Habermas Machine: Open-Source Code and Dataset

Alongside the peer-reviewed Science paper describing the Habermas Machine, Google DeepMind released the underlying study dataset and a prompted implementation of the system as open source.

What Is the Habermas Machine?

The Habermas Machine is an AI mediation system, named after philosopher Jürgen Habermas, designed to help groups of people with differing views converge on a group statement that reflects genuine common ground rather than simply the majority position. It pairs a generative language model, which drafts candidate group statements, with a personalized reward model that predicts how much each participant will agree with a given statement, refining the statement iteratively based on participant critiques.

What’s in the Repository

Study datasets, available as Parquet files (roughly 450 MB total) covering candidate statement comparisons, final preference rankings, position-statement ratings, and round-by-round survey responses from the underlying study.
A prompted implementation of the Habermas Machine, installable as a Python package, that uses an elaborate prompt and chain-of-thought approach rather than the fine-tuned models used in the original paper.
A demo notebook for trying the prompted version.

Some important differences from the version used in the published study, as documented in the repository itself:

The prompted version is not fine-tuned on the released dataset; it relies on prompting and chain-of-thought reasoning instead.
It currently uses a Gemini backend via AI Studio, though the codebase supports adding other LLM backends.
Its reward model generates a full ranking over candidate statements rather than the scalar scoring approach used in the paper, though additional reward models can be added.

Licensing

The code is released under the Apache License 2.0, and other materials (including the datasets) under Creative Commons Attribution 4.0 (CC-BY). Google DeepMind states this is not an official Google product and provides it “as is,” without warranty.

Why This Matters for Citizen Participation

An openly released dataset and reference implementation means the underlying method is inspectable and independently verifiable, not a black box. That matters for anyone considering AI-assisted deliberation for real citizen participation processes: you can examine what the model was actually trained and evaluated on, and adapt the prompted version to different contexts, languages, or issue types rather than relying on a vendor’s undocumented claims.

We plan, set up, and incorporate tools like the Habermas Machine into citizen participation processes for our clients. See our AI Tools for Society page for how this fits alongside our fact-checking work.

Official Resource

Need Help Implementing AI-Mediated Deliberation?

We Plan, Set Up, and Incorporate the Habermas Machine in Citizen Participation

We help organizations running citizen assemblies, participatory budgeting, or other deliberative processes evaluate whether AI-mediated deliberation fits their context, and if so, plan and set it up responsibly, including the parts the Habermas Machine doesn't cover on its own, such as fact-checking and topic moderation.

Talk to Us About Citizen Participation