There is a fundamental challenge at the core of modern governance. Policymakers and civic leaders want to make policy decisions that address the needs of their residents, but realize it’s impossible to hear from every individual they represent.

At the same time, many constituents feel frustrated that their voices aren’t being heard. They aren’t able to find the right local meetings to attend, feel underqualified to take part, or simply do not have the time. Even when people do participate, the only opinions that seem to matter are the ones shouted the loudest. As a result, fewer people attend town halls, leaders host fewer of them, and the gap between what people truly want and the policies that are ultimately implemented widens.

When we launched Sensemaking AI, our goal was to develop tools that would bridge this two-way communication gap between people and their policymakers. Over the past year, we’ve seen how our AI tool suite, powered by Gemini, can facilitate nuanced civic conversations at scale in ways that traditional approaches like town halls and standard surveys cannot. From our early pilots in Kentucky and Oklahoma to an interactive deliberative forum at Bloomberg CityLab in Madrid, Sensemaking AI is bringing vastly more voices into local governance processes, distilling nuanced perspectives and surfacing unexpected common ground across the political spectrum.

Today, we’re excited to announce the Jigsaw Partner Program: a collaboration between philanthropies and technology platforms, to help more policymakers benefit from AI-powered civic conversations.

This partnership will allow us to bring Sensemaking AI to communities of all sizes, even in a moment when many are facing significant budget constraints. Over the next year we will expand Sensemaking AI to policymakers and civic leaders in 30 cities, states, and countries around the world.

Cities and states on the forefront of civic innovation

As part of these 30 localities, we are announcing our first wave of confirmed civic conversations made possible by our Jigsaw Partner Program. With this diverse group of municipal hubs in the United States, Canada, and Poland, we are bringing Sensemaking AI to more domestic and international communities to help civic leaders listen to their constituents in new and exciting ways:

• Chattanooga, Tennessee
• Houston, Texas
• Bentonville, Arkansas
• Riverside, California
• Kitchener, Ontario, Canada
• San Rafael, California
• Schenectady, New York
• Gdańsk, Poland

“Building trust starts with listening. As we develop our 2027–2030 Strategic Plan, this partnership will help us hear from even more voices across Kitchener and ensure our plan for the future reflects the real experiences and aspirations of the people who call this city home.” — Mayor Berry Vrbanovic, City of Kitchener, Ontario, Canada

In addition to municipal deployments, the new Jigsaw Partner Program is supporting the National Governors Association (NGA) and The Napolitan Institute to offer Sensemaking AI in five new states, extending a partnership that began with a pilot earlier this year in Oklahoma and a recently completed civic conversation in North Carolina. Governors across the country face complex statewide challenges, from infrastructure investment to economic resilience, and this partnership will help meet the unique requirements needed to scale public input at the state level.

“By making it possible to listen to hundreds of thousands of constituents simultaneously, this platform will enable state leaders to see beyond politics and seek areas of agreement by creating genuine, public-informed policy that reflects the collective will of their communities.” — Brandon Tatum, CEO, NGA

The partner network expanding Sensemaking AI’s reach

Through the collaborative fund, Google.org is joining forces with leading philanthropic institutions — including Bloomberg Philanthropies and the David & Lucile Packard Foundation — who will support efforts to further explore the potential of expanding Sensemaking AI in municipal contexts.

“Many people have spoken up to their city and heard nothing back. Do that enough times, and you stop bothering. We’re supporting Sensemaking AI because a healthy democracy depends on people believing their voices matter.” — Ben Chou, program officer, Democracy, Rights and Governance, Packard Foundation

Beyond funding, our Jigsaw Partner Program enables policymakers to leverage Sensemaking AI through three foundational platform partners: Change.org, Make.org, and RMG Research. Each platform partner offers unique capabilities designed to help tailor Sensemaking AI to fit policymakers’ needs, get civic conversations up and running quickly, and do it all free of charge.

“With Jigsaw’s Sensemaking AI, it’s about bringing people across the spectrum together to figure out what they actually agree on, and putting that in front of the decision-makers who can make change happen.” — Duncan Lockard, GM of Civic Dialogues, Change.org

“Sensemaking AI gives us the multilingual precision and analytical power to facilitate robust dialogues at every level — from national to regional, state, and city levels — while ensuring that every citizen’s perspective is fairly represented.” — Alicia Combaz, Founder & CEO, Make.org

“Traditional polling gives us a snapshot of what people think. But Sensemaking AI reveals why they think it. By letting us hear people in their own words, instead of the language of politics, we’re able to see connections people make between topics and issues that we otherwise may not see.” — Scott Rasmussen, President, RMG Research

Additionally, the core Sensemaking AI codebase remains fully open-sourced on GitHub, allowing developers, academics, and civic technologists anywhere to inspect, fork, and build upon our offerings and to deploy them on their own.

An invitation to public leaders

Giving government leaders the tools to engage their constituents through sustained, public input can help improve policymaking. This is a challenge that AI is well suited for — both for helping communities feel heard, and enabling those elected to better serve them.

If you are a policymaker or civic leader seeking to understand your community’s priorities, we invite you to express interest in using Sensemaking AI through our Jigsaw Partner Program. Together, we are building a future where hearing every voice isn’t just an aspiration, but the baseline for effective leadership.

Authored by Angelo Carino, Head of Product, Jigsaw

Helping civic leaders hear every voice was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

In the modern digital era of data harvesting and automated profiling, this unencrypted metadata can be stitched together into surprisingly revealing profiles. A 2021 Federal Trade Commission (FTC) report revealed that major internet service providers (ISPs) continuously monitor unencrypted domain and browsing metadata, monetizing sensitive user categories such as location, demographics, and political affiliation with targeted advertising. Beyond commercial profiling, exposed domains can enable targeted scams, network censorship, and digital surveillance.

Closing this privacy gap makes the internet safer for everyone. In fact, it’s not just one gap: domain data leaks in two places during every connection — the initial DNS lookup and the unencrypted ClientHello in the Transport Layer Security (TLS) handshake. In 2018, Jigsaw partnered to help Android introduce DNS-over-TLS support in Android 9 (P). As encrypted DNS has gained significant momentum — now protecting over 25% of global web traffic — Jigsaw has again teamed up with Android to address the second critical leak: the TLS ClientHello.

A Blueprint for the Industry

Encrypted ClientHello (ECH) is a new internet standard published earlier this year that is designed to close this privacy gap. It hides the domain name using a secret encryption key that only the destination website can unscramble. Critically, though, not all web servers will offer ECH support. To avoid exposing only certain connections as ECH-protected, apps and browsers should use ECH GREASE — which sends fake, randomized ECH extensions to sites that don’t support ECH — so that every connection request looks the same.

Starting with Android 17, ECH GREASE will be enabled by default — opening the door to an industry-wide rollout of robust domain privacy to billions of mobile users worldwide. Android 17 sets a landmark precedent as the first major mobile OS to enable broad ECH support. To help app developers leverage this new capability, OkHttp, the open-source HTTP client powering millions of Android apps, has integrated ECH support into its core library.

Deploying a core networking upgrade across billions of devices requires confidence that the changes will not cause significant connection failures. To validate Android’s intended approach, Jigsaw empirically measured ECH performance at scale — showing how it can operate safely without breaking connectivity or slowing down connections.

Together, Android, OkHttp, and Jigsaw are proud to contribute to the industry’s shared momentum toward a more private web.

Ecosystem-Wide Support for ECH

Achieving lasting internet privacy requires alignment across operating systems, networking libraries, and applications. Leaders across the ecosystem have voiced strong support for bringing ECH into the mainstream:

“Enabling ECH by default across Android and major network libraries is a crucial step forward for user protection. We’re excited to see Android take this step and encourage developers to drive adoption across the entire ecosystem.” — Ehren Kret, CTO, Signal

“With OkHttp we strive to build network infrastructure that’s secure, efficient, and reliable. We were excited to integrate ECH because it improves user privacy.” — Jesse Wilson, OkHttp project lead

“Internet security must evolve continuously to match modern threats. Jigsaw’s ECH measurements helped validate deployment in Android 17, addressing a long-standing vulnerability and advancing our shared goal of a more private web.” — David Kleidermacher, VP of Engineering, Android Security & Privacy

“ECH support for Android is a huge step towards closing one of the largest remaining structural privacy holes left on the Internet.” — Nick Sullivan, Co-author of the ECH standard and Founder of Cryptography Consulting LLC (formerly Head of Research at Cloudflare)

Validating ECH Readiness at Global Scale

Before enabling ECH in the platform, the Android team needed empirical answers to three critical questions:

1. Does ECH impact performance?
2. Does ECH increase connection failures?
3. Do any ISPs or middleboxes around the world already block ECH traffic?

We previously published research on the performance of the first step in an ECH connection, the retrieval of connection configuration information over DNS. Building on this foundation, Jigsaw conducted global measurements to evaluate real-world outcomes and illustrated that ECH can be deployed without impact to server incompatibility or ISP interference.

Connection Success

To measure server resilience, Jigsaw tested ECH GREASE connections against the top 10,000 web domains globally. Our GREASE Compatibility Report revealed zero change in connection success rates compared to baseline TLS requests, demonstrating that modern servers handle ECH extensions without breaking.

Global Network Readiness

To understand current network behavior, Jigsaw evaluated HTTP requests across 202 unique countries and 740 unique ISPs, including heavily filtered environments. Our Multi-Country ISP Analysis demonstrated a virtually 0% network interference rate globally. Even in heavily restricted nation-wide networks like Russia’s and China’s, ECH GREASE requests maintained parity with baseline connections — indicating that intermediate middleboxes are not blocking the protocol.

While ECH protects user privacy by encrypting domain names from middlebox inspection, we recognize that managed networks — such as schools and enterprises — rely on network controls to enforce safety policies and block malware. To support these environments, Android provides administrative controls allowing network operators to manage ECH behavior via DNS, ensuring local network governance remains intact without compromising baseline privacy on the open web.

Open-Source Measurement Tooling & Data

We encourage developers to visit our repository on GitHub to review the reports, inspect the full global datasets, and utilize the interactive Jupyter notebooks to run your own custom ECH measurements. If you do, post on our GitHub discussion board to let us know how they helped you build the confidence to adopt ECH on your own services!

Paving the Way for a Confidential Internet

A confidential, secure, and open web is fully achievable when privacy and network performance are treated as complementary forces rather than trade-offs. By combining empirical testing with ecosystem collaboration, Android, OkHttp, and Jigsaw are demonstrating that modern privacy protocols can deploy at global scale without sacrificing speed or reliability.

We invite developers, platform operators, and security advocates to partner with us:

• Adopt Modern Networking Libraries & Best Practices: App developers should upgrade to OkHttp 5.5.0 and enable ECH. Developers building or configuring network stacks should also consult our technical recommendations on GitHub for optimizing HTTPS RR queries for the best performance.
• Evaluate Performance with Open-Source Tools: Explore Jigsaw’s open-source measurement tools on GitHub to test your own domain configurations and evaluate how your services perform with ECH enabled.
• Collaborate on Standards: Join us in advocating for ECH to be enabled by default across all major operating systems, networking stacks, and server infrastructure — ask your hosting providers to enable ECH support for your domains too!

When privacy is built into the core foundation of the internet, we protect not just network traffic, but the agency and safety of users everywhere.

By Maddy Hoffman, Senior Product Manager, Jigsaw

Closing a Critical Internet Privacy Gap for Billions of Users: Android 17 Rolls Out ECH Support was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

At Jigsaw, we identify the complex societal challenges that make people feel stuck in their lives — and build technologies that help them shape the future on their own terms. Over the past few years, we’ve focused on helping individuals and communities raise their voices and be heard, by each other and by their governments. Our novel Sensemaking AI captures opinions at scale with unprecedented nuance, helping policymakers to more deeply understand the people they serve and powering the largest town hall in America.

But disempowerment isn’t only a civic challenge. People yearn for more agency over their lives in many ways: not just as citizens, but as parents, as employees, and perhaps especially as patients. After all, most Americans feel overwhelmed by navigating care for themselves, or for loved ones. Bureaucracies may be complex or overlapping, information hard to find, and clinical staff spread too thin for most people to feel meaningfully in charge, even as they make life-altering decisions.

AI’s potential for human health is becoming clearer, from accelerating biomedical research and improving cancer screenings to saving time on documentation. Yet we believe it is AI’s ability to help people understand one another and solve shared problems that could truly expand human agency in healthcare. Starting with deep ethnographic research, we’re exploring how AI could empower patients and caregivers to make crucial care decisions together, reinforcing their capacity to shape what comes next.

Could better family conversations rewire the care economy?

Consider Piper, a retired principal who we met last February at her home in the Bronx. She is the primary unpaid caregiver for her father, who suffers from dementia and lives in a nearby nursing facility. She coordinates and advocates for his care, while also trying to make decisions with her siblings using a group messaging chat. End-of-life planning would reduce her stress and help her protect her father’s preferences, but the topic is taboo. “My father has an intense fear of dying,” she told us, “so he really does not want to talk about it.” Without this conversation, Piper and her siblings can only wait for an emergency and then react, unprepared and without a shared plan.

Some of the most consequential decisions in healthcare are made within families, especially for the 59 million Americans caring for another adult, like an aging parent or a spouse. Can a loved one still live independently? Should they still be driving? What steps should be taken after a new terminal diagnosis? What are their end-of-life wishes? Polls show that most Americans want to be having these conversations and making these choices together — but haven’t done so. Only 5% of U.S. adults have an advance care plan, which means that many major decisions are made during crises, when options may be more limited and preferred choices harder to implement.

Families like Piper’s lose agency when these conversations don’t happen, or happen badly: both caregivers and older adults can be left feeling stuck, unheard, and unhappy. But health systems also pay a price. Tensions or disagreements between older adults and families can be taken out on clinicians, already understaffed and at risk of burning out. Studies also suggest that rates of hospitalization may be higher for older adults without care planning, leading to higher costs.

As America ages and the nation’s caregiving crisis deepens, more families could find themselves in situations like Piper’s. Caregivers are, as one clinician told us, “the most effective technology” in the room, frequently bearing what the health system can’t. Yet the more they’re sandwiched by their responsibilities to aging parents, children, spouses, and careers, the more likely it will be that critical family conversations and decisions are crowded out — and the harder it will get for older adults to shape their own care. “Either you make the choice,” we heard from one expert, “or the choice makes you.”

But if we could break this vicious cycle, by equipping older adults and their families to plan their futures together, we could expand people’s agency over healthcare — and even start to remake America’s care economy.

What we’ve learned from caregivers & their families

Over the past year, we have immersed ourselves in the lives of families and professionals caring for older adults. Supported by our partners at Is It A Bird, our researchers and developers spent 150+ hours learning from more than 60 older adults, family caregivers, care managers, and community leaders in New York and Arizona: joining them on daily errands, chatting over meals or coffees, and visiting their homes and offices.

We wanted to deeply understand why family caregivers feel so stuck and overloaded, what older adults need to preserve their autonomy, what decisions families are trying to make together, and what’s at stake when those planning conversations fail. Now, as we continue our explorations, we are guided by three key learnings from this ethnographic research:

1. Control isn’t the same thing as agency.

From cooking and cleaning to appointments, finances, transportation, and medications, informal caregivers handle a vast range of daily tasks for the older adults they care for. Most we met were extremely organized, using notebooks, spreadsheets, and digital calendars to manage tasks and keep control. Last fall, we met Tina in New Jersey, where she was tidying her father’s apartment. She cares for him as well as her grandmother, and no detail escapes her notice: the layout of the clothes in the closet, the placement of medication bottles, the precise temperature of the fridge. “If I could,” she explained to us, “I would control everything.” A retired professional, Tina manages the quality of her father’s care very effectively.

Yet despite this everyday control, Tina feels burned-out and stuck. She doesn’t trust clinicians or family members to care for her father as well as she can, so she manages it all herself. She thinks of herself as “the head of the operation, the brain around here,” but also feels like a “slave” to her father’s needs. Crucially, the intensity of her control has led her to downplay her own wishes and needs. The woman who used to travel and go bungee jumping is gone. “I got no time to think of me,” she said. “I have not been able to get to the dentist for two years, and my tooth is hurting.” In this way, Tina was like many caregivers we met: already equipped with organizational systems that helped them stay on top of their daily tasks, but lacking the ability to think beyond “survival mode” and make proactive choices about the future. In control, but without meaningful agency.

2. Planning is more than a piece of paper.

Checklists or documents, like advance directives and wills, help families and older adults plan for the future. But when we spoke with care managers (specialized social workers who help patients and families navigate eldercare), they underscored that the strongest family plans rest on mutual understanding, agreement, and communication. In the words of Bethany, a care manager in New York with nearly 20 years of experience, “differing opinions between caregivers… wreaks havoc.” Sadie, another care manager, agreed. “Open and supportive family communication significantly improves care outcomes,” she explained, adding that care crises can often be prevented “when families are aligned… if your family is all on the same page, it’s easier to move forward.”

We observed this among caregivers and older adults, too. The families who seemed to have the most room to maneuver, and felt the most agency, were those who had frequent conversations together about the future. Martha, together with her sister, takes care of her elderly mother. “We knew something was going to happen and [so we] prepared,” Martha told us, “we’d talk about it.” When their mother had a fall, the family knew right away who would handle which tasks, who to contact, and which options were available. “There’s a big open space of communication,” Martha said. “That makes you prepared… I’m sure it helped so neither of us was surprised.” Talking about the future and making key decisions in advance, older adults and their families are more resilient to crises and more capable of determining what comes next for themselves.

3. Protect the beautiful frictions.

Caregiving for an older adult, whether a parent or spouse, is taxing work that demands attention to logistics, chores, health systems, relationships, and emotions. Many caregivers are exhausted. But the paradox is that much of this work is also deeply meaningful, a way for caregivers to show up for someone who matters to them, or work through their feelings. For Nicki and Diane, sisters in Phoenix now caring for their mother, cleaning out their father’s truck upon his death was a way for them to feel close to him, even as it was painful. Tina, on the other hand, who we met above, learned that her mother had prepared such a detailed list of funeral arrangements that upon her death, Tina simply had to “follow the instructions.” What was perhaps meant to smooth a difficult time left her feeling sidelined and powerless.

New technologies often promise the ability to reduce friction, automate tasks that seem painful and difficult, and sand away the rough edges. But that is not what families and older adults need most from AI. They want more time for what we call the beautiful frictions of caregiving, helping them embrace the challenging, meaningful parts of aging that they don’t know how to approach or struggle to find time for, like talking openly about wishes for future care. We believe that deep listening and collaborative design is essential to make sure that AI is sensitively supporting these beautiful challenges and helping families lean into each other, not accomplish more apart.

Earlier this year, we built on this research by inviting older adults, caregivers, and care managers to join us at Jigsaw and co-design possible AI solutions, grounded in their lived experiences. We are inspired by caregivers like Patrick, who believed that AI tools could really be “phenomenal” in helping his family understand one another and align on his uncle’s care.

We’re also learning from organizations with deep expertise in caregiving. Christine Peck is Vice President for Program at Lifespan of Greater Rochester, a pioneering nonprofit in upstate New York providing services to older adults and caregivers. “As both a professional and a caregiver,” she told us, “I embrace exploring how AI can be helpful in opening pathways of communication.” Peck has seen many families over several decades who “feel unprepared and delay important conversations.” In her view, AI could help older adults and caregivers by offering “a self-paced, safe environment to engage” with one another. “Ironically,” she said, “this technology can help us spend less time searching and more time with each other.”

Please join us as we explore how AI could help older adults, caregivers, and their families prepare for the future — and take their care journeys into their own hands.

By Ian Beacock and Shruti Verma, members of the research staff at Jigsaw

Note on privacy & AI: When our research participants share their lives with us, they consent to having their insights and stories shared with the public, in ways that don’t identify them. In turn, we protect their privacy by using pseudonyms and removing personally identifying information. The quotes in this post are taken directly from our interviews, but the images do not replicate individual respondents. Our design team has used AI creative tools to produce illustrations that reflect a range of participant profiles across our diverse dataset. We think that AI offers exciting opportunities like this to bring ethnographic research to life, while protecting privacy.

Helping families shape the future of caregiving with AI was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

At Jigsaw, everyone’s a researcher pursuing innovations [in engineering, design, research and beyond] that center human agency. Our Research Notes series offers a look inside our process — sharing early findings, hypotheses and in-flight experiments. We invite you to help shape the future of these ideas, turning our collective curiosity into a force for discovery.

Join us on LinkedIn for a discussion.

Public consultations take time. Gathering people together, listening to their thoughts, and analyzing the results — each step alone can take hours to days to weeks, even just to hear from a relatively small group of people. Yet there’s no true substitute for hearing from people in their own words when it comes to gathering rich, nuanced insights that can allow decision-makers to confidently make complex tradeoffs.

Over the past year, we’ve been applying our Sensemaking AI to overcome some of these fundamental limitations, to scale the town hall from a dozen participants to thousands and provide civic leaders with a more complete picture of the needs and priorities of their constituents within hours of a survey closing rather than weeks.

We’ve spent much of the year since our first pilot in Bowling Green honing our process and tools, aiming to provide leaders with better insights faster. So, when the team at Bloomberg CityLab approached us with the idea of running Sensemaking AI live for their audience of nearly 1,000 mayors and civic-minded people from all around the world at their annual CityLab summit in Madrid, we couldn’t pass up the opportunity. To take things one step further, and really put our tech to the test, we challenged this diverse audience of politicians, artists, public servants and creators to tell us what makes their city unlike any other place on earth.

Architecting for Scale and Concurrency

For our initial proofs of concept, we’ve relied on a fairly straightforward setup: leveraging the survey platform Qualtrics paired with back-end API calls to Google’s Generative AI API to create adaptive follow-up questions. This architecture worked perfectly for surveys where responses flowed in over the course of days to weeks.

But running a live session with nearly 1,000 people hitting a system simultaneously introduced an entirely different class of engineering challenges — chief among them being the API rate limits and potential concurrency bottlenecks. To bridge this gap for CityLab, we had to pivot from a passive survey structure to a highly resilient, real-time data pipeline.

We focused our infrastructure strategy on three core pillars:

• Intelligent Traffic Management: We moved our orchestration to a dedicated cloud layer, implementing an automated exponential backoff retry loop. If the AI model experienced a brief spike or rate-limit failure, the system would silently retry the request behind the scenes, smoothing out massive concurrent traffic bursts without impacting the end user.
• Deterministic Fallbacks: In a live environment, “failing gracefully” cannot mean an empty screen or an error message. We built an automated safety net: if an AI-generated question failed to load after three rapid attempts, the system instantly swapped in a contextually appropriate, pre-defined fallback question.
• Real-Time State Syncing: By utilizing a cloud database with real-time listeners, the front-end user interface remained entirely decoupled from back-end processing infrastructure. Whether a participant received a dynamically generated AI response or a pre-defined fallback, the state transition happened instantly and seamlessly on their device.

The result was an illusion of simplicity. The audience was guaranteed a seamless, interactive conversation, entirely unaware of the high-speed traffic control operating quietly in the background.

Sensemaking on Device & in Your Own Private Cloud

Conference Wi-Fi is notoriously spotty and AI resources can face brief periods of heavy demand, creating a risk that even if the survey went off without a hitch we might not be able to complete the analysis in the brief window of time we had to generate and show results back to the audience. To ensure we’d be able to process the data no matter what, we further extended Sensemaking AI to allow it to run with Gemma 4 open models, which can run entirely offline with near-zero latency.

We implemented Gemma 4 support by leveraging the OpenAI API specification, a popular interface which provides compatibility across models from multiple companies, and which is the standard for open weight model inference. This code will soon be made available in our Sensemaking GitHub for the community to use.

In addition to solving the immediate need of a backup solution in the event that our standard inference models succumb to heavy system load, this contribution opens up opportunities for deploying our Sensemaking tools in new contexts. The use of open-source and on-device models allows for strict control over the location where data is processed, a critical enhancement for communities who need full control over their own discussion data.

Generating Novel, Dynamic Visualizations for Dynamic Real-Time Presentations

Our Sensemaking AI delivers a thorough, interactive report of its findings from public consultations. What works well for careful reading at your desk, however, rarely works on a conference stage. Our partners on the CityLab team challenged us to develop a novel experience to pull the audience into the data in real-time.

Creating compelling and informative data visualizations, while maintaining ease of interpretability for a general audience, is among the most challenging UX design problems. We had two factors weighing in our favor however. First, doing the demonstration live meant we’d have the opportunity to walk the audience through the visualization and explain it step-by-step, so novelty didn’t have to be an impediment to understanding. And second, the dramatic advances in the capabilities of AI coding agents has dramatically lowered the cost of experimentation. We were ultimately able to build not just one, but several new visualizations of the data, each of which explored the data from a different angle, and highlighted different aspects of its structure.

We knew we wanted something sweeping and cinematic for the big stage, so for our first attempt, we leveraged the hierarchical nature of the categorization output from Sensemaking AI to render the data as a 3D tree. By adding a subtle amount of noise to buffet the nodes and a post-processing bloom pass courtesy of Unreal, we were able to transform the input into a glowing coral reef.

Inspired by the packed circle charts we’d previously worked on for our first Sensemaking AI proof-of-concept in Bowling Green, Kentucky, along with the recent launch of the Artemis II, our second approach to visualization transformed our categorized data into a glowing galaxy. To achieve this, we rendered the most commonly shared opinions from each topic as central “suns” orbited by “planets” of less commonly shared views. The overall layout leverages the same D3 packSiblings function used in Bowling Green. However, rather than using solid geometry, each sun and planet is set up as a spherical fibonacci lattice, allowing us to expose quotes from each member of the audience.

While both of these met the mark for our goals to create something novel and cinematic, they felt disconnected from an event focused on the future of cities. So we eventually landed on the idea of generating an interactive “city” from the data.

To lay out the city, we put together a custom greedy grid-allocation algorithm to pack topics and opinions onto a square grid, grouping our data into “neighborhoods” with individual buildings composed of “bricks” containing quotes from each member of the audience. We matched the tan and orange brand of the event via custom shaders with a subtle glow, and edge-glints to give the impression of a glass skyscraper under a bright sun.

To ensure we’d be able to seamlessly present the full set of results, we built a series of custom “scenes”, all leveraging a persistent canvas element and driven directly by the CSV output of the Sensemaking AI categorization process. Together with a secondary text file mapping transitions, this allowed us to walk the audience through specific points of interest from the stage. To ensure that the presentation would go smoothly, we then ran multiple trials with both synthetic data and real data gathered through pilot surveys.

While we ultimately stuck with the city scape visualization for our live presentation, we found that taken as a whole, this gallery of visualizations had an impact greater than the sum of their parts. Each visualization captured a unique perspective into the underlying data, which together helped us feel just a little bit closer to the rich multi-dimensional complexity that is public opinion.

Conclusion

Running Sensemaking live presented a number of technical hurdles, but it also illuminated a range of new ideas and possibilities.

To begin with, we’ve begun work to open source the code we used for our live survey in Madrid. More than just another survey platform, we’re exploring the possibility of offering a standalone platform where you’ll be able to launch new surveys with adaptive interviewing on your own infrastructure and kick off analysis and reporting with the click of a button. Through that release, we also hope to show how others might include our Sensemaking technology into other platforms their organizations already run and manage. And we’re excited to see what new applications are afforded by open model support.

Meanwhile, the gallery of visualizations we explored for this activation opened our eyes to new ways of pulling an audience into a conversation. Inspired by this, we’re beginning to explore new ideas unlocked by live conversations, like leveraging the full store of participant responses as an underlying knowledge-base for adaptive interviewing, further transforming the traditional survey into an at-scale conversation.

During CityLab we had the opportunity to learn from a room of a thousand on what made their homes unique. But what struck us in analyzing the results was just how much the audience’s answers had in common. Even as they told us about their unique traditions and foods, landscapes and festivals, the one thing that made their homes truly special to nearly everyone in the room was the people in them.

While it can often seem that public opinion is simply too divided or too complex to ever truly make sense of, what has struck us every time we’ve deployed Sensemaking over the last year is just how much communities — cities, states, and even entire countries or rooms full of strangers from all over the world — have in common.

By Peter Wiegand, Developer Relations Engineer, and Christopher Small, Sr. Software Engineer

Can AI Help Us Better Make Sense of Live Events? was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

A little over a year ago, Jigsaw set out to develop a novel way to deploy AI for one of the most critical, yet often overlooked aspects of our daily lives: civic engagement. This AI isn’t the kind you read about in the news that helps individuals with productivity or learning new skills, but instead helps communities talk to their elected officials, and in turn, enables those officials to better listen to their communities.

When public officials and representatives seek community input on critical issues, they are typically forced to choose between two imperfect options: traditional town halls, which lack representative attendance, or standard polling, which reduces complex human perspectives into narrow multiple-choice boxes.

We created Sensemaking AI to offer a new process designed to bridge this gap. And over the past year, we have deployed Sensemaking at the county-, state-, and national-level to help citizens and policymakers turn individual voices into actionable policy insights.

We’re excited to share more about the impact Sensemaking has had, including how it has helped local leaders in Kentucky and Oklahoma hear from their constituents and affect real policy changes. We’ll also share how we’re building on these early successes with new feature innovations, a first-of-its-kind live demonstration, and the next community where Jigsaw will put these tools into action.

The Next Chapter: Delivering Actionable Change in Chattanooga

To push the boundaries of what this technology can achieve, our next Sensemaking AI deployment will be in Chattanooga, Tennessee, in partnership with the office of Mayor Tim Kelly.

As Chattanooga navigates its ongoing growth, Mayor Kelly has identified transit, transportation, parks and green spaces, and urban development as critical pillars where he actively welcomes public input — and crucially, where he intends to take direct policy action based on that feedback.

This initiative will introduce an innovative multi-phase design to Sensemaking, in partnership with Change.org:

• Phase 1 (Broad Outreach): The wider Chattanooga public will be invited to share their open-ended thoughts on what might be improved in their neighborhoods.
• Phase 2 (Representative Deliberation): A demographically representative sample of residents will dive deeper into the ideas surfaced during the first phase. We are exploring hosting these deliberations over video conferencing to enable deeper discussion of proposals and trade-offs.

While our Sensemaking tools will synthesize the core conversation, we will enrich the analysis with AI-generated insights pulled from external datasets to further enhance the report into a data-rich playbook for municipal action.

“As Chattanooga continues to grow, our success depends entirely on making sure our residents are the ones driving that growth. Traditional civic feedback tools often miss the deep, day-to-day realities of our neighborhoods, but this partnership gives us a real opportunity to listen to Chattanoogans at an entirely new scale. I’m eager to hear directly from our community on the things that impact their daily lives the most — from the reliability of our infrastructure to the quality of our local parks.” → Tim Kelly, Chattanooga Mayor

Momentum & Proof Points: Sensemaking in Action

The upcoming work in Tennessee is built on a strong foundation of validated, real-world deployments over the past year. Here’s a round-up of how Sensemaking has helped policymakers make civic-informed decisions, brought people closer to those who govern them, and given the public a way to express their viewpoints in a nationwide conversation:

• County-level: Warren County, Kentucky — One year ago, Warren County leaders launched What Could BG Be?, a month-long digital town hall designed to shape the 25-year comprehensive growth plan for Bowling Green, the county’s largest city. Using Sensemaking, nearly 8,000 residents participated, representing roughly 1 in 10 locals. Together, they contributed over 4,000 distinct proposals and cast more than one million opinions. Today, that data is actively shaping the physical and economic landscape of Bowling Green. In April, more than 100 regional leaders from the public sector, private sector, NGOs, and other facets of the community released their BG 2050 strategic plan with Warren County’s government. The strategic plan includes a vision, values, and nine priority initiatives that were deeply informed by the What Could BG Be? process. Now, that data is being drawn on by the City-County Planning Commission as they work on the region’s land use and infrastructure plan for the next 25 years. The data has also been drawn on by Warren County Parks & Recreation for their future planning, the City of Bowling Green as they develop the future of the city’s riverfront, local schools that are building learning modules that utilize the dataset, and even local entrepreneurs who are responding to needs voiced by the community for more mixed-use developments.

“The depth of insight we received through this process completely changed how we view public infrastructure. We aren’t guessing what the community wants for the next quarter-century — we have a data-driven mandate directly from our residents.” → Doug Gorman, Warren County Judge-Executive

• State-level: Oklahoma’s American Dream — In Oklahoma, Governor Kevin Stitt used Sensemaking to launch his first ever statewide survey, where 1,000 representative residents mapped out their collective vision for Oklahoma’s future. The experiment highlighted the power of conversational AI in public spaces: 91% of Oklahomans found the AI-generated follow-up questions used in the survey helpful in refining their policy ideas, and 73% felt the process made state leaders feel more accessible to them.

“It was great to read people’s opinions on issues that were directly impacting them on a daily basis — the economy, infrastructure, health and human services, and community. Giving people the opportunity to open up, and then connecting what they’re feeling with what we’re doing is so valuable.” → Kevin Stitt, Oklahoma Governor

• National-level: We the People — As the U.S. approaches its Semiquincentennial, Jigsaw partnered with The Napolitan Institute on a national deliberation called We the People. Engaging over 2,400 Americans across all 435 congressional districts, participants discussed complex ideals surrounding freedom and equality. The conversation comprised more than 1.6 million words. Sensemaking distilled this massive volume of text into 26 core summary statements — the vast majority of these statements reached an agreement rate of over 80% across partisan divides, and 94% of participants reported feeling that their personal opinions were represented in the final output. Earlier this year, The Napolitan Institute alongside Jigsaw had the honor of being invited to share the learnings from We the People at the American Philosophical Society’s annual meeting. The project is being showcased in their exhibit ‘These Truths — The Declarations Of Independence,’ and the project itself will be archived in the Philadelphia-based society’s permanent archives for generations to come. Next, we’re excited to bring We the People to life at more in-person activations in the U.S. this year to showcase how people can explore this gallery of American voices.

Building on our momentum with policymakers, Jigsaw continues to explore new ways to elevate Sensemaking AI’s technical capabilities. In April, we tried something new with Sensemaking — a live facilitation between nearly 1,000 mayors and civic-minded people from around the world at Bloomberg’s annual CityLab summit in Madrid. You can learn more about that feat here.

Scaling Through Partnerships and Open Source

As of today, Jigsaw has also officially open-sourced our complete Sensemaking developer library on GitHub. By making our categorization, predictive agreement, quote ranking, and recursive summarization pipelines public, we are inviting developers, civil society organizations, and academics worldwide to deploy these tools on their own infrastructure.

Since open-sourcing the library, we’ve seen active engagement from the civic tech community. Pro-democracy organizations like The Forum replicated our Bowling Green methodology at the state-level in South Carolina, Nevada, and New Hampshire, elevating tens of thousands of voices in an open dialogue before pursuing deeper deliberations with smaller cohorts on residents’ priorities. Major civic tech platforms are incorporating aspects of Sensemaking technology and design, including Go Vocal, who inspired Jigsaw with their innovations in turn. We’ve also been partnering with Make.org and RMG Research to support their deployment of Sensemaking AI for public opinion research and citizen consultations at the city and state levels.

If you are invested in more responsive government and greater civic voice, we welcome collaboration. Policymakers and organizations interested in partnering can email our team at sensemaking-interest@google.com to share your context and deployment goals. Developers can visit our GitHub repository today to contribute to or apply our open-source tools.

The work is just beginning, and we look forward to what we can build together.

One year on: The momentum behind Jigsaw’s Sensemaking AI was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

As we reflect on Jigsaw’s Sensemaking deployments one year on, we’re excited to deep dive into the Jigsaw Sensemaking AI technology suite, to get under the hood of our tools and encourage replication.

Sensemaking AI is designed to support a range of discussions. These tools have been used to enable generative conversations about the future of a city, public opinion informed decision making at the state level, and deliberations which found hidden consensus across all of America.

The Technology Behind Our Pilots

The Sensemaking suite is built on three core features, each backed by significant engineering and user-experience research:

1. Adaptive Interviewing: Dig Deeper

Our Adaptive Interviewing feature transforms static surveys into a dynamic, two-way dialogue. It is an AI-powered conversational interviewer that guides participants to reflect deeper on their own thoughts, values and experiences.

To draw out rich, high-quality responses without introducing bias, the system operates under strict constraints:

• Brevity: Follow-up questions must be under 50 words to maintain focus and avoid introducing new information.
• Accessibility: Follow-up questions must be accessible at the 5th grade reading level.
• Neutrality: The AI avoids chatbot clichés (such as “thanks for sharing” or “that is an interesting take”) to establish a focused line of inquiry.

Adaptive Interviewing has been highly effective in our proofs of concept. In our Oklahoma pilot, 91% of participants felt that the AI follow-up questions successfully helped them explore their ideas further.

Adaptive Interviewing can be adjusted to meet the needs of a conversation. As examples, we have used Adaptive Interviewing to solicit more elaboration, encourage personal stories, or request suggestions to policy issues.

2. Automated Insights: Extract Themes

Traditional topic modeling only identifies what was discussed. Our models go beyond just topics to identify key opinion clusters — for example, distinguishing between those who favor “public transport expansion” for economic reasons vs. environmental reasons.

Our methodology deploys a recursive summarization pipeline. The AI first summarizes raw quotes into distinct opinions, then synthesizes those opinions into a cohesive topic summary and recursively creates a high-level executive report. Every summarization is strictly grounded in the raw data, enabling a direct drill-down to original quotes.

We validate the accuracy of Automated Insights by listening to our participants: in our We the People pilot, 94% of participants felt that their opinion had been accurately represented in the conversation.

3. Predictive Agreement: Find Consensus

To help decision-makers identify areas of alignment based on the conversation, we developed a Predictive Agreement engine. Powered by AI, the algorithm operates in two parts:

• Statement Generation: The system analyzes the entire deliberative corpus and generates opinion statements that are likely to have broad support across participants.
• Simulated Juries: It then runs a simulated vote, modeling how each participant would likely respond based on the thoughts they’ve previously shared.

In our We the People pilot, the Predictive Agreement engine generated 26 statements predicted to have high agreement. To validate our predictions, we ran a poll of the generated statements with a nationally-representative sample of participants, and found that 22 of these statements had over 80% agreement.

To learn more about the design of Predictive Agreement, see our Research Note.

Inside the Updated GitHub Library

To support the broader civic tech community, we have updated our open source code library on GitHub. We invite developers and researchers to explore our GitHub Library.

The repository’s main branch now holds all code necessary to run our Sensemaking AI suite on a completed public consultation, deliberation, or survey, and includes step-by-step documentation to assist you along the way. This serves as a stable, production-ready environment for developer reuse, enabling any organization to run the full pipeline self-service.

Risk mitigation

As we scale these tools, we maintain strict safeguards to mitigate against known risks.

• Viewpoint bias: LLMs can inadvertently amplify majority views or misrepresent minority dissent. Jigsaw continuously evaluates our summarization functions using diverse, balanced datasets to ensure proportionality.
• Hallucination mitigation: By strictly grounding every summary in direct participant quotes via our reference index, we minimize the risk of the model creating inaccurate narratives.
• De-identified data: To protect privacy, Sensemaking processes statements without collecting or tracking individual user accounts. Users own and control their data via their own Vertex AI accounts.
• Human-in-the-loop: Sensemaking is designed to work with people, not replace them. Project owners, facilitators, and platform partners can maintain complete control, reviewing and refining report outputs before they are shared.

The Future of Civic Tech: Unsolved Problems

While we have made massive strides, civic tech remains a frontier with several exciting, unsolved problems we hope to tackle next:

1. Video Support: Supporting live deliberations done over video, enabling a richer dialog and with real-time AI features.
2. Full AI Interviewing and Standalone Portability: Creating a click-to-run, standalone platform where anyone can deploy an adaptive survey in minutes without technical onboarding.
3. Model Accuracy & Live Translation: Improving cross-lingual alignment so multilingual groups can deliberate in real-time.
4. Future-Casting & Legislative Informing: Exploring how AI can help translate raw citizen feedback into specific, practical, and feasible policy suggestions, while also keeping the public informed on legislation.

We invite developers, researchers, and public officials to partner with us. Whether you have an idea for a new Sensemaking deployment, are planning to use our open source code, or would like to contribute to our research, we want to hear from you.

• Explore our code on GitHub Library.
• Visit our product website: Jigsaw Sensemaking AI.
• Get in touch to join our pipeline of new Sensemaking pilots: Email us at sensemaking-interest@google.com.

Sensemaking AI: Transforming public input into actionable policy insights was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

Research Note: How can developers configure their DNS queries to maximize privacy without introducing latency?

At Jigsaw, everyone’s a researcher pursuing innovations [in engineering, design, research and beyond] that center human agency. Our Research Notes series offers a look inside our process — sharing early findings, hypotheses and in-flight experiments. We invite you to help shape the future of these ideas, turning our collective curiosity into a force for discovery.

Join us on LinkedIn for a discussion.

TL;DR

• The problem: While the internet’s HTTPS protocol encrypts the information you give to websites, network observers can often still see the names of the sites you visit. A breakthrough standard called Encrypted ClientHello (ECH) is designed to hide those website domain names — a huge win for privacy. But it’s not clear whether turning ECH on by default might make connections slower or cause errors.
• What we tested: We measured how fast a normal website lookup is compared to a lookup that fetches the connection instructions (HTTPS RR) needed by the ECH privacy standard.
• What we found: For over 93% of popular websites, the HTTPS RR lookups were nearly as fast as DNS address lookups. In many cases, the metadata in the HTTPS RR lookups could even accelerate connection speed. Only 3.1% of websites were noticeably slower. However, a small sample of important websites failed to respond completely, breaking connectivity.
• What we recommend: Developers should leverage the HTTPS RR by default to prioritize performance and user privacy. As a stopgap measure until broken domains are fixed, they should cap the wait time for the HTTPS RR to mitigate the risk of timing out connections.
• Next steps: Jigsaw plans to publish additional measurement findings in the coming months, which we expect will confirm that turning ECH on by default does not break connections. We also intend to open-source our tools so others can run their own tests.

Closing a stubborn online privacy gap with Encrypted Client Hello

For years, internet users have assumed that the HTTPS lock icon in their web browser means their connection is secure and their browsing is private. In reality, a significant privacy gap still remains. While the content of a web page is encrypted, the destination — the domain name of the site itself — is still visible to Internet Service Providers (ISPs), WiFi operators, and other network observers, including malicious actors. We discussed the risks of exposure to unencrypted domain name data in a blog post last year and recognize that the dangers of this leaked data continue to grow.

Encrypted ClientHello (ECH) is a new internet standard just published last month and is designed to close this gap. However, ECH is not a simple toggle that can be flipped overnight. The internet is a heterogeneous ecosystem, where servers, browsers, and apps are configured differently, and each one will have to independently adopt ECH in order to offer this privacy benefit to users. On top of that, there isn’t an efficient method to test the real world implications of ECH across the internet. Network measurement requires specialized technical expertise, and the ECH standard is so new that most libraries don’t support it yet.

For developers and operating systems, this creates a complex puzzle: How do you configure ECH to maximize privacy without introducing latency or causing connection failures? Without empirical data on how ECH performs in the wild, developers may choose to adopt overly conservative policies that ultimately compromise the protocol’s effectiveness. To solve this, we at Jigsaw are applying our network measurement expertise to conduct a suite of tests to identify the risks and recommendations at each decision point of the connection process. Today, we are sharing the methodology and results of our first test.

A deep dive into the HTTPS Resource Record

Our research began at the very start of a web connection: the DNS lookup.

When a device attempts to connect to a website, it typically requests the IP address of the website from the Domain Name System (DNS). However, to encrypt the domain name, knowing the IP address isn’t enough; the device must also retrieve the encryption key. This key is stored in a modern DNS entry known as the HTTPS Resource Record (HTTPS RR). This record acts as a one-stop shop for connection instructions, allowing devices to discover encryption keys, supported protocols, and performance optimizations, which can collectively boost both performance and privacy.

This introduces a race. The device must decide whether to wait for the HTTPS RR (ensuring the domain name is private) or proceed immediately once it receives the records containing the IP addresses (ensuring the connection is fast). If the HTTPS RR takes significantly longer to return than the address records, that extra wait time introduces lag. But if the device doesn’t wait, the user loses the privacy benefit of ECH.

To quantify this trade-off, we measured DNS query latency against 10,000 of the most-visited domains on the internet. We specifically analyzed the time difference between receiving address records and receiving the HTTPS RR for each site. We also checked the content of these records to see if they contained additional information that could help speed up the connection process in later steps.

Since library support didn’t yet exist in the Go language for parsing HTTPS Resource Records, we ended up adding support in order to run our tests. We also used a version of curl (a popular command-line tool for web requests) created by the DEfO project.

A win-win for both privacy and performance

Our analysis yielded important insights that should reassure developers looking to implement ECH.

First, the vast majority of HTTPS RR queries are highly performant. For over 93% of queries, the HTTPS RR returned almost simultaneously (e.g. within 50 milliseconds) as address records. Even when the HTTPS RR arrived later, in most cases it would not slow down connections because it is not required until a later step in the process.

Second, in many configurations, using the HTTPS RR can even accelerate performance. We found that 60% of HTTPS RRs returned before the TCP connection would start. If those HTTPS RRs contained IP addresses, the device could begin the connection step immediately rather than waiting for the address responses. In addition, HTTPS RRs that advertise support for the QUIC protocol allow the device to bypass the TCP Handshake step and instead proceed to the TLS Handshake immediately. This saves a round trip between the client and the server.

Third, we also found that a small minority of domains exhibit increased latency. While most sites perform well, 3.1% of queries showed a noticeable delay (more than 100 milliseconds) between the return of address records and the HTTPS RR.

Finally, of all 10,000 domains we tested, a few dozen never returned a result. This includes popular sites, many of which are in the .gov domain. Waiting indefinitely for those HTTPS RRs would break connectivity to those sites, but the device or operating system can set a cap on the wait time to preserve connectivity.

Paving the way for a confidential internet

Developers can capitalize on these findings today to prioritize the privacy benefits of ECH while also reaping the performance benefits of HTTPS RRs. Because HTTPS RRs are performant for over 93% of queries, the delay caused by waiting for an ECH key is negligible for most users. Prioritizing the retrieval of the ECH key is the only way to ensure the user protection that the standard promises.

Alongside our measurement data, we have posted a detailed set of recommendations that developers and service managers can take now to enable HTTPS RRs by default while mitigating connectivity risks. The primary mitigation we recommend is to cap the wait time for HTTPS RRs as a temporary measure against results that never return.

At Jigsaw, we’ll continue leveraging our network measurement expertise to understand the impact of ECH on performance and latency. We have begun subsequent tests on the TLS handshake, and initial results show that ECH does not break connections. We plan to publish our full findings to confirm this assertion over the coming months.

In the meantime, we encourage developers to leverage our open source measurement tools to run their own experiments and validate their own configurations. Through this rigorous testing, we aim to pave the way for a confidential, secure, and resilient web for everyone.

By Vinicius Fortuna, Engineering Manager, Jigsaw

Research Note: How can developers configure their DNS queries to maximize privacy without… was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

Since its launch in 2018, Outline — Jigsaw’s open-source anti-censorship technology — has pursued a simple but powerful idea: empower anyone, anywhere, to access the global internet and share information freely. Defying the traditional barriers of geography or circumstance, Outline transformed the global digital landscape, expanding the internet freedom ecosystem from a handful of centralized services to a vast, distributed movement of resilient virtual private networks (VPNs) and apps operated by individuals, journalists, and human rights defenders. Today, we see Outline as an essential public good — one that deserves long-term, community-centered governance. Therefore, as of January 2026, Jigsaw will transition stewardship of Outline to a new, independent, nonprofit organization: the Outline Foundation.

A legacy of global impact

Prior to Outline’s launch, censorship-resistant VPN service was consolidated among only a handful of providers. Users had to trust these third party providers in exchange for access, and the centralization of service made each one easier for censors to identify and block. Outline upended this paradigm through its distributed model that empowers anyone to deploy their own VPN in the cloud, and currently enables over 140,000 self-managed VPNs. Outline has become a critical access point for people facing internet restrictions, particularly during crisis events — whether during the VPN blocks in Turkmenistan, the civil war in Myanmar, Russia’s invasion of Ukraine, or the Women, Life, Freedom movement in Iran. In these moments, demand for Outline spikes almost instantly, often serving as an early indicator of unrest or information disruption.

“Outline is attempting to fulfill the original design criteria of the Internet by keeping the system as open as possible in the face of resistance to that goal.”

– Vint Cerf, Google Chief Internet Evangelist

As Outline’s impact grew, we saw a path to making the core Outline technology into even more than a VPN tool. Developers began forking our open-source code, localizing and extending it with new features to meet their needs. Recognizing this momentum, we launched Outline SDK in 2023, enabling any developer to integrate Outline’s powerful network resilience technology directly into their apps and services. Outline SDK grew to become a common platform leveraged not only by all four leading anti-censorship VPNs but also by prominent news publishers, which use it to reliably reach audiences in highly restricted environments. Today, Outline technology helps over 30 million people connect to the open internet, offering a vital lifeline in places where information is most at risk.

Oliver Linow, Internet Freedom Specialist at Deutsche Welle, notes that, “Deutsche Welle takes the issue of internet censorship very seriously, as our online content is blocked in many countries. Our new DW Access app is developed specifically for people in countries with internet restrictions and uses Outline SDK to provide reliable connectivity. We trust the Outline technology because it is open source, auditable, adaptable, and community-informed.”

Valentina Aguana Villegas, Technical Project Coordinator at Conexión Segura y Libre, adds, “Outline SDK powers our newsreader app Noticias Sin Filtro, which we launched just before the 2024 Venezuelan presidential elections. Thanks to Outline SDK, Venezuelans can access the news and circumvent a new wave of digital censorship and internet blocks, without requiring any technical expertise. We’ve had over 130,000 downloads since our launch, and over 30 media partners whose websites or audio content can now be accessed through the app.”

A vision for a more collaborative future

As digital repression spreads, independent tools that provide safe and reliable access to the open internet are paramount. Outline SDK creates a space where circumvention experts, VPN service operators, and app developers can build on each other’s work to accelerate our collective impact against censorship. We envision a world where circumvention tools are no longer isolated defenses, but rather an interoperable and resilient ecosystem. In this collaborative model, Outline will serve as a force multiplier for the community’s collective strength, and internet censorship can finally become technically obsolete.

Achieving this vision requires operating from a neutral and community-governed place within the broader internet freedom community. The Outline Foundation will be the steward of the Outline portfolio, cultivating its open-source technology and future development and ensuring that Outline remains a trusted resource shaped by the people who rely on it.

The Outline Foundation will be led by Shabnam Aslam, as President. With more than fifteen years of experience in human rights, humanitarian action, and civic technology, Shabnam’s career has been shaped by frontline work in conflict zones and global efforts to defend digital freedom in some of the world’s most challenging environments.

“Outline has always been more than a VPN — it’s a critical infrastructure trusted by millions. I’m humbled to steward Outline into its next chapter as we build a community-governed home that is independent, transparent, and resilient. My focus is on servant leadership and ensuring Outline remains a trusted technology grounded in the communities who rely on it.”

– Shabnam Aslam, President, Outline Foundation

Jigsaw will continue to be a contributing partner in this new model. In particular, Vinicius Fortuna, Jigsaw’s Engineering Lead for Outline, will serve on the inaugural board and continue to inform the organization’s technology strategy. Along with Vinicius, the founding board members include Adam Fisk of Lantern, and Martin Zhu of nthLink — both of whom lead VPNs serving users around the globe who experience internet restrictions — with plans to add new members in the coming year.

Join us in shaping the future of internet freedom

We are deeply grateful to the many partners, contributors, developers, and community members who have shaped Outline into the global resource it is today.

As the Outline Foundation takes the vision forward, we, together with the foundation’s leadership, encourage and invite:

• Developers to continue contributing to and building on the open-source codebase
• Organizations and funders to partner with the Outline Foundation in strengthening global access
• Community members to help define the future of a truly interoperable censorship-resilience ecosystem

We are incredibly proud of what the Outline community has accomplished, and we are even more excited for it to thrive and grow under the stewardship of the Outline Foundation.

By Scott Carpenter, Managing Director, Jigsaw

Introducing the Outline Foundation: An Independent Home for Outline was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

For decades, organizations across government, business, and civil society have sought a genuine understanding of complex public sentiment. However, traditional tools present a fundamental challenge: polls can lack the necessary nuance, and focus groups cannot achieve the necessary scale. New AI innovations allow us to synthesize and extrapolate deep, nuanced data, helping to resolve the long-standing limitation between achieving depth and maintaining breadth in public consultation.

That’s why we partnered with the Napolitan Institute, a nonprofit that seeks to elevate the voices of everyday Americans, to attempt something new: We the People, a conversation among a representative group of Americans — including participants from all 435 congressional districts — about what it means to be American today. This pilot attempts to merge the nuance of a focus group with the power of a poll, reimagining how organizations can listen to the public.

Today, we are sharing our learnings from our pilot conversation, which was on freedom and equality, as well as an interactive report that helps you explore the conversation for yourself.

The Focus Group: The Power of Being Heard

The Napolitan Institute invited over 2,400 Americans to engage in a new form of online dialogue, where they responded to a set of questions on what freedom and equality mean to them today, and reflected on how their lived experience informed their responses.

This microcosm of America shared roughly 10,000 views, creating a rich tapestry of personal experiences and different perspectives — the complexity you’d expect if a cross-section of Americans were given a new platform to explain their thoughts in their own words and respond to one another.

Not only did this group show a clear desire to share — with participants sharing over 1.6 million words in total and spending an hour in the conversation on average — but they also reported an appetite to be heard. The percentage of participants who said they “feel heard by other Americans” jumped from 40% to 68% after the experience.

The Focus Group: Discovering the Opinion Map

As the Napolitan Institute’s technology partner, we used the latest AI models to enable this conversation:

• Individual Expression: AI served as a partner, asking personalized follow-up questions to prompt participants to think more deeply, ensuring they shared their personal experiences and nuanced views.
• Collective Reflection: The AI translated the 10,000 viewpoints from the individual expression phase into an organized opinion landscape that allowed participants to listen to one another and respond to the breadth of the group’s input.

Our priority metric was the participants’ subjective experience: despite the large group size and wide array of viewpoints shared, did they feel represented in the conversation? A resounding 94% indicated they did, a stark contrast to the 47% of participants who said they feel represented in general opinion polls.

The Poll: Synthesizing the Discourse

Against the backdrop of the nation’s 250th anniversary and building on decades of polling experience, our partner, the Napolitan Institute, hypothesized that areas of shared agreement exist within the varied opinion landscape of the scaled focus group. They put forth a challenge: Could AI distill the breadth from the scaled focus group into a concrete set of short, poll-ready statements likely to have broad support across participants?

From the 1.6 million words of the conversation, we used AI — with human-in-the-loop support from the Napolitan Institute — to generate a set of statements that were informed by the discourse, covering a range of views, and predicted to have broad support across participants. When invited to agree or disagree on the final set of 26 statements: the majority of the statements (85%) had over 80% agreement.

This left a critical open question: Do statements generated directly from the We the People conversation have broad appeal beyond the group? Putting the same set of statements to a different representative group of Americans in a traditional poll, our partners at the Napolitan Institute found similarly high levels of support — with all statements receiving over 75% agreement — demonstrating that the statements produced through the We the People process, surfaced with AI, resonated beyond our pilot conversation.

Moving Forward

Working alongside our partners at the Napolitan Institute, we are encouraged by the fundamental takeaways from this pilot: that people are eager to share when genuinely invited, and that they deeply value being heard. Explore the interactive report and our learnings on the Napolitan Institute’s website. It’s your opportunity to engage with the conversation directly — not just the data, but the stories.

The conversation on freedom and equality is just the first, with another planned for early next year by the Napolitan Institute as we approach America’s 250th anniversary in 2026. We see exciting new opportunities to enable public consultation that combines authentic voices and perspectives with AI’s strength in synthesis and scale across every sector of society.

Nuance at Scale: The ‘We The People’ Pilot was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.

At Jigsaw, everyone’s a researcher pursuing innovations [in engineering, design, research and beyond] that center human agency. Our Research Notes series offers a look inside our process — sharing early findings, hypotheses and in-flight experiments. We invite you to help shape the future of these ideas, turning our collective curiosity into a force for discovery.

Join us on LinkedIn for a discussion.

In our We the People pilot, AI enabled participants to share and respond to a vast array of opinions and personal ideas — in total more than 1.6 million words! This project, conceived in collaboration with the Napolitan Institute, aimed to merge the substance of a focus group with the quantitative scale of a poll. From his decades of polling experience, Scott Rasmussen, founder of the Napolitan Institute, hypothesized that from within that huge breadth of perspectives, we’d find areas of shared agreement. He prompted our team with a research question: Could AI distill the breadth of 1.6 million words into a concrete set of short, poll-ready statements likely to have broad support across participants?

Earlier this year, our colleagues at Google DeepMind published a paper in Science titled AI can help humans find common ground in democratic deliberation (Tessler, Bakker, et al. 2025), which demonstrated an approach to identifying shared agreement in a controlled lab experiment in the UK. From initial open text responses, they used AI to generate several summary statements and predict which would garner the strongest support. AI then further refined these statements based on iterative feedback from participants. The participants ultimately approved these AI generated statements at higher rates than those written by human facilitators, demonstrating the success of their method.

Our challenge? DeepMind’s system was designed and tested with around five to six participants per conversation. We engaged thousands: Over 2,400 people across the 435 US congressional districts participated in our first We the People conversation.

As we detailed in an earlier post, this conversation took place over three rounds, each with a different purpose and unique application of AI to support that purpose:

• Round 1: Individual Expression — Participants responded to open-ended questions about America’s founding ideals of freedom and equality, and elaborated on their views in response to AI-generated follow up questions.
• Round 2: Collective Reflection — Using AI to categorize responses into broad topics and distinct opinions, we created an interactive summary of responses from Round 1. Participants were then invited to explore what others said, respond to selected quotes, and rank opinions.
• Round 3: Discourse Sourced Polling — Based on responses from Rounds 1 and 2, we used AI to generate declarative statements likely to have broad support among participants. These were presented back to participants, who were then able to agree or disagree with each statement.

This Research Note explains how we took inspiration from DeepMind’s work to distill this massive and far-reaching discussion into 26 simple statements representing areas of broad agreement among We the People participants.

Step 1 — Drafting statements

By the end of Round 2, participants had shared over 1.6 million words worth of text, more than would fit in the context window of even the most powerful AI models. To overcome this challenge, we began organizing the data by the opinions identified in Round 1.

For each opinion, we collected all the corresponding Round 1 quotes together with responses to those quotes from Round 2. We then prompted AI to generate a set of statements for each opinion, with the goal of fully and accurately capturing the shared sentiments from the discussion on that opinion. This resulted in a total of 96 statements, expanding upon more nuanced aspects of the original opinions.

Step 2 — Predicting preferences

With 96 statements thus generated — more than we could reasonably expect participants to react to — we sought to predict which would garner the strongest support.

For each participant, we prompted an AI with all of their responses and asked it to make two key predictions about the statements:

• whether that participant would likely agree or disagree with each statement, and
• how that participant would likely rank the full list of statements within each topic.

These two predictions together helped to distinguish nuances. While someone might agree with 10 statements, they could nevertheless feel much more strongly about 2 or 3 of them.

Broadly, this technique of predicting preferences builds on the work of Tessler, Bakker et al. (2024) and Fish et al. (2023), who demonstrated that these so-called simulated juries can extrapolate from plain-text position statements to help make decisions that align with collective participant preferences.

Critically, these predictions are only used to help us generate a shortlist of statements for participants to react to in Round 3. For AI to support and not supplant human agency, it’s essential that participants have the final say in selecting the statements that best represent their shared beliefs.

Step 3 — Ranking statements

Based on predicted individual preferences, our next step was to identify a set of statements representing the strongest points of shared agreement across the participant group. Specifically, this meant optimizing for two goals:

1. Each statement would likely be supported by most participants
2. Every participant would likely support many statements

To achieve this, we combined two existing algorithms to add statements to the set one by one, similar to how one might draft a sports team.

For the first pick, we followed Tessler, Bakker et al. (2024) in using the Schulze (2003) method. This modern and robust ranked-choice voting mechanism fulfills a number of important criteria, notably Condorcet completion, which simply states that if there exists a candidate that beats every other in a head to head match, then that candidate should win. This method identified the single “best” statement likely to garner strong support across participants.

However, as with a sports draft, it’s essential that subsequent picks bring unique skills to your team as a whole. After all, what good is drafting 7 star quarterbacks, without any receivers to throw to? To ensure that additional selections brought something different, we considered how many statements each participant is likely to support, prioritizing adding statements that allow more participants to say “yes” to more things.

To facilitate this, once the top statement was selected by the Schulze method, additional selections were made using Proportional Approval Voting (PAV). This method, specifically designed for multi-candidate elections, balances maximizing how many participants approve of each statement with maximizing how many statements each participant approves. Formally, the method satisfies so-called Justified Representation (JR) and Extended Justified Representation (EJR) axioms, mathematically rigorous criteria which capture these intuitive notions of representation breadth and depth. Informally, to extend our drafting metaphor, this means ensuring that each draft pick adds a new skill to the team as a whole.

This resulted in a ranking of the statements for each topic, based on how likely each statement was to serve as a point of shared agreement.

Step 4 — Cross-topic statements

Up until this point, we had generated statements scoped per opinion, and therefore all statements aligned to one of the main topics surfaced after Round 1. However, we also wanted to allow for ideas from different topics to combine in new and interesting ways, to allow for the possibility of surfacing more nuanced perspectives.

To accomplish this, we prompted an AI with statements selected from Step 3 and asked it to generate a set of additional statements that combine ideas across topics. These cross-topic statements were also ranked according to the process above.

Step 5 — Final selections

To obtain the final list of 26 statements for participants to respond to, the Napolitan Institute went through a human-in-the-loop process to make the final selections. Our goal with this project has always been to combine the scale of AI with the wisdom of human judgement, and so it was vital that the final set of propositions go through careful human review to balance predicted support with a range of positions.

This process involved reviewing the statements topic by topic while balancing the following criteria:

• The rank outcome of the algorithm described above
• Redundancy with already selected comments
• Readability and clarity
• Degree to which the statement encompassed the most highly ranked opinions from Round 2 responses
• A balanced number of statements per topic

Results & Conclusions

In Round 3 of the conversation, we presented the final set of statements back to participants and asked them whether they agreed with each one. Of the 26 statements generated through the process above, 22 had over 80% agreement among participants and 85% of participants agreed with 20 or more of the statements. Moreover, an astonishing 94% of participants felt their opinions were represented in the conversation, and 77% found more common ground than they expected. We invite you to visit our full report to experience the conversation.

We also welcome future research to build upon this work. A few research questions raised by this work include:

1. How might we improve simulated juries to more accurately capture individual participant preferences?
2. How can we design tasks to produce data with greater predictive power for simulated juries?
3. How can we design processes like this to find more substantive common ground, especially on contentious issues?
4. How can we faithfully and productively surface differences of opinion alongside areas of agreement?

This work has significant implications for how we surface and understand public opinion, providing a proof of concept for how AI might facilitate and synthesize conversation at scale. Ultimately, we hope these methods offer a foundation for other practitioners and researchers to build upon, leading to more nuanced public discourse and shared understanding.

By Christopher Small, Sr. Software Engineer; Rachel Xu, Research Manager; Lucy Vasserman, Head of Engineering, Jigsaw

Research Note: Can AI help thousands of people discover shared values? was originally published in Jigsaw on Medium, where people are continuing the conversation by highlighting and responding to this story.