Elamur Research⌕
☰
home / papers / 2026 / Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding

Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding

Paper · 2026 · Wajdi Zaghouani

Summary

LLMs can open deliberation to far more people than turn-taking and limited facilitators normally allow. Systems such as the Habermas Machine already write group statements that participants often prefer. The author warns that LLM-only mediation can flatten minority views, push toward agreement for its own sake, and weaken legitimacy when people cannot see or challenge how they were represented.

The paper proposes a conceptual "symbiotic" human–AI blueprint with three layers:

  • AI observation that amplifies diversity,
  • facilitation in which every clause of a synthesized statement is traced to its contributors,
  • final ratification kept in human hands, with veto rights.

The specification also includes:

  • graded coverage, diversity and erasure metrics that weight minorities by salience rather than size;
  • a provenance pipeline combining cross-encoder similarity with knockout regeneration;
  • inference-time trade-off control via PARM;
  • equity-aware contestation workflows and adversarial robustness tests;
  • an ablation-based evaluation plan that relies on human raters.

No experiments are reported. The contribution is a precise, testable specification for accountable deliberation technology.

Key points

  • The framework defines eight requirements (R1–R8), among them diversity retention, contestability, provenance, reframing support, human primacy, robustness and governance fit.
  • The motivating evidence comes from two studies:
    • The Habermas Machine was tested with over 5,700 participants, and more than 56% preferred its statements.
    • In DeliberationBench, simple best-of-N selection beat the best multi-LLM protocol roughly sixfold (p < 0.01) at lower cost, which argues for parsimonious aggregation.
  • CounselBench (2,000 expert evaluations) shows that LLM judges overrate outputs and miss problems at the span level. The author therefore recommends human raters with span-level auditing.
  • Adding knockout regeneration raises costs by about 50–60% over similarity-only provenance, while facilitator workload grows sublinearly.
  • Stated limitations include:
    • no implementation yet;
    • annotation studies still needed to calibrate metric thresholds;
    • provenance remains open to gaming (robustness tests draw on DAMASHA);
    • field deployment issues such as dropout and multilingual quality.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Northwestern University in QatarDoha · Qatar39 works · 148 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Social Sciences · 1,954 worksSafety Research · 626 worksPolitical Science and International Relations · 206 works

Cite

Cite as Wajdi Zaghouani. 2026. Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding. Proceedings of The 2nd Workshop on Language-driven Deliberation Technology.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/37crtjgcefv3

@inproceedings{zaghouani2026accountable,
  title = {{Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding}},
  author = {Wajdi Zaghouani},
  booktitle = {Proceedings of The 2nd Workshop on Language-driven Deliberation Technology},
  year = {2026},
  doi = {10.63317/37crtjgcefv3},
  url = {https://doi.org/10.63317/37crtjgcefv3}
}

References

24 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Michael Henry Tessler, Michiel A. Bakker, Daniel Jarrett, et al. 2024. AI can help humans find common ground in democratic deliberation. Science, 386(6719): eadq2852. doi: 10.1126/science.adq2852. — [doi:10.1126/science.adq2852](https://doi.org/10.1126/science.adq2852) find
  2. Scott E. Page. 2025. Everyone, everywhere, all at once: LLMs and the new physics of collective intelligence. Collective Intelligence, 4(3). doi: 10.1177/26339137251367733. — [doi:10.1177/26339137251367733](https://doi.org/10.1177/26339137251367733) find
  3. Damon Centola. 2022. The network science of collective intelligence. Trends in Cognitive Sciences, 26(11):923–941. doi: 10.1016/j.tics.2022.08.009. — [doi:10.1016/j.tics.2022.08.009](https://doi.org/10.1016/j.tics.2022.08.009) find
  4. Alessandro Volpe. 2025. Toward an artificial deliberation? On Google DeepMind’s Habermas Machine. Ethics and Information Technology, 27(4):45. doi: 10.1007/s10676-025-09854-1. — [doi:10.1007/s10676-025-09854-1](https://doi.org/10.1007/s10676-025-09854-1) find
  5. Jürgen Habermas. 1984. The Theory of Communicative Action, Volume 1: Reason and the Rationalization of Society. Beacon Press, Boston. find
  6. Nicolás Palomo Hernández. 2025. Towards automating deliberation? The idea of deliberative democracy embedded in Google’s Habermas Machine. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), 8(2), pages 1951–1960. doi: 10.1609/aies.v8i2.36687. — [doi:10.1609/aies.v8i2.36687](https://doi.org/10.1609/aies.v8i2.36687) find
  7. James S. Fishkin. 2011. When the People Speak: Deliberative Democracy and Public Consultation. Oxford University Press. find
  8. Alvaro Oleart and Nicolás Palomo Hernández. 2025. Why AI technosolutionism harms democracy and deliberation. Journal of Deliberative Democracy. doi: 10.16997/jdd.1839. — [doi:10.16997/jdd.1839](https://doi.org/10.16997/jdd.1839) find
  9. James Fishkin, Valentin Bolotnyy, Joshua Lerner, Alice Siu, and Norman Bradburn. 2024. Can deliberation have lasting effects? American Political Science Review, 118(4):2000–2020. doi: 10.1017/S0003055423001363. — [doi:10.1017/s0003055423001363](https://doi.org/10.1017/s0003055423001363) find
  10. James S. Fishkin. 2018. Democracy When the People Are Thinking: Revitalizing Our Politics Through Public Deliberation. Oxford University Press. find
  11. Baijiong Lin, Weisen Jiang, Yuancheng Xu, Hao Chen, and Ying-Cong Chen. 2025. PARM: Multi-objective test-time alignment via preference-aware autoregressive reward model. In Proceedings of the 42nd International Conference on Machine Learning (ICML 2025). find
  12. Christopher T. Small, Matthew X. Zhou, Gal Yona, David Lazer, and Ashish Goel. 2023. Opportunities and risks of LLMs for scalable deliberation with Polis. arXiv preprint arXiv:2306.11932. doi: 10.48550/arXiv.2306.11932. — [doi:10.48550/arxiv.2306.11932](https://doi.org/10.48550/arxiv.2306.11932) find
  13. Guizhen Chen, Liying Cheng, Anh Tuan Luu, and Lidong Bing. 2024. Exploring the potential of large language models in computational argumentation. In Proceedings of ACL 2024, pages 2309– 2330. doi: 10.18653/v1/2024.acl-long.126. find
  14. Vaarunay Kaushal and Taranveer Singh. 2025. DeliberationBench: When do more voices hurt? A controlled study of multi-LLM deliberation protocols. arXiv preprint arXiv:2601.08835. doi: 10.48550/arXiv.2601.08835. — [doi:10.48550/arxiv.2601.08835](https://doi.org/10.48550/arxiv.2601.08835) find
  15. L. D. M. S. Sai Teja, N. Siva Gopala Krishna, Ufaq Khan, Elizaveta Goncharova, and Vasudeva Varma. 2025. DAMASHA: Detecting AI in mixed adversarial texts via segmentation with human-interpretable attribution. arXiv preprint arXiv:2512.04838. doi: 10.48550/arXiv.2512.04838.
  16. Yahan Li, Jifan Yao, John Bosco S. Bunyi, Adam C. Frank, Angel Hwang, and Ruishan Liu. 2025. CounselBench: A large-scale expert evaluation and adversarial benchmarking of large language models in mental health question answering. arXiv preprint arXiv:2506.08584. doi: 10.48550/arXiv.2506.08584. — [doi:10.48550/arxiv.2506.08584](https://doi.org/10.48550/arxiv.2506.08584) find
  17. David Dalrymple, Joar Skalse, Yoshua Bengio, Stuart Russell, et al. 2024. Towards guaranteed safe AI: A framework for ensuring robust and reliable AI systems. arXiv preprint arXiv:2405.06624. doi: 10.48550/arXiv.2405.06624. find
  18. Shuai Ma, Qiaoyi Chen, Xinru Wang, Chengbo Zheng, Zhenhui Peng, Ming Yin, and Xiaojuan Ma. 2025. Towards human-AI deliberation: Design and evaluation of LLM-empowered deliberative AI. In CHI Conference on Human Factors in Computing Systems (CHI ’25), 23 pages. doi: 10.1145/3706598.3713423. — [doi:10.1145/3706598.3713423](https://doi.org/10.1145/3706598.3713423) find
  19. Robert Tibshirani, Guenther Walther, and Trevor Hastie. 2001. Estimating the number of clusters in a data set via the gap statistic. Journal of the Royal Statistical Society: Series B, 63(2):411– 423. find
  20. Christoph Riedl and David De Cremer. 2025. AI for collective intelligence. Collective Intelligence, 4(2). doi: 10.1177/26339137251328909. — [doi:10.1177/26339137251328909](https://doi.org/10.1177/26339137251328909) find
  21. Maged M. Khader, Ali Al-Sharafi, Moath H. Al-Sioufy, Wajdi Zaghouani, and Ali Al-Zawqari. 2024. Munazarat 1.0: A corpus of Arabic competitive debates. In Proceedings of the 6th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT) @ LREC-COLING 2024, pages 20–30.
  22. Anita Williams Woolley, Christopher F. Chabris, Alex Pentland, Nada Hashmi, and Thomas W. Malone. 2010. Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004):686–688. doi: 10.1126/science.1193147. — [doi:10.1126/science.1193147](https://doi.org/10.1126/science.1193147) find
  23. Siwar Laabar and Wajdi Zaghouani. 2024. Multidimensional insights: Annotated dataset of stance, sentiment, and emotion in Facebook comments on Tunisia’s July 25 measures. In Proceedings of the Second Workshop on NLP for Political Sciences @ LREC-COLING 2024, pages 22–32. find
  24. Heba Al Heraki and Wajdi Zaghouani. 2025. Analyzing digital polarization on hijab: A dataset of annotated YouTube comments. In Proceedings of the Nineteenth International AAAI Conference on Web and Social Media (ICWSM 2025), pages 2350–2360. AAAI Press. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zaghouani-2026-accountable-human-ai-deliberation-llms/