Detoxifying Toxic Communication: A Design Science Approach to Responsible AI
Hossein Arshadi Soufiani, Henry M. Kim, Hjalmar Turesson, Syed Mohammad Arham Noman, Anav Setia
cs.CY · cs.CL
Quick verdict
DSR artifact detects toxicity and rewrites it via DistilBERT + mT0-XL-ORPO; promising metrics but Twitter-only evidence, STA unreported, no real deployment.
Summary
This paper applies Design Science Research (DSR) to build a responsible-AI artifact that detects and rewrites toxic messages in workplace chat instead of deleting them. The pipeline fine-tunes DistilBERT and DistilRoBERTa classifiers on a binary adaptation of the Davidson et al. (2017) Twitter corpus (24,802 tweets, 80/10/10 split; 3 epochs, learning rate 5e-5, seed 42), then passes flagged text to the pretrained mT0-XL-Detox-ORPO sequence-to-sequence model, which generates non-toxic paraphrases under 10-beam search. Reported results: 0.965 accuracy and 0.979 F1 for toxicity detection, and a mean BERTScore F1 of 0.930 (precision 0.969, recall 0.948) for semantic preservation of rewrites, with style-transfer accuracy (STA) said to confirm outputs are non-toxic but without a reported number. The paper contributes design principles (content preservation, tone consistency) and a planned, not yet executed, organizational deployment evaluation. Key limitations: all quantitative evidence comes from Twitter data rather than workplace communication; only one combined metric pair is given for two classifiers with no per-class breakdown on an imbalanced dataset; no baselines are compared; the detoxification evaluation inputs are unspecified; and several citations appear mismatched to their claims. Code and models are stated to be publicly available on GitHub.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fine-tuned DistilBERT and DistilRoBERTa classifiers detect binary toxicity with 0.965 accuracy and 0.979 F1 on an adapted Davidson et al. (2017) dataset.
- The mT0-XL-Detox-ORPO module rewrites flagged messages with strong semantic preservation (mean BERTScore F1 0.930; precision 0.969, recall 0.948).
- A DSR-grounded detect-then-detoxify design with principles of content preservation and tone consistency operationalizes responsible AI better than delete/block moderation.
- This is the first DSR artifact to perform automatic detoxification while explicitly operationalizing responsible AI principles for workplace communication.
- Transformation-based moderation supports conversation continuity and respectful discourse — an assertion drawn from prior literature, not yet tested with this artifact in organizations.
Strengths
Unusually transparent training protocol
Section 3.1 reports the full fine-tuning recipe for the classifiers: Davidson et al. (2017) data collapsed to binary, 80/10/10 splits, 3 epochs, lr 5e-5, weight decay 0.01, stratified sampling with seed 42, ~30 minutes on a Colab T4 — plus a public GitHub repo with training scripts and evaluation notebooks (Section 3.3).
Honest boundary-setting on evaluation maturity
Section 3.1 explicitly labels the API layer and user-feedback loop as 'planned but unimplemented', and Section 3.4 labels the organizational deployment study (baseline → intervention → withdrawal) as conceptualized but not executed — no overclaiming of deployment evidence.
Clean DSR framing with instantiable design principles
Sensible modular architecture
Weaknesses
Core detoxification success metric reported without a number
Ambiguous, non-class-level classification results
No baselines on either side of the pipeline
Citation integrity problems
Workplace claims rest on zero workplace data
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This arXiv paper (cs.CY/cs.CL; the front matter notes an updated version in the HICSS-60 proceedings, January 2027) presents a Design Science Research (DSR) artifact for workplace communication: a detect-then-detoxify pipeline that flags toxic messages and rewrites them into non-toxic paraphrases rather than deleting them. Four central claims: (1) fine-tuned DistilBERT/DistilRoBERTa classifiers detect binary toxicity with 0.965 accuracy and 0.979 F1 (Section 3.3); (2) the mT0-XL-Detox-ORPO rewriter preserves meaning, with mean BERTScore F1 0.930 (precision 0.969, recall 0.948, Section 3.3); (3) the artifact operationalizes responsible-AI design principles — content preservation and tone consistency (Figure 1); (4) a novelty claim that no prior DSR artifact performs automatic detoxification while operationalizing responsible AI principles (Section 1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the detoxification system in this paper actually do?
It monitors workplace chat messages, flags toxic ones using fine-tuned DistilBERT/DistilRoBERTa classifiers, and then rewrites the flagged text into a polite, semantically equivalent paraphrase with the mT0-XL-Detox-ORPO model — replacing delete-or-block moderation with transformation.
How well does it perform?
On a Twitter-derived dataset (Davidson et al. 2017), the classifiers report 0.965 accuracy and 0.979 F1, and the rewrites score a mean BERTScore F1 of 0.930 for meaning preservation. However, the style-transfer-accuracy value is not reported in the paper, there are no baseline comparisons, and no real-world deployment or user study has been conducted.
Has the system been tested in a real company?
No. The organizational evaluation (measuring toxicity baseline, applying the intervention in tools like Slack or Teams, then withdrawing it) is described only as a future plan; all quantitative evidence comes from the public Twitter dataset.
Is the code and data available?
Yes — the paper points to a public GitHub repository (github.com/thearhamn/detox_org_com) said to contain training scripts, model configurations, the dataset and evaluation notebooks; the classifiers and the detoxification model are built on publicly available datasets and HuggingFace checkpoints.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–5 appear in the extracted text only as captions; any per-model classification metrics in Figure 4 and score distributions in Figure 5 could not be inspected, so reported-number completeness for those visuals is unverified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.