, Engage-to-Unlock 107, and Time-Matched Unlock 94.Source: §3.2 · p. 6

3.3. Tasks and Materials

  • Participants wrote an essay of at least 200 words from four source summaries, then judged eight AI-generated passages based on those same sources without AI assistance.Passages included sound and flawed items; flaws involved evidence errors or reasoning errors, and item soundness was counterbalanced.Source: §3.3 · p. 7

3.4. Procedure

  • The procedure included surveys around writing and evaluation, disabled clipboard transfers to and from external windows during writing, and allowed participants to work at their own pace.Both tasks had suggested durations of 20 minutes, and pilot data were excluded from the main analysis.Source: §3.4 · p. 7

3.5.1. Inferential Framework and Planned Comparisons

  • Planned contrasts compared Engage-to-Unlock with Standard Chatbot and Time-Matched Unlock, using Holm correction within each outcome’s two-comparison family.Participants passing attention checks were analyzed as assigned, including Time-Matched Unlock sessions affected by the reported delivery deviation.Source: §3.5.1 · p. 8, §3.5.2 · p. 8

3.5.3. Writing Outcomes

  • Writing measures covered essay length, discourse structure, semantic dispersion, lexical frequency, and source reuse; surveys assessed agency, ownership, workload, and workflow disruption.Semantic dispersion estimates were descriptive rather than formal tests of between-condition differences or equivalence; AI-only essays served as a descriptive reference, not a randomized condition.Source: §3.5.3 · pp. 8–9

3.5.5. Passage Evaluation Performance

  • Passage evaluation measured overall accuracy and conditional error-classification accuracy, while duration and interaction logs captured task time and writing behavior.The conditional classification outcome concerns correctly detected flaws and does not by itself establish improvement across all flawed passages; prompt analysis distinguished Teach-me from Tell-me queries.Source: §3.5.5 · p. 10, §3.5.6 · p. 11

4. Results

Across the writing conditions, Engage-to-Unlock participants used AI more actively, while essay content and downstream writing characteristics showed a mix of similarities and differences. Engagement-responsive unlocking also differed from simply matching unlock timing: most Time-Matched Unlock participants received access before meeting the engagement criterion.

4.1.1. Model Unlocking Status.

  • Among the 60 Time-Matched Unlock participants who received an unlock, 38 (63.3%) had not met Engage-to-Unlock’s discourse criteria when access became available.Thus, matching unlock timing did not reproduce the engagement state that triggered unlocking in Engage-to-Unlock.Source: §4.1.1 · p. 11

4.1.2. AI Interaction and Text Integration

  • Engage-to-Unlock participants submitted more prompts than those in Time-Matched Unlock or Standard Chatbot, while their verbatim AI-text coverage was lower than Standard Chatbot’s.Teach-me accounted for most prompts in both unlock conditions, and Tell-me availability did not necessarily lead to its use.Source: §4.1.2 · pp. 12–13

4.2.1. Semantic Content and Diversity

  • Exploratory PC1 analysis separated the five corpora; Engage-to-Unlock essays were closer to Human-Only than Standard Chatbot essays, while Engage-to-Unlock did not significantly differ from Time-Matched Unlock.Participant-condition semantic dispersion estimates were similar, while AI-only essays had substantially lower estimated dispersion than Human-Only essays.Source: Figure 2 · p. 12, §4.2.1 · pp. 13–14

4.2.2. Source Reuse and Lexical Patterns.

  • Verbatim reuse of shared readings was low, providing little support for direct copying as an explanation of semantic overlap; shared ideas or paraphrased source content could still contribute.AI-assisted essays also showed descriptive word-use similarities to the AI-only reference, which the analysis could not distinguish from effects of shared topics and sources.Source: §4.2.2 · p. 14

4.2.3. Essay Length and Argumentative Structure

  • Essay word counts did not differ significantly across experimental conditions (F(3, 216.55) = 1.43, p = .234; ω2 = .004, η2 = .012), despite average lengths ranging from 277.11 words in Human-Only to 308.32 in Standard Chatbot.The AI-only corpus had a median of 496.0 words (SD = 38.34).Source: §4.2.3 · pp. 14–15

4.2.4. Discourse Elements Diversity and Density.

  • Total discourse-element counts differed by condition, and Engage-to-Unlock yielded significantly fewer counterclaims than Standard Chatbot, but not Time-Matched Unlock.The remaining planned contrasts for discourse outcomes were not statistically significant.Source: §4.2.4 · p. 15

4.3. Passage Evaluation Performance

Engage-to-Unlock participants classified detected flaws more accurately than Standard Chatbot participants, but this did not translate into significantly better complete diagnosis across all flawed passages. They evaluated passages faster and more efficiently, while spending more time writing without increasing overall task duration.

4.3.1. Judgment and Error-Classification Accuracy

  • 90.43% versus 78.12%: Engage-to-Unlock participants more accurately classified flaws they had detected than Standard Chatbot participants, but complete diagnosis across flawed passages did not significantly improve.Overall judgment accuracy ranged from 61.61% to 66.36%, with no significant differences between conditions.Source: §4.3.1 · pp. 15–16

4.3.2. Evaluation Time and Efficiency

  • Engage-to-Unlock participants completed passage evaluation in 17.41 minutes on average, 4.80 minutes less than Standard Chatbot and 2.88 minutes less than Time-Matched Unlock.They showed higher values than Standard Chatbot across all three accuracy-per-time measures; versus Time-Matched Unlock, only conditional error-classification efficiency differed significantly.Source: §4.3.2 · p. 16, Figure 4 · p. 17

4.4. Time Investment and Subjective Experience

  • Engage-to-Unlock participants spent 4.77 minutes longer writing and 4.80 minutes less evaluating than Standard Chatbot participants, a cross-task difference-in-differences of −9.57 minutes.Total active task duration did not differ significantly across conditions; the corresponding difference-in-differences versus Time-Matched Unlock was not significant.Source: §4.4 · pp. 16–18
  • Engage-to-Unlock participants reported greater effort than Standard Chatbot participants, but not Time-Matched Unlock participants; perceived workflow disruption did not differ across AI-assisted conditions.There was no evidence that Engage-to-Unlock disrupted writing flow more than the other AI access conditions.Source: §4.4 · pp. 18–19

5. Discussion

Engage-to-Unlock shifts effort toward early human contribution while preserving AI interaction, and the discussion frames capability-specific access as an alternative to uniform friction or all-or-none AI access. The authors note limits in isolating some comparisons and in generalizing beyond a short-term writing experiment.

5.1. Designing Where Humans Think: From Effort Reduction to Effort Reallocation

  • Evaluation is not automatically productive effort: judging AI output requires sufficient task and domain knowledge.The discussion cautions against treating downstream evaluation as valuable solely because it involves additional human work.Source: §5.1 · p. 19
  • Engage-to-Unlock participants completed evaluation faster without lower evaluation performance, potentially because earlier engagement front-loaded sensemaking needed to judge AI output.The authors interpret Standard Chatbot participants’ additional evaluation time as possibly reflecting work deferred from initial writing.Source: §5.1 · p. 19
  • The design goal is to allocate effort where independent reasoning, authorship, or judgment matters, rather than maximize automation or human effort.The authors contrast this with cognitive forcing functions that add friction at AI-assisted judgment.Source: §5.1 · p. 19

5.2. Beyond Binary AI Access: Capability Boundaries in Human–AI Collaboration

  • Prior studies report both impaired and improved later unassisted performance, depending on how AI assistance is provided, motivating capability-specific rather than binary access.The cited examples include safeguards, writing assistance, and AI examples intended to support learning.Source: §5.2 · pp. 19–20
  • Engage-to-Unlock restricts direct generation while preserving AI-supported interaction; participants reported greater effort than Standard Chatbot, with no detected increase in overall task time.This separates access to a specific capability from withholding AI altogether or changing the underlying model.Source: §5.2 · p. 20

5.3. Limitations and Future Work

  • The comparison with Time-Matched Unlock is limited by unlock-delivery deviations and separate-phase recruitment, and the study is short-term and writing-focused.A timer constraint affected unlock delivery for 14 of 94 Time-Matched Unlock participants; the authors call for research on longer-term effects and other domains.Source: §5.3 · p. 20

6. Conclusion

In a 398-participant argumentative-writing experiment, Engage-to-Unlock made generative assistance available after writers stated a position and supporting argument. Compared with Standard Chatbot, it shifted effort toward writing and was associated with faster passage evaluation, more prompting, and some differences in error classification.

6. Conclusion

  • Compared with Standard Chatbot, Engage-to-Unlock participants spent more time writing and reported greater effort, then completed passage evaluation faster without significant differences in frustration or temporal demand.The unlock followed writers’ expression of a position and supporting argument.Source: §6 · pp. 20–21
  • Among correctly detected flaws, Engage-to-Unlock participants classified errors more accurately than Standard Chatbot participants, while the comparison with Time-Matched Unlock was not statistically significant.Source: §6 · p. 21
  • Engage-to-Unlock participants submitted more prompts than those in the other AI-assisted conditions, mainly in Teach-me mode, but fewer than half of those who gained Tell-me access subsequently prompted it.The authors suggest the prompting pattern may indicate greater engagement with conversational guidance.Source: §6 · p. 21
  • The findings motivate studying contribution-based AI access and how capability boundaries shape engagement and effort across tasks.The authors identify longer-term effects on learning, independent judgment, and skill retention as areas for future work.Source: §6 · p. 21

B. Task Duration Distributions

Figure 6 presents task-duration distributions across conditions.

B. Task Duration Distributions

D. AI Assistant System Prompts

The appendix reproduces the writing assistant’s system prompts, including its shared role and principles and separate policies for guidance and generation. It also documents figures describing task durations and Human-Only essay word frequencies.

D. AI Assistant System Prompts

  • The assistant used a shared foundational prompt and separate policies for Guidance (Teach-me) and Generation (Tell-me) modes.The system was powered by gemini-3.1-pro-preview.Source: D. AI Assistant System Prompts · p. 26

D.1. Shared Assistant Role and Principles

  • The same assistant followed the currently active mode, and prior responses remained attributed to that assistant even when modes changed.Participants could switch modes when available.Source: §D.1 · pp. 26–27
  • Figure 6’s panels show writing, passage-evaluation, and combined task-duration distributions; no significant condition difference was detected in combined time.The figure uses violins, participant dots, boxplots, and mean markers.Source: Figure 6 · p. 27
  • The assistant could access the live essay draft, conversation history, and task sources, while its principles required concise, transparent responses and source-grounded claims.It was also instructed to preserve participants’ intended meaning and distinguish their ideas from sources and AI suggestions.Source: §D.1 · p. 27
  • Figure 7 plots mean normalized frequencies of the 85 most frequent Human-Only content words, with shaded bands showing ±1 SEM.Frequencies are occurrences per 1,000 words, averaged across essays.Source: Figure 7 · p. 28

D.2. Guidance Mode (Teach-me) System Prompt

  • Teach-me guidance supports participants in articulating, clarifying, and reflecting on their thinking without supplying task-specific argumentative content or essay text.Source: §D.2 · p. 28
  • Teach-me also bars leading questions, evaluation of participants’ ideas, artifact production, and hidden scaffolding toward a position or argument.Source: §D.2 · pp. 28–29

D.3. Generation Mode (Tell-me) System Prompt

  • Tell-me responds directly with task-specific ideas, reasoning, source-based information, organization, revisions, and draftable prose, grounded in provided sources.Its objective is to help participants advance and complete the argumentative writing task through direct content-level assistance.Source: §D.3 · p. 29

E. Sample Conversation Transcripts

The section presents participant–assistant conversation records from Task 1 and explains how participant codes identify conditions and assigned user IDs.

Participant: B151 | Condition: Engage-to-Unlock

B151 used Teach-me mode to examine the Swedish trial’s methods and question its sleep and stress measurements. After meeting the unlock criteria, the participant switched to Tell-me but submitted no prompts and completed the essay without AI-generated prose.

Loading 2610.01518v1…