RecalibrateGPT: AI Fatigue Resilient Conversational Interfaces

AI summary

Overview

Research area: Human-Computer Interaction (cs.HC), specifically conversational user interfaces and human-AI interaction.

Technical level: Intermediate. The paper combines a qualitative user study, a quantitative usability evaluation, and a system implementation that uses sentence-transformer embeddings, cosine similarity, and KL divergence, but it explains these at a conceptual level.

Scope (one sentence): The paper presents RecalibrateGPT, a conversational interface that adds five single-click "cross-turn operators" acting on full conversation history, and reports two pilot studies with the same 12 advanced LLM users showing reduced perceived workload and high perceived usability.

The paper was published as arXiv:2609.00506v1 [cs.HC] on 01 Sep 2026, authored by Nikhil Wani (OpenThreads AI Research), and designated for the 39th Annual ACM Symposium on User Interface Software and Technology (UIST Adjunct '26), November 02–05, 2026, Detroit, MI, USA (DOI: 10.1145/3830397.3841834, ISBN 979-8-4007-2855-6/2026/11, CC BY 4.0).

What This Paper Is About

The authors argue that LLM interfaces trap users in a repetitive "type → read → retype" loop: people restate goals, correct drifts, and manage growing context across turns, which produces conversational AI fatigue, cognitive load, frustration, and eventually task abandonment. Their goal is to treat this as an interaction-flow problem rather than only a model-quality problem, and to remove it with one-click controls that recalibrate the whole conversation instead of just the last response. RecalibrateGPT introduces five such operators, each aimed at a distinct fatigue type, invoked through a single AssistiveButton across three palette layouts.

Key Contributions

1. A taxonomy of four conversational AI fatigue types derived from a formative qualitative study with 12 advanced LLM users: retyping fatigue (F1), scanning fatigue (F2), decision paralysis (F3), and context drift (F4). The paper states that single-coder thematic analysis identified these as recurring themes, each clustering around multi-turn interaction breakdown.

2. Five cross-turn operators — Anchor, Replay, Delta (Diff), Scope, and Steer — that act on the full conversation history with a single click and return structured JSON response templates. The paper claims these close two gaps in prior work: existing operators act only on a single-turn previous response, and none are designed around session-level conversational fatigue.

3. An interaction design for invoking them: a Universal AssistiveButton as the single entry point, and a three-way mode toggle that switches the operator palette between three layouts — Vertical (a pill-shaped sidebar left of each response), Arc (a fan menu radiating from the AssistiveButton), and Tablet (a persistent horizontal strip below each response). The authors state that no prior system defines operators acting on full conversation history across three geometric palettes tuned to fatigue context.

4. Two design objectives and two pilot studies grounding the system: DO1 (multi-turn response calibration) and DO2 (single-click corrective control), followed by a within-subjects quantitative evaluation reporting NASA-TLX and SUS results.

Main Findings

• Four fatigue types emerged from free-text reports. The formative study collected 96 reprompt examples (M=8.0 per participant) from 12 advanced LLM users (7M, 5F; ages 22–39, M=28.4, SD=4.6). Representative quotes: retyping fatigue, "I kept retyping the same constraints because the answer drifted" [P2]; scanning fatigue, "I was rereading to find the one line that actually mattered" [P11]; decision paralysis, "I kept asking follow-ups because I couldn't tell what to pick" [P9]; context drift, "I was fighting the interface, not solving the problem" [P3].

• Perceived cognitive workload dropped by half. In Study 2, participant-level mean NASA-TLX was M=2.7 under RecalibrateGPT versus M=5.4 under standard chat. The paper reports the drop occurred for all 12 participants, shown as a paired-slopes plot in Figure B.

• Perceived usability was high. Mean SUS was M=86.5 on the 10-item, 0–100 scale, which the paper places in the "excellent" usability band above 80.

• Participants uniformly preferred the system. The paper states participants consistently selected RecalibrateGPT as less fatiguing across the operator comparisons.

• Three operators were most often named useful. Anchor (n=4), Replay (n=3), and Delta (n=3) were the operators most frequently identified as useful during debrief.

• Layout preferences were split. Participants used Vertical n=5, Arc n=4, and Tablet n=3 as their preferred layout.

• The authors frame these as directional, not generalizable. Given the pilot scale, they explicitly report the results as directional evidence of feasibility rather than generalizable effects.

Methodology in Plain English

The work proceeded in two studies with the same 12 participants, all recruited by purposive sampling through social and academic networks. Inclusion criteria required daily LLM use (at least 5 days per week, at least 6 months sustained), active experience across at least 2 conversational AI platforms, English fluency self-rated at least 4/5, and self-reported use of LLMs for high-stakes information-seeking such as medical or legal questions. Users with exclusively single-turn, casual LLM experience were excluded.

Study 1 (formative, qualitative). Participants completed an online survey describing in free text their interaction frustrations and most-retyped corrections, producing 96 reprompt examples. Thematic analysis of these responses yielded the four fatigue themes and two design objectives.

Study 2 (quantitative, within-subjects). The same 12 participants evaluated a simulated healthcare scenario, the management of a new Type 2 diabetes diagnosis. For each of the five operators, they saw a side-by-side pair: Condition A, a standard-chat baseline, and Condition B, the RecalibrateGPT intervention — 10 screens in total. Each pair was shown in the participant's preferred layout, with operator order and condition presentation order randomized per participant to reduce order effects. After each comparison, participants picked the less fatigue-inducing interface and completed parallel 7-point NASA-TLX ratings for both conditions; the six NASA-TLX dimensions were averaged into an unweighted score per operator-condition, and the five operator scores were then averaged into one participant-level mean per condition. After all five comparisons, participants rated the RecalibrateGPT panel on the 10-item SUS, followed by a semi-structured interview. Sessions ran online, approximately 35 minutes per participant, in three phases: Onboarding (5 min), Paired Evaluation (20 min), and Usability and Debrief (10 min). Participants were compensated, the protocol was reviewed and approved by an internal ethics review process, and all participants provided informed consent.

Operator backend. The operators were implemented in Python over the OpenAI API (GPT-5.5), with all operators sharing sentence-transformer embeddings. Anchor computes goal alignment using cosine similarity between the embedding of the original goal and the latest response, and reorients low-similarity outputs. Delta compares response semantic distributions using KL divergence and flags goal-relevant removals. Replay summarizes the conversation history into established facts, open questions, and a next step. Scope clusters sentence embeddings from the latest response to surface subtopics, then expands the selection on a second tap. Steer computes unresolved gaps as goal sub-aspects whose maximum cosine similarity to any history turn falls below a threshold τ, then converts them into three tappable follow-up questions.

Mapping of operators to fatigue types: Anchor targets context drift (F4); Replay targets scanning fatigue (F2); Delta targets retyping fatigue (F1); Scope targets retyping fatigue (F1); Steer targets decision paralysis (F3). The paper notes that the Scope and Steer operator interfaces are included in supplementary material.

Why This Matters

Impact on research. The paper reframes conversational AI fatigue as an interaction-flow cost that interfaces can remove, rather than purely a model-quality issue. It also claims a specific gap in prior HCI systems — DirectGPT (Masson et al., 2024), Cells and Lenses (Kim et al., 2023), prompt-into-diagram work (Jiang et al., 2023), verifiable editing surfaces (Laban et al., 2024; Suh et al., 2023), and pipeline tools (Arawjo et al., 2023; Wu et al., 2022) — namely that the unit of manipulation in those systems and in commercial interfaces (ChatGPT, Claude, Gemini) remains a single response, leaving fatigue unaddressed. RecalibrateGPT proposes the conversation, not the response, as the unit of manipulation, and offers a fatigue taxonomy that later work can test and extend.

Real-world applications (as motivated by the paper):

• High-stakes clinical diagnosis, where the paper cites Rajashekar et al. (2024) and notes that accuracy matters (Wang et al., 2021; Wani et al., 2018).
• Legal interpretation, named alongside clinical diagnosis as a high-stakes domain where drift and retyping are costly.
• Everyday high-stakes information-seeking by LLM users who ask medical or legal questions, the inclusion criterion used to recruit participants.
• Any multi-turn session where users must repeatedly restate constraints, find one relevant line, choose among follow-ups, or fight against context drift.

Industry relevance. The paper argues that many user actions currently require full re-prompting by typing, which increases per-token inference cost and compounds the burden of every additional turn. Interfaces that convert the loop from "type → read → retype" into "type → read → click" could, in the authors' framing, reduce both user burden and token spend. The paper does not report measured cost savings.

Future Directions

• Proactive interfaces. The authors state future work will explore interfaces that dynamically surface the right operator as fatigue emerges, rather than waiting for the user to invoke it.
• Larger-scale validation. The paper calls for validating the approach through larger-scale studies, since the current results come from a pilot with 12 participants and are described as directional evidence of feasibility.
• Whether fatigue types generalize. The taxonomy of four fatigue types was derived from 12 purposively sampled advanced users in one formative survey; an open question is whether the same four themes and the same operator-to-fatigue mapping hold for other populations, including the single-turn casual users who were explicitly excluded.
• End-to-end effects across a full session. Study 2 measured per-operator paired comparisons within one simulated healthcare scenario; the paper does not report whether these gains persist across longer, real multi-turn sessions or translate into reduced task abandonment, which is the outcome the introduction motivates.

Target Audience

This paper benefits HCI and conversational-interface researchers and designers, particularly those working on multi-turn LLM interaction, prompt-reuse controls, and user fatigue in AI tools. Product designers and engineers building chat interfaces will find the operator palette layouts, the AssistiveButton placement logic, and the operator-to-fatigue mapping directly applicable. Practitioners in high-stakes domains such as clinical and legal AI may find the healthcare scenario design and the workload/usability instruments useful as a template. Readers looking for benchmark-style model-performance results will not find them here: the evaluation is a 12-participant pilot reporting NASA-TLX and SUS, and the paper reports no significance tests, effect sizes, or measured inference-cost savings.

Authors’ abstract

Large language models are powerful, but their interfaces often devolve into a type $\rightarrow$ read $\rightarrow$ retype loop, creating conversational AI fatigue, cognitive load, and eventual task abandonment. To mitigate this, we present RecalibrateGPT, a system introducing five cross-turn operators (Anchor, Replay, Delta, Scope, and Steer) that each target a distinct fatigue type, recalibrating LLM responses through a structured panel by acting on the full conversation history with a single click. Users invoke these operators through the AssistiveButton in one of three operator palette layouts: Vertical, Arc, or Tablet. We conducted two pilot studies with the same 12 advanced LLM users. An initial formative qualitative study identifies a taxonomy of four fatigue types (retyping, scanning, decision paralysis, and context drift) and derives two design objectives for RecalibrateGPT. A follow-up quantitative evaluation finds it reduces perceived cognitive workload by half (NASA-TLX = 2.7) at high perceived usability (SUS = 86.5), suggesting AI fatigue is not just a model-quality issue but an interaction-flow cost that interfaces can remove.