Source-linked AI summary
In-Browser Human-AI Co-Writing for Counterspeech: An Exploratory User Study of Agency, Perceived Response Quality, and Time
Tammy Schmidt, Luise Koch
TL;DR
Writing counterspeech can be demanding, while automated generation may compromise users’ voice and responsibility. This paper evaluates an in-browser co-writing assistant that combines user input with editable language-model suggestions. In an exploratory study, assisted responses had higher descriptive mean ratings across five quality proxies, while composition time increased.
Problem
Writing counterspeech can be cognitively and emotionally demanding, while evidence about interactive AI tools that support users without replacing their judgment, voice, or responsibility remains comparatively limited.
Source: §1 · p. 1Method
The in-browser assistant requires user input and offers three editable language-model suggestions, which users can select, revise, and post themselves.
Source: §1 · p. 1Results
Assisted responses had higher descriptive mean peer ratings across all five quality criteria, with differences ranging from 0.19 for relevance to 0.88 for empathy.
Source: §4.3 · p. 7Takeaways & Limitations
User input, choice among alternatives, and editing were associated in participants’ accounts with perceived agency and responsibility, while some empathy-oriented suggestions conflicted with personal voice.
Source: §6 · p. 9No participant posted counterspeech in a live environment, so the study does not measure real-world outcomes and supports provisional design insights rather than causal or generalizable effectiveness claims.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Writing counterspeech that challenges online hate speech can be cognitively and emotionally demanding, whereas fully automated generation may weaken users' sense of voice and responsibility. We present an in-browser AI-assisted co-writing tool that enables users to contribute their own ideas and offers three editable suggestions generated by a large language model using dynamically retrieved hate speech-counterspeech examples from the Multitarget CONAN dataset. We evaluated the tool in an exploratory within-subject study with ten participants, complemented by think-aloud observations and semi-structured interviews with half of the participants. Participants wrote responses to three self-selected hate speech instances without assistance and then repeated the task with the tool. Final responses were subsequently peer-rated for empathy, non-toxicity, relevance, specificity, and persuasiveness. Participants evaluated the interface positively, and qualitative accounts linked user input, selection among alternatives, and small edits to perceived agency and responsibility. Assisted responses received higher descriptive mean ratings across all five proxies, with the largest difference for empathy. Mean composition time increased from 2.52 to 3.88 minutes. Qualitative findings further revealed a tension between empathy-oriented outputs and users' desire to express legitimate anger or direct disagreement. We contribute exploratory design guidance for treating AI-assisted counterspeech as deliberative co-writing rather than automated text production.
1 Introduction
Online hate speech can harm targets and wider communities, while counterspeech offers a constructive alternative to removal or sanctions. This paper studies an in-browser co-writing assistant designed to support users without replacing their judgment, voice, or responsibility.
1 Introduction
- Counterspeech challenges hateful narratives with non-aggressive responses that can support targets, encourage reflection, and model civil interaction.Platforms commonly use removal or account sanctions, but these measures are contested and do not enable users to participate constructively.Source: §1 · p. 1
- Composing counterspeech can be time-consuming, emotionally demanding, and rhetorically difficult, despite large language models’ ability to generate fluent, contextually relevant responses.Prior work has focused more on generated outputs and model configurations than on interactive tools that preserve users’ judgment, voice, and responsibility.Source: §1 · p. 1
- The in-browser assistant requires user input, offers three editable suggestions, and leaves selection, revision, and posting to the user.The paper’s contributions include an interaction design and exploratory evidence on user experience, response ratings, and design implications.Source: §1 · p. 1
- The study asks how users experience the interface in terms of usability, agency, ownership, and voice, and how assisted and unassisted responses differ descriptively in quality proxies and composition time.The questions frame an exploratory evaluation rather than a test of real-world counterspeech effectiveness.Source: §1 · pp. 1–2
2 Related Work
Related work distinguishes perceived linguistic qualities from real-world counterspeech effectiveness, reviews automated generation approaches, and motivates interactive tools that support human judgment. The paper’s system combines dynamic example retrieval with browser-based co-writing.
2.1 Counterspeech Strategies and Evaluation
- Counterspeech effectiveness can mean changing a speaker’s stance, de-escalating an exchange, supporting targets, encouraging bystanders, or improving discussion tone.The paper treats empathy, non-toxicity, relevance, specificity, and persuasiveness as perceived linguistic qualities, not direct measures of real-world effectiveness.Source: §2.1 · p. 2
2.2 Automatic Counterspeech Generation
- Automated counterspeech generation uses fine-tuning, prompt engineering, and retrieval-augmented generation, with no universally superior method reported.Prompting large pretrained models has outperformed smaller fine-tuned models in some counterspeech settings.Source: §2.2 · p. 2
- The system extends static Multitarget CONAN in-context examples by dynamically retrieving semantically similar examples for each prompt.The study does not independently test whether dynamic retrieval improves generation quality.Source: §2.2 · p. 2
2.3 Interactive Tools for Counterspeech
- Prior interactive systems include practitioner support through monitoring and retrieval, and CounterQuill’s multi-step educational co-writing; this prototype instead places co-writing in the browser.The prototype emphasizes substantive user input, choice among three editable suggestions, and perceived agency, while prior work identifies concerns about authenticity and responsibility.Source: §2.3 · p. 2
3 System
The system is designed as a user-controlled co-writing tool, with ethical safeguards and explicit user initiation. Its browser interface sends requests through a backend and retrieval module to an externally hosted language model.
3.1 Interaction Design
- The interaction design treats the extension as a co-writing tool, using system steering, alternatives, and post-editing to support user agency, ownership, and responsibility.The design draws on a design space for intelligent writing assistants proposed by Lee et al. (2024).Source: §3.1 · p. 3
3.2 Ethics-by-Design
- The extension requires explicit user initiation, minimizes collected data, and leaves decision-making with users; users must consent to input transfer to an externally hosted model.Prompt safeguards aim to reduce harmful or toxic output, while reliance on a pretrained model and external API limits developer control over model behavior and downstream data processing.Source: §3.2 · p. 3
3.3 System Architecture
- The modular architecture comprises a browser frontend, backend service, retrieval module, and externally hosted LLM, as shown in Figure 1.The assistant uses llama-3.3-70b-versatile through the Groq API, with temperature 1.0 and provider defaults for other parameters.Source: §3.3 · pp. 3–4, Figure 1 · p. 4
- Users provide their own ideas or wording, choose a responder role, style, and approximate length, then compare, select, revise, and copy one of three suggestions.The Info page explains data processing, accountability, and appropriate use; users acknowledge these conditions before opening the Generator.Source: §3.3 · p. 3
- The backend validates inputs, retrieves examples, constructs model messages, sends them through the Groq API, and returns three parsed suggestions.Prompt components have intended roles and a priority order, with safeguards and default guidelines above user input and retrieved examples lowest; their individual effects were not measured.Source: §3.3 · pp. 3–4
- Retrieval uses Multitarget CONAN’s 5,003 hate speech–counterspeech pairs and adds the five nearest semantic matches to each prompt.The five-demonstration setting was an implementation choice, and its effect was not evaluated.Source: §3.3 · p. 3
4 Experiments
The exploratory study compared participants’ unassisted and AI-assisted counterspeech, using peer ratings, composition time, and measures of user experience and perceptions. Participants generally evaluated the prototype positively, while assisted responses had higher descriptive mean quality ratings but took longer to compose.
4.1 Experimental Setup
- The fixed-order within-subject study had ten participants, who chose three of eight hate-speech instances and completed unassisted responses before assisted ones.Repeated exposure and practice may affect comparisons; peer ratings assessed final edited responses, not isolated model performance.Source: §4.1 · pp. 4–5
4.3 Results
- Participants generally evaluated the prototype positively, with mean UEQ-S ratings of 2.2 for pragmatic quality, 1.7 for hedonic quality, and 1.9 overall.Given the sample size, results are an initial assessment of the interaction experience rather than a robust comparison with other products.Source: §4.3 · p. 6
- All five qualitative participants requested broader response styles, describing tension between consistently de-escalating suggestions and expressing their own emotional position.Participants also called suggestions generic, even though assisted responses had slightly higher mean peer ratings for specificity.Source: §4.3 · p. 7
- Participants linked contributing ideas, choosing among three suggestions, and revising a response to perceived agency and responsibility.Accounts were illustrative; the study collected no systematic interaction logs to compare selection and editing behavior.Source: §4.3 · p. 7
- Assisted responses had higher mean peer ratings across all five quality proxies, with differences ranging from 0.19 for relevance to 0.88 for empathy.Low inter-rater agreement means the condition means are descriptive patterns, not reliable evidence of improved response quality.Source: §4.3 · p. 7, Table 1 · p. 8
- Composition time increased from 2.52 minutes unassisted to 3.88 minutes assisted; nine of ten participants took longer with the extension.The mean increase was 1.36 minutes, or approximately 54%.Source: §4.3 · p. 8
5 Discussion
The exploratory findings connect user control with perceived agency while revealing tensions around empathy, response quality, and time. Fixed-order comparisons and inconsistent ratings constrain interpretation.
Interaction Design
- Participants linked substantive input, choice among suggestions, and editing with agency, although prior LLM familiarity was not measured and editing was limited.The authors suggest supporting more iterative, phrase- or argument-level revision.Source: §5 · p. 8
De-Escalation and Authenticity
- Empathy-oriented suggestions produced the largest descriptive increase in peer-rated empathy, yet some users found formulaic empathy toward hate speakers inauthentic or invalidating.The authors caution against treating empathy as a universal design objective and suggest offering varied non-toxic rhetorical strategies.Source: §5 · p. 8
Perceived Response Quality and Time
- User satisfaction may depend on suggestion quality and prompt configuration, which the study did not isolate through component ablations.Suggested future tests vary interface controls, system guidelines, and retrieved examples.Source: §5 · p. 8
- All five mean peer ratings were higher for assisted responses, but fixed order and low inter-rater agreement prevent attributing the differences to the tool.Raters interpreted the evaluation criteria inconsistently.Source: §5 · p. 8
- The extension increased mean composition time by 54%, but insufficiently fine-grained logs leave the reasons for the increase undetermined.Comparison of suggestions, experimentation, and revision are possible explanations; the prototype is a deliberative co-writing aid, not a demonstrated time-saving tool.Source: §5 · p. 8
6 Conclusion
The paper presents an in-browser AI-assisted counterspeech co-writing tool and explores its use with ten participants. The findings position the approach as deliberative co-writing rather than automated text production.
6 Conclusion
- The study explored an in-browser AI-assisted counterspeech co-writing tool with ten participants, examining user experience and descriptive comparisons of assisted and unassisted responses.The tool is presented as a co-writing aid rather than an automated text-production system.Source: §6 · pp. 8–9
- Assisted responses received higher descriptive mean peer ratings across all five quality proxies, while composition time increased and the fixed-order design precludes causal interpretation.Participants evaluated the interface positively, but low inter-rater agreement also limits interpretation of rating differences.Source: §6 · p. 9
- User input, choice among alternatives, and editing were associated with perceived agency and responsibility, while generic empathy-oriented suggestions sometimes conflicted with personal voice.The authors offer exploratory design guidance for supporting constructive, personally accountable responses while preserving agency and voice.Source: §6 · p. 9
Limitations
The study’s small, convenience-based sample and constrained procedure limit the robustness and transferability of its findings. It evaluates the complete prototype in a dataset-based setting, not real-world counterspeech outcomes or individual components.
Limitations
- The ten-person convenience sample comprised English-proficient adults recruited through the authors’ networks, limiting robustness and transferability to broader populations.Only half contributed think-aloud and interview data; the sample may not represent professional counterspeakers or users from different linguistic and cultural contexts.Source: Limitations · p. 9
- The fixed unassisted-then-assisted order, repeated stimuli, and participant-selected exposure prevent separating the extension from order effects and reduce comparability.Stimuli and retrieved examples came from CONAN rather than live conversations, lacking platform cues and surrounding social context; no live posting or real-world outcomes were measured.Source: Limitations · p. 9
- Evaluating the complete prototype prevents attributing findings to individual technical components or design choices.Results may differ substantially with different configurations or combinations of model, prompt, safeguards, retrieved examples, temperature, interface controls, or external service.Source: Limitations · p. 9
AI Assistance Disclosure
The authors used generative AI tools to assist with language editing, structural revision, and code refinement, and reviewed and revised all generated suggestions.
AI Assistance Disclosure
- The authors used OpenAI GPT-4/o, Google Gemini, and Claude for editing, revision, and code refinement, then critically reviewed and revised generated suggestions.The authors state that they take full responsibility for the submitted content.Source: AI Assistance Disclosure · p. 9
Offensive Content Warning
The appendix warns that its screenshots and study examples contain potentially distressing hateful comments. It also documents the extension interface and supplementary study materials.
Offensive Content Warning
- Appendix screenshots and study examples include hateful comments targeting people with disabilities, Jewish people, LGBTQ+ people, migrants, Muslims, People of Color, women, and others.The appendix warns that this content could be offensive and cause distress.Source: Offensive Content Warning · p. 12
- The appendix depicts the extension’s default and filled Generator pages, sample generated counterspeech suggestions, and its Info-page consent statements.The screenshots are identified as Figures A.1–A.4.Source: Figure A.1 · p. 13, Figure A.3 · p. 13, Figure A.2 · p. 13
- Supplementary tables document ethical considerations for the browser extension, Groq API study prompts and system message, hate-speech instances, exploratory authenticity items, and peer-rated response-quality criteria.Table A.1 adapts ethical considerations from Fjeld et al. (2020).Source: Table A.1 · p. 14, Table A.2 · p. 15, Table A.3 · p. 16
- Appendix materials also report inter-rater agreement, UEQ-S ratings, participant composition times, and a comparison with the UEQ-S benchmark.The agreement and UEQ-S summaries appear in Tables A.7–A.8; time and benchmark figures appear in Figures A.5–A.6.Source: Table A.7 · p. 18, Table A.8 · p. 18, Figure A.5 · p. 18