From battleground to common ground

BridgingBot shows how AI designed to mediate could nudge toxic online spaces towards healthier discourse

What if, just as an online argument tips into hostility, a trusted mediator could step in—paraphrasing the shouting, surfacing shared concerns, and steering the conversation back toward mutual understanding?

That’s the vision behind BridgingBot, a prototype developed by Jeff Fossett of the Plurality Institute and Harvard University. Powered by large language models (LLMs), it’s designed not to censor, but to model what constructive conflict can look like online.

Fossett presented BridgingBot at the LLMs in Public Discourse conference in February, co-hosted by the Plurality Institute and the Council on Tech and Social Cohesion in partnership with Google.org. The event brought together researchers, developers, platform leaders, and civil society organizers to explore how LLMs might promote pluralism and bridge divides in digital public life. All the conference presentations can be viewed here.

From moderation to mediation

The internet doesn’t polarize people just because they disagree. It polarizes people because it rewards toxic forms of disagreement, by design.

Tackling toxicity has often meant what Fossett describes as “content moderation type strategies,” like removing, demoting, or labeling problematic posts. But those approaches tend to be reactive. “LLMs present a lot of new opportunities for engaging more proactively—to mediate conflict and help make disagreements more constructive online,” he said.

Rather than trying to shut down the disagreement, BridgingBot seeks to make it more constructive by 1) listening and paraphrasing to ‘translate’ through the toxicity and 2) surfacing potential areas of common ground. In short: it acts like a good mediator.

Doing what users want

Users have frequently signaled that they want more healthy, less toxic civic discourse online. As Beth Goldberg, head of R&D at Google Jigsaw noted during the LLMs in Public Discourse workshop, a survey of 2,000 Americans found that people “actually want to have common ground and dialogue with one another,” but instead often encounter hostile, binary conversations that push them into silence or withdrawal.

This can lead to ‘false polarization,’ according to Duke University Sociologist Chris Bail, author of Breaking the Social Media Prism. The prism’s greatest danger, he says, “is not simply that it fuels extremism, but also that it mutes moderates.” This results in users overestimating the ideological extremity of those who don’t share their views.

This desire for higher-quality online discourse is reflected in Google’s Jigsaw team’s Sensemaker toolkit—an evolving set of LLM-powered tools designed to expand participation and understanding. It helps communities make sense of large-scale conversations by summarizing, surfacing points of agreement and disagreement, and offering actionable insights grounded in the original discussion.

The truck in the bike lane

Fossett shared a vivid example of BridgingBot in action from Reddit’s r/Boston. A user posts a photo of a delivery truck parked in a bike lane. The first commenter lashes out: “Grow up! This guy’s doing his job—he has 50 stops to make. Anybody who thinks this is a problem is an idiot.” A second user replies in kind: “Anyone who agrees with you is a [bleep]. This guy needs to obey the law.”

What’s actually being debated here? Is it about laws? Jobs? Safety? Civic priorities?

Enter BridgingBot. It reads the thread and posts a short intervention:

“It looks like you’re debating whether it’s okay to block a bike lane briefly while doing your job. One of you emphasizes the driver's work constraints; the other highlights safety and legality. Could improved city infrastructure reduce these kinds of trade-offs?”

Fossett explained that this kind of intervention draws on what he called “translation and de-escalation,” describing it as “translating not just between languages, but between ways of thinking or ways of speaking.” The goal isn’t to resolve every disagreement, but to “help users find common ground and build mutual understanding,” he said.

Built with, not for, moderators

Fossett underscores how closely the developer team is working with the Reddit moderators in the testing to ensure interventions are tailored to specific community norms and dynamics. This includes:

1. Offline testing with real Reddit threads and moderator feedback.

2. Randomized control trials, including lab-based experiments via the open-source Deliberate Lab platform.

These trials will test not just whether BridgingBot reduces toxicity, but whether it changes how people feel in the conversation.

Do they feel heard? Do they stay longer? Are they more curious? Do they report lower defensiveness? These are the metrics of civic health—not just safety, but engagement.

In a time when trust in platforms is fraying and online polarization is accelerating, BridgingBot points to how AI might be able to do more than silence harm.

Each time BridgingBot steps into a thread, it models the skills of a human mediator—paraphrasing, reframing, and surfacing common ground around underlying needs. These are the same techniques that transform conflict offline—and now, with the help of LLMs, we have the chance to bring them online by designing systems not just to moderate, but to mediate.

Lena Slachmuijlder Co-Chairs the Council on Tech and Social Cohesion.