Exploring How AI Might Enhance Our Digital Public Squares

In April 2024, Jigsaw convened leading experts from civil society, academia, and the public and private sectors to consider how emerging AI technologies might help us improve the quality of online discourse. Today, together with our coauthors representing nearly 20 organizations, we are excited to share some of the most compelling ideas that emerged from that summit, in the form of a position paper on AI and the future of our digital public squares. The paper identifies a number of cutting-edge applications of large language models (LLMs) for online discussions, reflects on key risks to mitigate, and proposes directions for future research and investment.

The summit convened a unique cross-section of partners from this vibrant emerging field of research and development: technologists and practitioners, as well as representatives from media, government, and academia, all trying to combine technology with online conversation and deliberation. Innovations in this field offer real promise for strengthening societies, from AI-enabled platforms for crowdsourced wisdom to LLMs that can help find consensus and aid peace negotiations between seemingly irreconcilable groups. We’re publishing this paper in an effort to help galvanize this dynamic field globally around a shared agenda and set of opportunities.

From its earliest days, Jigsaw has been passionate about using technology to surface more voices and deliver on the internet’s promise. The potential for LLMs to substantively improve the resilience of our digital public squares makes this work a natural next step — one that we’re eager to embark on together with a diverse set of partners.

AI & better futures for online discourse: a collaborative agenda

The following reflections are intended to help develop a collaborative research agenda for this rapidly evolving ecosystem. Together with our coauthors, we hope to see these ideas pursued, challenged, and built upon as we collectively explore the potential for LLMs to improve digital public squares.

In what follows, we show that LLMs have specific qualities that can be harnessed to enhance digital public squares. They are, for instance, especially good at helping people make sense of complex and nuanced dialogues involving a large number of people. They may also be able to help more people feel heard and meaningfully involved in online conversations.

Among the many possibilities for LLMs to play a constructive role in our digital discourse, we have identified four we believe to be the most promising for strengthening online, pluralistic conversations. Since this paper is intended to encourage further exploration, each opportunity is anchored around a core challenge — a spark meant to inspire further research, collaboration, and investment.

Collective dialogue systems (CDS)

How might LLMs enhance collective dialogues to help people feel that they are being heard and that their participation matters?

Over the past few years, researchers and practitioners have been exploring new applications of what is often called deliberative technology: tools and features meant to enable higher-quality, more constructive, and more effective public discussions, both online and offline. One example of deliberative technology is the collective dialogue system, or CDS. These platforms, such as Polis or Remesh, join the qualitative richness of classic focus groups with the quantitative scale of polling. They allow large numbers of people to participate meaningfully in online discussions by, for instance, sharing their opinions or voting on the statements of others. Some CDSes may also help identify emerging areas of agreement, or help participants learn about the existence of different perspectives within their communities. CDSes can help gather collective feedback on a scale once unimaginable, but they are still confronted by major challenges: education and onboarding are time-consuming, CDSes can be costly to host and run, and it can be difficult for policymakers and participants to quickly find meaningful patterns from the data.

LLMs are particularly well-suited to addressing many of these challenges. One example might be the use of proprietary LLMs attached to specific CDSes that are trained with relevant data for specific deliberations. Facilitators and participants could then query these LLMs, helping to make education more personalized and less expensive. Other opportunities for LLMs might include live translation for multilingual populations, intelligent prompting features to help encourage participation, and better forms of summarization and visualization for complex data. Yet CDSes are not a panacea — opinion aggregation is only part of any meaningful deliberation, and LLMs should be used to enhance or expand this wider process of collective reasoning and constructive disagreement, not replace it. Transparency and reliability will also be essential for LLMs to be perceived as trustworthy and helpful additions to collective dialogues.

Bridging

How might LLMs help reimagine digital public squares to enhance bridging while retaining constructive disagreement?

Building trust in fragmented societies is part science and part art, so mediators and designers of discursive spaces have long employed the concept of bridging. This means finding common ground between people with different views, values, and identities. Recent breakthroughs in AI have made it possible to scale bridging to larger online spaces, providing new opportunities to increase mutual understanding and/or reduce polarization.

Bridging algorithms are already commonly in use, powering features like X’s Community Notes or YouTube’s Notes. These algorithms identify and elevate the crowdsourced context notes that resonate most with broad audiences. Now, LLMs make it possible to identify the rhetorical attributes that contribute to trust and understanding in what people post online (e.g. personal stories, compassion, nuance, or reasoning). This, in turn, creates new opportunities for app and platform designers to create ways of surfacing these comments in ways that drive healthier engagement and conversation. Studies have shown that re-ranking for constructiveness or curiosity may increase understanding and reduce animosity. Yet bridging systems pose the risk of amplifying content that is widely appealing but ultimately insubstantial, even at the expense of meaningful discussion.

Community-driven moderation

How might LLMs help moderators save time, avoid burnout, and achieve better results for their communities in ways that are seen as legitimate and trusted?

As many different digital public squares flourish across a wide range of platforms, there is also an opportunity for LLMs to better support smaller online communities and help them have the kinds of conversations they want to be having amongst themselves. Paid and volunteer human moderators play an essential role in leading these communities, but the work can be taxing and often involves working within one-size-fits-all content policies or tooling that may fall short of communities’ unique needs. With LLMs, however, group moderators could benefit from tools that allow them to set and enforce dynamic and tailored guidelines for the conversations happening in their communities.

For the first time, LLMs allow for the possibility of more flexible or “composable” tools that are more sensitive to nuanced content and can be trained to follow a specific community’s norms when flagging potentially harmful posts. LLMs can also be used in ways that help moderators proactively shape community health, including simulated discussions that can help train new moderators, or real-time author feedback features to help group members communicate more constructively. For these LLM tools to be legitimate and useful, however, they will have to be deployed in ways that are mindful of potential risks and designed with transparency in mind. LLM training sets may, for instance, overindex on certain languages or modes of speech and could exhibit bias against disadvantaged groups. If their use and function is not transparent, LLM moderation tools could lead to lower trust levels within communities. And the processes that establish community norms for LLMs to follow will need to be inclusive and fair if they are to be legitimate bases for AI-enabled moderation.

Proof of humanity

How might new humanity-verification methods keep up with the emerging capabilities of LLMs while protecting privacy and ensuring accessibility?

The early internet fostered a culture of free expression among anonymous and pseudonymous participants from around the world, on a wide range of bulletin boards and online forums. Over time, the rise of bots, spam, and deceptive accounts led to new issues with trust and authenticity. Today, there is a reasonable concern that LLMs could amplify these challenges by making it easier and cheaper for bad actors to manipulate and defraud users. Given the pervasiveness and potential benefits of LLMs for online discourse, it will be important to meet these potential abuses with a technical response that can keep pace with AI’s rapid development.

Proof-of-humanity systems constitute one such response. For years, reCAPTCHAs have helped protect against bots by asking humans to solve interactive visual puzzles. But studies have now shown that LLMs are able to solve these problems as well as humans. One potential avenue for solving this technical problem might be personhood credentials using zero-knowledge proofs: a cryptographic method that allows one party to prove to another that a statement is true, but without sharing any further information. This could allow for verification of human status while ensuring user anonymity. Advanced proof-of-humanity systems also bring new potential risks, including coercive consent for biometric data, security breaches or data leaks, and new vectors for surveillance. As a starting point for the secure and equitable deployment of these systems, a set of shared principles should be developed. Initial principles might include interoperability (making credentials portable across platforms), security (ensuring the integrity of systems and personal data), proportionality (using these systems judiciously and minimizing their use when not essential), and self-sovereignty (ensuring user consent and ownership).

Together with our partners, we believe that AI has great potential to benefit online discourse in the years ahead — but for our digital public squares to flourish, this work will need to be guided by responsible innovation, close attention to risk, and an open and collaborative spirit. The above reflections are offered in this spirit, to spark further conversation and exploration.

For those interested in delving deeper into these challenges and opportunities, the full position paper is available as an open-access preprint here.