Source-linked AI summary
Probing How Users Interact with Turn-Level Design Frictions for AI Chatbots
Helen Weixu Chen, Katy Ilonka Gero
arXiv:2608.22427v1cs.HC
TL;DR
Chatbots can let users move from minimal prompts to ready-to-adopt text. This paper develops six turn-level friction probes and finds that all increased workload, task duration, and psychological ownership relative to a conventional chatbot.
Problem
Method
The paper develops six turn-level design probes organized around eliciting user contribution, restricting access to generated content, and reshaping system output.
Source: when familiar chatbot workflows are… · p. 2Results
All six friction probes increased workload, task duration, and psychological ownership relative to a conventional AI chatbot.
Source: ownership, recall, and recognition… · p. 2Takeaways & Limitations
The consequences of friction depend on where added effort is directed and how each constraint fits users’ goals and workflows.
Source: ownership, recall, and recognition… · p. 2The study leaves open how different friction mechanisms play out in longer, more naturalistic settings.
Source: and the model · p. 2
Abstract
from arXiv · showhide
AI chatbots can help people write faster, but they can also encourage overreliance by making it easy to turn minimal input into usable text. We study turn-level design friction: intentional constraints added to each chatbot exchange that slow, limit, or redirect how users request, access, or use model responses. We designed six friction probes, organized around three mechanisms: eliciting user contribution, restricting access to generated content, and reshaping system output. In a within-subject study with 24 participants, all six probes increased workload, task duration, and perceived ownership relative to a conventional AI chatbot, while their effects on recall and recognition were more selective. We further found that participants adapted to friction in different ways, and that the same constraint could support or obstruct involvement depending on users' goals and workflows.
1 Introduction
The paper examines turn-level chatbot frictions as a way to preserve user involvement, presenting six probes and findings that show both increased effort and ownership and varied effects on recall, recognition, and workflow.
1 Introduction
- The introduction identifies passive consumption of ready-to-adopt text from minimal prompts and a lack of comparative understanding of friction mechanisms, required effort, costs, and user adaptation.It frames design friction as an intentional way to slow, interrupt, or redirect an otherwise effortless workflow.Source: §1 · p. 2
when familiar chatbot workflows are deliberately disrupted.
- The paper defines turn-level friction as constraints within chatbot exchanges and develops six probes across three mechanisms: eliciting user contribution, restricting access, and reshaping output.Two probes represented each mechanism.Source: when familiar chatbot workflows are… · p. 2
- The probes were evaluated in a within-subject study of 24 participants completing short writing tasks that involved requesting, interpreting, and incorporating AI-generated material.The mixed-methods evaluation measured workload, task duration, and psychological ownership.Source: when familiar chatbot workflows are… · p. 2
ownership, recall, and recognition with interaction logs and semi-structured interviews.
- All six probes increased workload, task duration, and psychological ownership versus a conventional chatbot, while effects on recall, recognition, and workflow varied with effort placement and fit to users’ goals.The findings caution that adding effort alone did not make an interaction productive.Source: ownership, recall, and recognition… · p. 2
- Friction supported involvement when it kept users engaged in formulating, interpreting, or assembling responses, but obstructed them when effort shifted to satisfying constraints or managing the interface.Participants also adapted their workflows and redistributed work between themselves and the model.Source: ownership, recall, and recognition… · p. 2
and the model.
- The paper contributes a mechanism-based framing and empirical account of friction’s benefits, costs, and workflow consequences, offering preliminary evidence while proposing longer, naturalistic studies.Source: and the model · p. 2
2 Background and Related Work
Prior work links LLM assistance to reduced cognitive effort and changes in users’ reasoning, contribution, and ownership, while showing that interaction structure can support involvement. Design friction offers a way to redirect attention and preserve user contribution, motivating reusable constraints at the level of chatbot turns.
2.1 Impacts of LLM Assistance on User Involvement
- Cognitive offloading shifts mental work to external tools, while automation bias leads people to follow even incorrect AI recommendations rather than evaluate them [7].Source: §2.1 · p. 3
critical-thinking performance varying with the time available for the task.
- In writing, producing substantial text from few creative decisions may weaken expressive intent [34], while ownership and accomplishment track perceived contribution and control.Post-processing AI output can help users retain ownership, whereas a dominant chatbot role can distance them from the result.Source: critical-thinking performance varying… · p. 3
- LLM-assisted essay writing was associated with weaker neural connectivity than unassisted writing [33].Source: critical-thinking performance varying… · p. 3
unassisted writing and poorer immediate recall of participants’ own essays.
- Assistance effects are not inevitable: beginning independently can support critical thinking when time permits, and prompts to analyze, explain, and decide can support reflection and problem solving [58].Source: unassisted writing and poorer… · p. 3
2.2 Design Friction
- Design-friction interventions use obstacles or “microboundaries” to interrupt automatic behavior and create opportunities for reflection.Examples include goal reminders, time limits, lockouts, delayed access, feed redesign, and additional actions before continuing.Source: §2.2 · p. 3
- Across digital-wellbeing settings, friction has reduced distraction and mindless scrolling, supported intentional use and agency, and improved attention or memory, but can cause annoyance and frustration.Source: §2.2 · p. 4
- Prior AI interventions have used forcing functions and output cues to preserve judgment [7], or structured code, writing, and prompt workflows to elicit user effort [30] [28].Source: §2.2 · p. 4
- We explore friction as a reusable chatbot-interaction layer, comparing six turn-level probes that vary how users contribute to, access, interpret, or use model responses.Prior interventions are often tied to a particular task or behavior.Source: §2.2 · p. 4
3 Friction Design
The authors reviewed prior friction designs to identify recurring interaction mechanisms, then used those mechanisms to develop and refine six probes for evaluation.
3 Friction Design
- The authors developed and iteratively refined six friction probes guided by three recurring mechanisms: eliciting user contribution, restricting access, and reshaping system output.These probes were retained for evaluation.Source: §3 · p. 4
3.1 Deriving Friction Mechanisms from Prior HCI Work
- The review covered 24 HCI papers from the past five years across technology contexts and treated each distinct intervention as a design, yielding 73 unique designs.Source: §3.1 · p. 4
- The three mechanisms accounted for most designs; four designs instantiated two mechanisms, producing 77 design–mechanism entries, while incoherent cases remained under Other rather than forming a fourth mechanism.Source: §3.1 · p. 5
- Eliciting user contribution requires users to provide additional information, reasoning, judgment, or action.Source: M1. Eliciting user contribution · p. 5
M1. Eliciting user contribution
The supplied passages describe requiring more than the default interaction and delaying, gating, or conditionally withholding content or functionality.
M1. Eliciting user contribution
- One described constraint requires more than the default interaction would ordinarily require.Source: M1. Eliciting user contribution · p. 5
- Another constraint delays, gates, or conditionally withholds otherwise available content or functionality.Source: M2. Restricting access · p. 5
M2. Restricting access
M2 focuses on restricting access by changing the amount, form, organization, or communicative role of chatbot content.
M2. Restricting access
- M2 restricts access by changing the amount, form, organization, or communicative role of what the system provides.Source: M3. Reshaping system output · p. 5
M3. Reshaping system output
The six design probes were intended to redirect users’ effort toward the task, their reasoning, or model-generated content—not to discourage chatbot use. They implement three friction mechanisms through distinct constraints on user contribution, access to responses, and system output.
3.2 From Friction Mechanisms to Six Design Probes
- The authors selected six probes—two for each mechanism—favoring constraints that support task engagement over obstruction, bypassable designs, or effort spent managing the interface.They rejected output alterations that impaired readability and long delays that seemed disproportionate to short, iterative chatbot turns and risked feeling punitive.Source: became difficult to interpret and use · p. 5, merely making the interaction more… · p. 5
3.3.1 Eliciting User Contribution.
- Think Multiplier requires a prompt of at least 20 words, rejects prompts repeating the same sentence more than three times, and caps the reply at around half the input length.The output cap is intended to keep users as primary authors while the AI plays a subordinate role.Source: §3.3.1 · p. 6, Figure 2 · p. 6
- User Echo requires users to draft a tentative response before the model replies, after which the model gives constructive feedback and adds ideas or information.The design aims to preserve user involvement in planning and idea generation.Source: §3.3.1 · p. 6
3.3.2 Restricting Access.
- Hold Reveal blurs the response until users hold four keys near the keyboard’s corners, requiring both hands and making simultaneous text selection or copying impractical.Source: §3.3.2 · p. 6
links physical effort to attention and sense-making [4].
- Timer Lockout reveals model output for 15 seconds after users slide a timer, then re-locks it; the design also disables copy-paste.The timed window was intended to allow reading and gist formation while limiting prolonged skimming; disabling copying followed initial testing.Source: links physical effort to attention and… · p. 6, Figure 3 · p. 7
3.3.3 Reshaping System Output.
- Less is More limits each turn to one sentence, while Elenchus refuses to answer directly and instead asks questions that challenge the user’s premises.The concise-output design draws on minimalism and cognitive load theory to encourage active meaning-making.Source: §3.3.3 · p. 7, longer, document-level responses [21] · p. 7
3.4 Implementation
- The interface uses React (Vite), a Node.js backend, and GPT-4o via the OpenAI API; each friction condition is a modular React component.Source: §3.4 · p. 7
4 Experimental Method … indicate higher values for the interface).
The study evaluated six turn-level design frictions through short, bounded writing tasks in a within-subject design. Surveys, interviews, and memory tests assessed participants’ experiences and recall of their writing.
4.1 Participants
- The sample included 24 adults aged 18–44, and 23 reported prior experience using generative AI tools for writing.Reported tools included ChatGPT, Gemini, Claude, and DeepSeek.Source: §4.1 · p. 7
4.2 Apparatus
- Sessions ran in person or remotely, with audio, screen, on-screen interactions, exchanges, and task time recorded; each lasted about an hour.In-person sessions used a provided laptop in a quiet lab, and participants received a $30 Amazon gift card.Source: §4.2 · p. 8
4.3 Procedure, Design, and Tasks
- Twenty-four participants experienced all six frictional interfaces, ChatGPT, and a no-model writing condition in a counterbalanced within-subject study.The eight conditions were ordered using a balanced Latin square.Source: §4.1 · p. 7, §4.3 · p. 8
beginning the experimental tasks.
- Participants completed eight short writing tasks, producing 50–80-word paragraphs with each assigned interface before evaluating the condition.Prompts were adapted from IELTS-style tasks and covered diverse topics; a word-count check prevented proceeding until the minimum was met.Source: beginning the experimental tasks · p. 8
can be found in Appendix B.
- After each writing task, participants rated workload using NASA-TLX subscales and personal ownership of their writing on a five-point scale.They also completed a semi-structured interview about their experiences, perceived difficulty, ease of use, and impressions of ownership and control.Source: can be found in Appendix B · pp. 8–9, their experiences while those… · p. 8
each design influenced their sense of ownership and control during writing.
- After the interview, participants completed 50 arithmetic problems, then recalled their writing and took a sentence-recognition test.Recall used the eight writing topics as cues; recognition asked participants to identify their exact sentence among their own sentence and three GPT-4o-generated foils.Source: each design influenced their sense of… · p. 9, writing to memory encoding and… · p. 9
4.4 Data Analysis
- Quantitative interface differences were tested with Friedman tests and, after significant omnibus tests, pairwise Wilcoxon signed-rank tests with Holm correction.Interview transcripts and interaction logs were analyzed inductively; recall responses were segmented and independently scored by two researchers on a 0, 0.5, or 1 rubric, with disagreements resolved by discussion.Source: §4.4 · p. 9, indicate higher values for the… · pp. 9–10
5 Results
All friction conditions produced significantly higher overall workload than the ChatGPT baseline. Hold Reveal analyses used 23 participants because P3 could not complete that interface’s task; other conditions used all 24 participants.
5 Results
- All participants completed the study, but P3 could not complete the Hold Reveal task, so those analyses used 23 participants and other conditions used 24.Source: §5 · p. 10
5.1 Workload
- All friction conditions had significantly higher overall workload than the ChatGPT baseline (p≤.007).Overall workload was calculated as the unweighted mean of the five NASA-TLX subscales.Source: §5.1 · p. 10
- Overall workload was the unweighted mean of the five NASA-TLX subscales.Subscale-level results are reported in Appendix E.Source: §5.1 · p. 10
conditions increased workload over the Write by Yourself condition.
Friction conditions increased several workload subscales, while participants described added effort as either supporting their thinking or obstructing it. Physical demand and frustration were linked to specific interface actions and mismatches with users’ needs.
conditions increased workload over the Write by Yourself condition.
- All friction conditions increased Mental Demand (p≤.002) and Effort (p< .001); most also increased Physical Demand (p≤.011), while Performance differences were nonsignificant.Frustration was higher relative to ChatGPT; User Echo was the exception for Physical Demand (p= .06).Source: conditions increased workload over the… · p. 10, Figure 5 · p. 11
5.1.1 Added Cognitive Effort Could Support or Interfere with Thinking.
- Some participants said Think Multiplier and Elenchus helped them think critically and write in their own words.Think Multiplier prompted more writing from participants’ own thoughts, while Elenchus’s questions gave direction to their thinking.Source: §5.1.1 · p. 11
- Others said extra effort hindered reasoning or caused cognitive overload, shifting attention from developing ideas to coping with lost information.Participants described needing extra effort to achieve the same performance and being occupied with sudden loss of information in Timer Lockout and Hold Reveal.Source: §5.1.1 · p. 11
5.1.2 Physical Demand was Most Pronounced for the Access-restricting Interfaces.
- Hold Reveal and Timer Lockout increased Physical Demand by 5.63 and 5.69, respectively, relative to ChatGPT.Participants linked the demand to embodied actions for accessing output; one said Hold Reveal occupied both hands and hindered scrolling.Source: §5.1.2 · p. 11
5.1.3 Frustration Emerged When Friction Felt Obstructive or Disproportionate.
- Think Multiplier showed the largest frustration increase relative to ChatGPT (Δ𝑀= 7.58), closely followed by Timer Lockout (Δ𝑀= 7.54); User Echo showed the smallest (Δ𝑀= 2.29).Participants attributed frustration to different aspects of the friction.Source: §5.1.3 · p. 11
- Some participants found Think Multiplier’s effort disproportionate to its assistance, describing writing at length for a shorter response or writing for its own sake.One participant called supplying 100 words to get 50 back a poor trade-off.Source: §5.1.3 · p. 11
- Timer Lockout frustrated participants by interrupting access to generated responses, including before they had finished reading.Participants described having to reveal the response again as a waste of time.Source: §5.1.3 · p. 11
- Elenchus frustrated some participants who already knew their argument’s direction and wanted supporting information rather than its style of assistance.A participant wanted information or material to back up their thoughts.Source: §5.1.3 · p. 11
continued questioning frustrating rather than helpful.
All friction interfaces and Write by Yourself increased perceived ownership relative to ChatGPT; participants linked ownership to their contribution to ideas, expression, and thinking, not only to typing the final words.
5.2 Ownership
- All friction interfaces and Write by Yourself significantly increased ownership versus ChatGPT (M=2.00), with Write by Yourself producing the largest increase.On a 5-point scale, ownership ratings increased by 2.71 for Write by Yourself, 2.25 for Elenchus, 1.92 for Think Multiplier, 1.38 for Less is More, 1.21 for Timer Lockout, 1.04 for User Echo, and 1.00 for Hold Reveal.Source: §5.2 · p. 12
5.2.1 Definition of Ownership.
- Participants described ownership as involving whose ideas, expression, and thinking shaped the response, beyond who typed its final words.Source: §5.2.1 · p. 12
5.2.2 What Shaped Ownership Across Conditions.
- Participants associated greater ownership with contributing response substance and with interfaces that limited how much finished prose the model supplied.They described Elenchus as requiring them to develop the response’s substance, and described Think Multiplier and Less is More as limiting ready-made text.Source: §5.2.2 · p. 12
to develop more of the response themselves. … (Appendix E).
Friction interfaces increased task duration, while recall and recognition effects differed across conditions. Participants described ownership as arising either from contributing their own ideas or from relying less on model-written prose, and logs showed adaptations in interaction and writing workflows.
to develop more of the response themselves.
- User Echo had a comparatively small ownership increase but was second-most often named for greatest ownership, as participants valued stating their idea before model elaboration.Participants described ownership as preserving an idea’s origin and direction even when the model contributed to its eventual expression.Source: to develop more of the response… · p. 12
- Timer Lockout and Hold Reveal made model text harder to reuse, prompting users to reconstruct or write responses themselves; ownership could arise from reduced reliance rather than support for developing ideas.One participant described Timer Lockout by saying, “I just wrote it myself.”Source: to develop more of the response… · p. 12
5.3 Task Duration
- All friction interfaces took longer than baseline (p < .001), and each had higher average task duration than Write by Yourself; Think Multiplier averaged 3.41 minutes versus ChatGPT’s 3.15.Elenchus averaged 2.96 minutes; User Echo, Less is More, Timer Lockout, and Hold Reveal had broadly similar completion times.Source: §5.3 · p. 13
5.4 Recall Performance
- Write by Yourself had the highest average recall score (M=0.88), and most interfaces improved recall compared with ChatGPT (M=0.38), except User Echo.Source: §5.4 · p. 13
5.5 Recognition Performance
- ChatGPT recognition accuracy averaged 0.63; only Less is More (ΔM=0.38, 95% CI [0.17, 0.59], p= .02) and Elenchus (ΔM=0.33, 95% CI [0.14, 0.55], p= .04) significantly improved recognition.Remaining conditions showed no significant differences, though all averaged above ChatGPT; the authors suggest the recognition task may have been easier than free recall.Source: §5.5 · p. 13
5.6 Workflow Adaption Patterns
- Interaction logs showed four recurring adaptations in which participants changed how much they interacted with the model and redistributed writing work across turns.Source: §5.6 · p. 13
Participants often adapted by turning a paragraph-level writing … Our goal was to study turn-level friction in AI chatbot interactions. We used short
This section describes how participants adapted to turn-level friction by changing their prompts and contributions, compressing outputs, or adjusting how much they requested and whether they used AI. It discusses how friction’s effects depended on users’ goals and workflows, including whether it discouraged unnecessary reliance or supported work with needed information.
5.6.1 Breaking the Writing Task into Smaller Steps.
- Participants broke broad writing requests into stages; Less is More averaged 6.04 turns (SD=3.62), compared with ChatGPT (M=2.13, SD=1.30).User Echo (M=3.46, SD=2.54), Elenchus (M=3.42, SD=2.21), and Think Multiplier (M=3.38, SD=1.69) also involved more turns on average than ChatGPT.Source: Participants often adapted by turning… · p. 13
5.6.2 Bringing User Ideas Earlier into the Interaction.
- User Echo brought users’ ideas earlier into the exchange, where participants supplied substantive positions and then asked the model to elaborate, check, or integrate them.This introduced an intermediate user-created artifact rather than moving directly from a question to model-generated prose.Source: Another adaptation was to place… · pp. 13–14
5.6.3 Compressing Model Output to Fit Access Constraints.
- When output was hard to retain or revisit, participants shortened or segmented it, adapting its size and format more than the writing task’s substance.Examples included reducing a summary to a couple of sentences, requesting short bullets, or splitting information into individual sentences.Source: §5.6.3 · p. 14, response · p. 16
5.6.4 Working Around Friction to Restore Familiar Workflows.
- Some participants worked around constraints to restore familiar chatbot behavior, adding repeated characters or pasting prior model text without necessarily adding new input.Hold Reveal and Timer Lockout did not lead participants to send more prompts than with ChatGPT; participants instead often adjusted the size of the information requested before access was restricted again.Source: §5.6.4 · p. 14, response · p. 16
6 Discussion
- All six friction probes increased perceived ownership, workload, and task duration relative to ChatGPT, while memory effects were selective and participants adapted in several ways.Most probes improved recall, but only Elenchus and Less is More significantly improved recognition of participants’ own wording; adaptations included decomposing tasks, contributing ideas earlier, compressing responses, and working around constraints.Source: §6 · p. 14
increase involvement to when and how it can do so productively.
- The same friction could support or obstruct involvement depending on users’ goals: Socratic questioning impeded users who had formed an argument and wanted supporting information or later editing help.For users who had not formed a position, questioning could instead keep them from immediately outsourcing their reasoning.Source: increase involvement to when and how… · p. 14
supporting information, drafting, or revising.
- Participants’ design preferences suggest turn-level friction may work better as an optional, switchable repertoire than as a uniform rule across changing task goals.They described systems that could shift between reflective questions and direct assistance, such as questioning during ideation but providing examples on request.Source: supporting information, drafting, or… · p. 15
rule applied throughout an interaction.
- User Echo elicited early user contributions but did not significantly improve recall over ChatGPT, whereas Think Multiplier did; only Elenchus and Less is More significantly improved recognition.The authors caution that the probes differed in several ways, but suggest friction may need to preserve user involvement after model content is generated, not only require an initial contribution.Source: rule applied throughout an interaction · p. 15, response · p. 15
response
- With Hold Reveal and Timer Lockout, participants requested progressively shorter outputs, broke needed information into individual sentences, or wrote without generated responses.One participant summarized their Timer Lockout experience as, “I just wrote it myself.”Source: response · p. 16
- Access friction appeared to raise the cost of repeatedly seeking, revisiting, or reusing AI assistance rather than encouraging more iterative collaboration.Source: response · p. 16
- For discouraging habitual or unnecessary reliance, access costs may be useful; for helping users understand needed information, friction may work better during interpretation, formulation, or assembly of responses.Source: response · p. 16
- Reduced help-seeking did not necessarily mean deeper engagement: Timer Lockout and Hold Reveal could distract participants, who focused on maintaining or recovering access instead of developing ideas.Source: response · p. 16
Writing tasks were short and bounded. … F Usage Mirror
This section connects findings on turn-level frictions with prior design mechanisms that constrain or reshape chatbot interactions. It also presents the study prompts, interview and comparison materials, and a Usage Mirror for self-monitoring interaction metrics.
Writing tasks were short and bounded.
- Short writing tasks provided a bounded setting to examine requesting, interpreting, and incorporating output, but friction value depended on users’ goals and workflow stage.Participants noted that support needs changed across task stages; a friction useful during ideation could obstruct information seeking.Source: Our goal was to study turn-level… · p. 16
Static rather than adaptive friction.
- Static constraints could frustrate users when their needs changed, especially when moving between deeper engagement and efficient information access.The authors suggest exploring legible, adjustable constraints such as lighter-weight nudges; the Usage Mirror prototype points to one possible direction.Source: The current prototypes applied… · pp. 16–17
Optional frictions and user workarounds.
- Future work could examine stronger frictions as well as frictions users voluntarily leave on, including user-controlled toggles and when they activate them over days or weeks.The paper raises voluntary use as an alternative to escalating friction.Source: Most frictions can be avoided if the… · p. 17
8 Conclusion
- Across 24 participants, all six turn-level friction probes increased perceived ownership, workload, and task duration, while memory effects varied by design.Participants also adapted their workflows to different forms of friction.Source: §8 · p. 17
A Exploratory Corpus of Prior Friction Designs
- The exploratory corpus contained 73 unique designs from 24 papers; four designs instantiated two mechanisms, producing 77 design-mechanism entries.The corpus was used to derive the three friction mechanisms.Source: A Exploratory Corpus of Prior Friction… · p. 21, Table A.1 · p. 21
listed under both categories, resulting in 77 design-mechanism entries.
- Prior designs restricted access by requiring users to follow a moving dot before continuing to a target application, or by withholding AI suggestions until users made an initial unaided decision [7].Source: listed under both categories,… · p. 24
- A code interface provided hierarchical pseudocode instead of a complete code solution [30]; other interfaces used color-coded highlighting, inline explanations, or presented divergent outputs.Source: listed under both categories,… · pp. 25–26
- Social-media designs let users hide selected post types with feed filters and block recommended posts, while YouTube’s search-first Focus Mode hides homepage recommendations and removes related videos.Source: listed under both categories,… · p. 25
- Designs included a timer communicating how long unintended use continued, a prompt asking users to reflect on emotional state, and a proposed content-status indicator.A node-based interface visually traced prompts and outputs.Source: listed under both categories,… · pp. 26–27
B Writing Prompts
- Writing prompts include AI’s advantages and disadvantages in environmental protection, nuclear power for future energy needs, space spending amid poverty, and creativity versus academic knowledge in education.The supplied prompt list also includes universal basic income and facial recognition in public spaces.Source: B Writing Prompts · p. 27, §5 · p. 27
C Semi-structured Interview Questions
- The interview materials ask about participants’ overall experience with the models and what features they would include in a writing model of their own.Source: C Semi-structured Interview Questions · p. 28, §5 · p. 28
D Recall Scoring Pipeline
- A table presents examples used in the recall-scoring pipeline.Source: Table D.2 · p. 28
E Per-intervention Scores
- The per-intervention materials compare interventions with ChatGPT on workload, workload dimensions, ownership, task duration, and recall; recognition accuracy is also shown across interfaces.The supplied captions identify these measures but provide no scores or outcome comparisons.Source: Table E.3 · p. 28, Table E.4 · p. 29, Table E.5 · p. 29
F Usage Mirror
- Usage Mirror visualizes session duration and the character-count ratio of user input to LLM output, aiming to promote self-monitoring and voluntary self-correction without interrupting workflow.The metrics appear in the interface’s bottom panel.Source: Figure F.3 · p. 32