A collage of circuit boards interwoven with aerial photographs of water treatment facilities, agricultural lands, industrial sites and mining machinery.

Privacy by Architecture Makes AI Compliance Work

A yellow school bus is seen in Chicago, Illinois, United States, on July 28, 2026. (Photo by Marcin Golba/NurPhoto)

In the US, Caution Rules on AI Ahead of K-12 School Year

A young child holds up a smart phone in front of their face using both hands.

This Week's Meta Settlement Will Reverberate Across the World for Years

This image is a collage with a colourful Japanese vintage landscape showing a mountain, hills, flowers and other plants and a small stream. There are 3 large black data servers placed in the bottom half of the image, with a cloud of black smoke emitting from them, partly obscuring the scenery.

Deborah Lupton, "Servers in a Landscape" / 

'Technology Freedom Cooperative' To Build Infrastructure for Human Rights and Justice Groups

Meta CEO Mark Zuckerberg (C) ahead of testifying before the Senate Judiciary Committee hearing on protecting children from sexual exploitation online, in the Dirksen Senate Office Building in Washington, DC, USA, 31 January 2024. (Photo by Ritzau Scanpix/Sipa USA)(Sipa via AP Images)

Meta Owes the Rest of the World Its Teen Protections

A massive sculpture of Icarus lies in front of the Concordia Temple in the Valley of Temples in Agrigento, Sicily, on January 5, 2019. Shutterstock

A massive sculpture of Icarus lies in front of the Concordia Temple in the Valley of Temples in Agrigento, Sicily, on January 5, 2019. 

Make AI Companies Criminally Liable for Preventable Harm

The EU and US flags in front of a European Commission logo.

The EU and US Can’t Agree on Tech — or Trust

Nicklas Akers, California Senior Assistant Attorney General, left, and Deputy Attorney General Megan O'Neill exit court after tech giant Meta reached a settlement to end a landmark trial over teen social media addiction in Oakland, Calif., Wednesday, Aug. 26, 2026. (AP Photo/Noah Berger)

Assessing the Landmark Meta Settlement

Signs for company Cambridge Analytica in the lobby of the building in which they were based on March 21, 2018 in London, England. (Photo by Chris J Ratcliffe/Getty Images)

The $459 Million Cambridge Analytica Reckoning Buried in Meta's Teen Safety Settlement

The TikTok logo is displayed through a magnifying glass in this photo illustration in Ontario, Canada, on February 5, 2026. (Photo by Thomas Fuller/NurPhoto via AP)

Understanding Brazil’s $29.7 Million TikTok Fine Over Children’s Data

Three human silhouettes are depicted mid-stride, each comprising of a collage of hardware used to develop AI including circuit boards and CPU chips. In the background there is a black and grey gradient with smoke. Beneath the runners from the left, there are two rectangle panels which run across the image, one is red and the other is blue. 

To Avoid a 'Tokenocracy,' Ensure Popular Access to AI Systems

CHONGQING, CHINA - AUGUST 10: In this photo illustration, the logo of Meta is displayed on a smartphone screen on August 10, 2026 in Chongqing, China. (Photo by VCG/VCG via AP )

What Meta’s US Settlement on Child Safety Means for Europe

Meta CEO Mark Zuckerberg attends UFC Freedom 250 on the South Lawn of the White House, Sunday, June 14, 2026, in Washington. (AP Photo/Alex Brandon)

What's in Meta's Teen Social Media Settlement—and What Hinges on Its Rivals

Image of Solana Larsen with text, Managing Editor Solana Larsen

Solana Larsen Named Managing Editor of Tech Policy Press

A child stands near a games stall. (Photo by Maxim Konankov/NurPhoto via AP) 

(Photo by Maxim Konankov/NurPhoto via AP)

Banning AI Toys for Four Years Won't Make Kids Safer

A yard sign expressing opposition to the construction of a data center is seen outside of a home in Mifflinville, Pa. on July 3, 2026. (Photo by Paul Weaver/Sipa USA via AP Images)

Data Center Discontent Drives State Legislation Surge

L to R: Open AI CEO Sam Altman, US President Donald Trump, CEO of DeepMind Demis Hassabis, South Korea's President Lee Jae Myung and German Chancellor Friedrich Merz attend a working session during the G7 Summit in Evian, France, on June 17, 2026. (Jeanne Accorsini/Sipa via AP Images)


AI Systems Are Getting More Powerful. The Ability to Verify Must Keep Pace.

The Capitolio Nacional in Bogotá, where Colombia's Congress is considering Bill 025 to regulate artificial intelligence.

Colombia is Preparing a Poor Copy of the EU's AI Act

French President Emmanuel Macron. (Manon Cruz/Pool Photo via AP)

France’s Constitutional Council Ruling Tests the Limits of Social Media Bans

United States Treasury Secretary Scott Bessent speaks to members of the media outside the White House in Washington, Thursday, Aug. 20, 2026. (AP Photo/Jacquelyn Martin)

What the Bond Market Disruption Means for Rebuilding American Tech Policy

Chris Mills Rodrigo is a fellow at Tech Policy Press

As students and teachers return to elementary, middle, and high schools in the United States over the next month, they are unlikely to find consistent, clear rules about what kind of artificial intelligence-powered programs they can use to help complete or deliver their coursework.

The rapid rise of AI for generating text and images—paired with the natural desire of most kids to get 

 done as fast as possible—has left teachers and school administrators scrambling to develop policies for how the technology should be integrated into education.

Ahead of the 2026-27 academic year, many school districts have adopted a cautionary approach to the technology, establishing guardrails and promoting AI education but shying away from specific, system-wide rules about which programs can be used, delegating in-classroom decisions to individual teachers and leaving at-home use largely untouched.

Others have failed to approve any guidelines at all.

The aversion to setting explicit rules on AI products is leaving teachers in the lurch. Sixty percent of educators said that their districts’ policies were unclear in 

 2025 survey. Just 18 percent of teachers reported getting any kind of formal guidance from school administrators on how AI tools should be used in a 

Opaque direction appears to be trickling down to students. While seven in 10 teenagers use AI for their schoolwork according to a recent Common Sense Media 

, less than a third have discussed what AI is and how it works with a teacher in class. Just 44 percent reported receiving guidance on when it should or should not be used for coursework.

Chicago Public Schools, the nation’s third-largest district, released what it calls a 

 on generative AI in 2024 that includes principles for integrating the technology into classrooms including transparency and accessibility.

The document, which has been updated several times since release, places a major emphasis on understanding what AI is and how it works, and the district has integrated AI literacy into its K-12 curriculum.

Students are encouraged to use approved AI tools as a “co-pilot” and to always double-check their output. CPS maintains a 

 of approved educational technologies that contains AI-powered tools, but notably does not include popular systems like ChatGPT and Claude. The guidebook does, however, include age restriction information provided by those platforms and others like Gemini and Perplexity.

CPS’s current AI policies have received some criticism for not being restrictive enough. One parent group dubbed Illinois Families for Public Schools has successfully 

 from half of the candidates on the ballot for Chicago School Board this fall to further limit AI use in the school district if elected. The pledge includes enacting a three-year moratorium on generative AI use, screen time restrictions, and the appointment of a chief privacy office.

The nation’s ninth largest school district, Houston Independent School District, also released a 

Alongside outlining principles for interactions with AI like prioritizing learning and ensuring academic integrity, the guidebook says students under 14 years old can learn about generative AI and use pre-approved tools as part of classroom assignments while supervised by a staff member, while those above that age can access approved AI tools with parental consent and permission from their teacher.

The guidance emphasizes AI literacy, saying that the goal of the document is to help students “understand what Generative AI is and is not, its opportunities and limitations, and ways to apply the tools judiciously to augment their own knowledge, creativity, and problem-solving abilities in their studies.”

 with OpenAI to give students and teachers access to an education-focused version of ChatGPT called 

 nine campuses “Future 2” schools with new programming aimed at preparing students for an “AI-enabled world.” Up to 100 schools in the district are reportedly set to adopt that model by July 2031.

Chicago and Houston have something in common—both consulted with the New York-based company AI for Education while developing their guidebooks.

AI for Education has helped over 40 districts across the country develop generative AI guidelines, CEO Amanda Bickerstaff told Tech Policy Press.

The organization encourages districts to adopt guidelines that both include rules on the limits of acceptable use and stress student education about how generative AI actually works.

“Every day, it just becomes clearer to me that AI literacy cannot be done well without guardrails, and guardrails cannot be done well without AI literacy,” Bickerstaff said. “Literacy must go hand in hand because the technology is too new. It works too differently than how you think it does, and it is too disruptive to be able to just come out with one or the other.”

Another district where that model has been deployed in partnership with AI for Education is Charleston County, a South Carolina school district with nearly 50,000 students.

The school district developed and adopted an AI policy over the last two years based on what it calls a S.E.E. framework. Before using a tool, the district wants students to ask themselves: Is it safe? Is it ethical? And, is it effective?

This policy has then been used to make practical decisions over what technologies should be allowed in schools and who should have access to them.

Guidelines “empower the school system to then act towards a common goal, which for us is to integrate generative AI into the system appropriately,” deputy superintendent Luke Clamp told Tech Policy Press.

Going into this academic year, teachers in the district will have access to MagicSchool, an AI-powered tool designed to help with lesson planning and creating classroom activities. Students may get access to the program down the line as well.

“We don't see this tool, [or] any other tool, to be a replacement for an in-person instructor in the classroom,” Clamp stressed.

Charleston has also focused on AI literacy for educators, which will then be passed on to students. Many of its teachers joined professional development sessions organized by AI for Education this summer, while others participated in an AI course at Clemson University. The district recently launched a resource 

 on its website with more information for teachers, students, and parents.

The attention paid by the school district to AI comes from a belief that the technology will transform education and the workplace, and that if students are not taught about how to use it they will lag behind others.

“We did not want to be behind the eight ball, if you will, on how we integrate with it,” Clamp said.

Los Angeles Unified School District (LAUSD), which serves over half a million students, released AI 

 for educators and students in April of 2024 requiring that tools be screened by the district for data privacy issues and that educators get parental consent for specific apps.

Alongside those guidelines, LAUSD launched a proprietary chatbot named “Ed” before 

 it down three months later. In February, Superintendent Alberto Carvalho had his home raided by the FBI, 

 in relation to his connection to the company behind the $3 million chatbot.

The district has seemingly made updates to its AI policies since the 2024 guidelines were released. As of a Los Angeles Times 

 in September of 2025, students under 13 were barred from using the technology in the classroom and those above that age were required to sign a responsible use policy, although those rules do not seem to be publicized on the district’s 

 screen time before second grade and instituted restrictions for older students that will 

 be rolled out this year and next. While not directly related to AI, this new policy will affect its use—and includes a 

 of the district’s contracts for classroom technology.

At Miami-Dade County Public schools, the third largest district in the country, the situation is even more hazy. The school board 

 directing the superintendent to develop guidance for AI in schools “including a tiered framework teachers can apply to classroom tasks and assignments,” in August of 2025. No such rules appear to have been released, at least publicly. In the interim, though, the district 

 with Google to offer Gemini to the district’s high school students and teachers.

North of Miami on the I-95 interstate highway, Orange County Public Schools—the eighth largest district in the nation, boasting over 200,000 students—released an 

 this summer allowing students to use the technology for research but restricting them from submitting graded assignments produced by it. The district places responsibility for developing guidelines on how AI should be used in the classroom to teachers.

“It’s like building a plane while we’re flying,” 

 Orange County District 2 School Board Member Maria Salamanca at the school board meeting where the policy was approved. “It’s moving so quickly, and I think that this board, above all else, mostly cares about safety and guardrails.”

As public opinion shifts against AI—typified by 

 opposition to data centers—there appears to be growing pushback to integration of generative AI into K-12 schools.

The New York City Department of Education released a 

 on generative AI in March that hit on many of the same notes as the guidelines covered earlier in this piece—setting green, yellow, and red light designations for implementations of the technology but largely leaving student use up to individual teachers.

The proposal was almost immediately met with strong backlash: hundreds of parents 

 saying it did not go far enough, and teachers organized for a moratorium.

The city’s education department said it was pressing 

 on new guidance for the time being, and has yet to release updated rules with students set to return to classrooms imminently.

Momentum for moratoriums on AI in schools has been building. The organization FairPlay recently released a 

 signed by over 200 organizations and experts calling for a five-year hold on AI citing risks to learning and mental health. The Brookings Institute also released a 

 concluding that “at this point in its trajectory, the risks of utilizing generative AI in children’s education overshadow its benefits.”

A handful of school districts have heeded that call, at least partially.

The Katy Independent School District outside of Houston announced in July that it was 

 generative AI tools for K-6 students. The Haddonfield School District in New Jersey 

School districts aren’t the only bodies trying to set rules around generative AI in education.

 official guidance or policies on AI in classrooms. A smaller subset of those actually require schools to adopt specific policies.

 a lot like what districts have been developing independently: guardrails on AI use, a focus on technology literacy for students and teachers, and an emphasis on the role it will play in future workplaces. Some of these state frameworks have proven 

The Center for Democracy and Technology has 

 for states to take on a greater role in the integration of AI into education, pushing them to focus on risks around data privacy and harmful use cases, restrictions on vendors, and greater transparency around technology that is allowed in class.

The federal government has also weighed in with the 2025 

 on “Advancing Artificial Intelligence Education for American Youth,” which allocates resources for AI literacy training for teachers, creates a task force dedicated to AI in education, and directs the National Science Foundation to prioritize training initiatives that expand AI use.

A laundry list of technology firms—including Google, NVIDIA, and xAI—

 to providing resources for youth training on AI under the executive order.

Artificial intelligence presents new privacy challenges, but treating AI as if it lives outside the parameters of ordinary privacy law is fallacious. The emerging lesson from privacy regulators and enforcers is increasingly clear: there is no AI exceptionalism. The same principles that have long governed responsible data use still apply. What changes is how and where those principles need to be built. In these nascent days, AI offers an opportunity to move beyond “traditional” compliance by embedding privacy, transparency, and user control into the architecture itself.

“Traditional” privacy compliance has often focused on collection, use, disclosure, and retention policies lined up with what a company says in its statements and privacy policies. The emerging reality with AI complicates this picture. AI infers, summarizes, remembers, ranks, and increasingly acts across blurred boundaries.

That shifts AI into a category closer to delegated action, where the user is not controlling nor even observing each action. And this delegated action relationship creates a different risk profile and attack surface for privacy harms. The core concern is not only that a model may process personal data. It is that the system may decide what data it needs, access more than the task requires, move information across contexts, and act before the user fully understands what happened. What data gets inferred, shared, or disclosed might go beyond what a user is willing to share.

Despite the broader AI exceptionalism arguments, regulators are not treating AI as an exotic category outside privacy law. Instead, regulators are identifying where familiar privacy principles become harder to operationalize. Across guidance and enforcement actions, the same concerns keep surfacing: AI collapses boundaries between collection and use, between one purpose and another, between user instruction and agentic initiative.

Regulators globally have used different language to describe these risks, yet key privacy principles remain consistent and expectations to integrate them into AI-related contexts are high. A few in-progress examples of articulating privacy rules in the context of emerging AI use cases include:

 cautioning that AI agents should have clear purposes for collecting and processing information and that selecting what databases agents can access is fundamental to data minimization requirements.

 emphasizing human oversight as a crucial accountability mechanism for output data and inferences made by an AI model.

Hong Kong’s Office of the Privacy Commissioner for Personal Data

 underscoring the importance of access right controls and continuous risk assessments as cornerstones for compliance.

Singapore’s Personal Data Protection Commission

 clarifying that web-scraping datasets containing personal data can be justified by the “Publicly Available Exception” in lieu of always needing to seek consent.

Australia’s Office of the Information Commissioner

 highlighting that privacy obligations apply to information inputted into AI systems, just as well as the information that is outputted.

In the United States, the Federal Trade Commission’s past enforcement actions provide relevant guidance for privacy design considerations when deploying AI:

 settlement order shows the pre-deployment testing the company must conduct before deploying AI for high-risk uses.

 matter demonstrates that data minimization and retention limits should be baseline 

 requirements to lessen the attack service and the risk of personal data exposure from a breach.

 with the FTC highlights that connected services cannot quietly collect, use, and disclose sensitive personal data for unknown purposes.

 matter demonstrates the need to respect users’ desire to delete data and ensure it’s not used for training models.

US state regulators have also been busy releasing guidance and developing new laws. Even without new laws, state Attorneys General can enforce existing unfairness and deception laws to address AI-specific harms.

Three key commonalities emerge from the privacy principles that regulators and enforcers are emphasizing across guidance and enforcement actions.

The first is that purpose limitations must become architected into AI services, and not merely contractual. In ordinary privacy programs, a purpose limitation often appears as a sentence in a notice: we use data to provide and improve services. For AI services that blur lines and can agentically act on a user’s behalf, this is not enough. “Improve productivity” or “assist the user” can become an over-broad and unclear permission to do nearly anything unless translated into technical boundaries.

In agentic systems in particular, setting clear purpose limits should be architected to honor users’ privacy and to reduce the risk of security failures. This means constraining which tools each agent can call, which data sets it can retrieve, which memories it can use, whether data can move between context, when the user must confirm an action, and whether outputs can be used for training or secondary analytics. This helps operationalize privacy requirements while adhering to longstanding security principles such as giving the least privilege to each role.

The second thread is accountable deployment. Regulators are skeptical of the idea that autonomy dissolves responsibility. An organization that builds, buys, deploys, or integrates AI remains responsible for how personal data is used. That means pre-deployment review cannot stop at asking whether the model is accurate; it must ask whether the system behaves within the privacy boundaries the organization promised.

Testing should include privacy-specific evaluations. Does the agent access data outside the task? Does it retain information unnecessarily? Does it resist prompt injection making system boundaries unreliable? Can the organization reconstruct what data was accessed, which tools were called, and why? Does the system perform differently across groups? What happens when the agent is wrong?

A third thread is that transparency and use control must move from notice to interface. Privacy notices cannot do all the work. Users need controls at the point of action. For AI services that store and use memory, this means developing memory dashboards that surface what the AI system remembers and allow users to edit or delete it. It means permission receipts explaining what data an agent accessed for a task. It means pre-action confirmations before an agent sends a message or discloses information. It means mode-switching between personal and workplace contexts. It means “do not remember this” and “why did you use that?” controls embedded into the product experience. These tools not only enhance compliance, indeed they help the user to improve the system experience.

These threads stem from principles enforcers and regulators have consistently noted are a core part of privacy requirements. They are not decorative compliance features, rather they are how privacy principles should be reflected in the design and operation of AI models and systems. The challenge now is to apply these principles to newer risks created by persistent memory, expanded permissions, and increasingly agentic systems.

Unlike the internet advertising economy, where privacy and profit have often been in tension, AI presents an opportunity to align them. Privacy, transparency, and user control can become part of the product itself. Features that strengthen trust, improve the user experience, and make AI systems more commercially attractive. The question for AI developers and deployers then, is not whether existing privacy principles can survive new technology. It is whether entities will seize the opportunity to translate those principles into the architecture, and ultimately the value proposition, of their products.

Audio of this conversation is available via your favorite podcast service.

How can human rights and social justice advocates and investigators protect their data and technology infrastructure in an age of rising authoritarianism in the United States and abroad? Earlier this month, a coalition of human rights, justice, and data organizations 

 the Technology Freedom Cooperative, a project designed to help nonprofit organizations protect evidence, reduce their dependence on big tech platforms, and "build independent capacity to work with large language models." 

What follows is a lightly edited transcript of the discussion.

My name is Maheen Khan. I'm the director of technology at the Invisible Institute.

I'm Patrick Ball. I'm the director of research at the Human Rights Data Analysis Group.

I am pleased the two of you could join me today, and we're going to talk a little bit about this new project that you are working on together, the Technology Freedom Cooperative, which also, as I understand, involves a range of other partners. We'll hear a little more about that. But I think I might just first ask you each to describe what your own organizations do so that the listener has a little bit of grounding for who you are, how you come to these questions that we're going to discuss today. Maybe Maheen, I'll start with you.

I work at the Invisible Institute. We are an investigative journalism nonprofit based out of the Woodlawn neighborhood in Chicago. And our main thing that we do is long-form investigations into policing misconduct. And part of the way that we do that is that we've gotten a ton of police misconduct data over the past decade through Freedom of Information Act requests, also known as FOIA requests. So yeah, our mission is kind of to use data and storytelling to help people hold their government institutions accountable, oftentimes that being the police. And yeah, I'm the director of technology, so I help with a lot of data analysis stuff that we do for stories and any kind of tech work that comes up. For example, we have a data tool called the Citizens Police Data Project where people can look up any Chicago police officer and see their misconduct history, that we've published on a tool.

So the Human Right Data Analysis Group provides assistance so that groups working on justice and accountability have the strongest possible scientific foundations for that argument. So we started our work abroad. My first project was in El Salvador in 1991, and we've worked in over 30 countries since then, trying to put together rigorous data analysis, inferential statistics, machine learning, and most recently AI analysis so that our partners can make arguments about violence by the state. So we work with truth commissions, with tribunals that are trying former officials. I've testified in the trials of three former heads of state, three former presidents, for example, and we work with lots and lots of partners. We've worked with the Invisible Institute for 10 years. And it's been an enormously fruitful collaboration and they are exactly the kind of partners that we have the most to offer because they go get all this amazing data and they have really clear things to do with it, arguments to make, stories to tell, communities to engage, and that's perfect. So this is a really happy collaboration for us.

We're going to talk about this new collaboration, which invites in other partners as well, this Technology Freedom Cooperative. I think before we get into the meat of why and what it's about, maybe since we're talking about what the orgs do, et cetera, who are your other partners in this cooperative?

So one is a group in Mexico called Data Cívica. Data Cívica does a similar kind of collection of largely public records, although they do some investigation, some collaboration with groups on the ground to study how the government of Mexico uses funding. They work a lot on the transparency records. They look a lot at violence, both violence against women, el feminicidio, the killing of women because of their gender. They look a lot at disappearances and they look a lot at the recovery of the human remains of disappeared people. That's one of the projects we've collaborated with them most closely.

So our partner in New Orleans is Innocence and Justice Louisiana, which is a group that spun out of the Innocence Network. They largely work on getting people who have been unjustly convicted exonerated and released from prison. And we work with Kilimetro Cero in Puerto Rico that has had a series of just stunning successes in their work exposing and demanding reforms of violence from the Puerto Rican police, who have done a lot of pretty terrible stuff over the last decade. And Kilimetro Cero has managed to get a lot of reforms into place to reduce that violence.

This cooperative has a bunch of different purposes. You talk about bringing them together to protect critical evidence, to reduce dependence on big tech platforms, to build independent capacity, to work with large language models. I want to talk a little bit about each one of those things and what you plan to do around them. But I thought I might start by just asking what's the motivation for this cooperative to come together right now? Was there a specific set of events or I suppose a confluence of things that have happened that made you feel this was the moment to launch this new initiative?

Well, I think a big part of the timing is, yeah, with this current administration, the Trump administration and their hostility towards journalists and just people trying to keep records. So I definitely think that was one of the catalysts of why now. Even though I know Patrick and the folks at HRDAG, and I think so many other groups that have accumulated a lot of data over however many years, I do think everyone's really been thinking about our reliance on these big tech companies, that they don't have any sort of allegiance to us or human rights abuse, uncovering human rights. If they decide on a day that they want to close our accounts, they can do that on a whim. So I think these are things that I know a lot of people have been thinking about for quite some time as we've been slowly descending into fascism. But I think this has felt like the right moment, where we're seeing how quickly we're descending into fascism, like speed up.

Yeah, there was a particularly triggering event for us near the beginning of this administration when the Trump foreign policies team decided to use one of the most powerful levers in the foreign policy powers, which is to place a person on what's called the sanctioned persons list. And the sanctioned persons list was a creation of the 9/11 global war on terror era policies in which putting someone on this list means that they are basically no business that does any business in dollars is allowed to do business with them. This was a list created for the world's worst terrorists. It was created for Osama bin Laden. Yeah, well, the Trump administration came in and decided to use this against the international criminal court. They placed several of the senior officials and judges at the international criminal court on this list. And the effect of that is those people lose access to doing any business in dollars. So that means you can't use a bank, you can't have a cell phone account.

You can't do things that you have to do in dollars because every multinational company in the world does business in dollars, so you just can't do business with them. So people, the story that one judge told is that she came home and said, "Hey, Alexa," and nothing happened because Amazon turned her account off, Microsoft turned her accounts off, and Google turned her accounts off. And this just should be an absolute red flag for anybody who thinks that democracy is strengthened by vigorous debate, because people debating with the Trump administration are now literally in the same category as the world's worst terrorists. This is not just an analogy, indeed, we're using exactly the same tools against them. And I looked at that and said, "Well, what if they turn that against domestic groups?" Certainly, many administrations going back into the 1970s have used de-banking as a tool against domestic critics. So a domestic critic, someone that the administration doesn't like, is prohibited from being part of the banking system.

So if you were a cannabis retailer in a state where that was legal, you couldn't have a bank account. If you're a sex worker or an organization of sex workers, you can't have a bank account. If you were a supporter of Chelsea Manning, you couldn't have a bank account. This is all really well documented. There's a great new book about it called Transaction Denied, I recommend for folks. But what it does is it really should open our eyes to the kind of non-legal powers. I don't want to say illegal because it's not clear if they're illegal or not, but they're not inside the justice system. These powers that the administration has to manipulate private sector actors against the critics of the government. And those of us who are critics of the government should take that very seriously and ask, "Who are we going to do business with?"

Patrick, Maheen, what I understand you're going to do is try to build infrastructure for organizations who are trying to collect human rights violations, evidence, state violence evidence, perhaps attempt to take on wrongful convictions, other abuses of power. Maybe bring this home for us, Maheen, in the context of your work there in Chicago. We've seen of course a significant state intervention there and extraordinary spectacular ICE CBD raid there. We've seen a lot of activity around that in your city. I don't know, how can you imagine this infrastructure potentially helping you?

Yeah. Well, to take a step back, Invisible Institute was founded in 2014. And one of the catalysts for that was that the founder, Jamie Kalven, had filed a lawsuit that resulted in establishing that in Illinois, police misconduct files are public information. You are allowed to FOIA for those. And that was the beginning of Invisible Institute. Because of that lawsuit, the org got a ton of police misconduct files in all these different CSVs. And the question was, okay, this is a lot of data and okay, now anyone can theoretically file a FOIA request to get this data. But FOIA requests can cost money, it can take a while to get that data back, and it can be kind of difficult to parse through that. So that was why we built a tool where people can just look up officers and see their misconduct history. But yeah, since 2014, we filed so many more FOIA requests and have been adding on to the different types of data that we get.

We also have another tool where you can see employment histories of officers within a state. So yeah, this has been an accumulation of data that we've collected for over a decade. And if one day we were to just lose access to all of that, we would just lose so much of the work that we've done for over a decade. And I think related to, like with Midway Blitz, which occurred in Chicago, which was when a lot of customs and border patrol agents, ICE agents were in Chicago, we were part of a collaborative with some other newsrooms in Chicago where we were collecting video and photo footage from people who were witnessing ICE agents detaining people. And that's also data that we wouldn't want to lose. So yeah, I think for us, it is really preserving all of this data that we've accumulated over a really long time that I think would be very hard to...

We couldn't just file a FOIA request to get it all back. So I think for us, it is really like we're not necessarily scared that the government will see these files that we've got from FOIA requests because we're getting them from a government agency anyways, we're nervous about us ever losing access if... A lot of our stuff is hosted on Google Drive or Dropbox, if they decide to close our account tomorrow, we would lose all of that. So for us having it in a server that's separate from all these big tech companies and it's hosted by us and by a really amazing infrastructure that Patrick and other folks at HRDAG have spent months building, for us, that's where it feels so important. We've been collecting so much data for so long and we want to make sure to preserve that and have access to it.

Patrick, walk me through the mechanics. How is this set up technically? Who owns what physically? Where are you keeping information? How are you splitting things up across jurisdictions? What's the kind of technical approach?

So the idea is first, as you start out with the lead, is to distribute the information, to decentralize it across lots of different places, and lots of different places and lots of different owners and lots of different jurisdictions. Those things all go together in the notion of place. But I want to say it's not just different houses or offices or co-location facilities, it's different legal regimes and different countries and different disaster zones, because disaster is also one of the things you have to worry about when we're talking about data redundancy. So what we do is, we say, "Okay, each of the five organizations in the cooperative starts out with a server." And that server, they put whatever information they want to put into our archive into that server. They scrape it down from their Dropbox accounts or their G drive accounts or their Amazon S3 accounts or Microsoft Azure or SharePoint, or whatever the tool is that they're using in the cloud to store the data.

We make copies of all that data on their local server. Now, they can keep using the copies. There's nothing wrong with using those cloud copies as long as we have access to their accounts. So we're not trying to displace that usage necessarily, we want to be sure that we can respond if the companies cut us off for whatever non-legal reason they make that decision to do that. So okay, so now we have the data on this local server. Great. Let's bundle it up, bundle it up in little chunks, and the chunks are about a gigabyte each. And the reason we do that is it's just easier to handle things when they're in standard size pieces, to handle them and move them around. If we have hundreds of thousands of tiny little files, that becomes a big sort of catalog headache. You have to keep that proper.

So the first thing you do is you make a catalog of it all, put it in a database, and then bundle it up into these one gigabyte chunks. And each of these one gigabyte chunks has a bunch of important properties. First of those properties is that we know what's in it. So we have a catalog. The second is that we know what's in it in a rigorous way, which means we have a digital sort of fingerprint, called a hash, that represents the authenticity of each file and we store that separately. The next thing that we know is that we know who owns that data. Whose is it? It's the Invisible Institute's data. And we know that because that whole bundle carries a digital signature which can only be created by the Invisible Institute. So the Invisible Institute has signed that commit. We know it belongs to them. So we know the content of it, we know it belongs to them.

We do one more thing here, which is that we take the digital signature of the whole catalog, which, for nerds watching, is the Merkle root of the hash tree across all the files that we've put in this bundle, and we store that on the Bitcoin blockchain. And the reason we do that is that once that's on the Bitcoin blockchain, now we have a timestamp. So that's another guarantee that this data existed at or before the time that that timestamp was created. None of that is, the term in cryptography is repudiable. It cannot be repudiated. All of that is provable mathematically, that these are the files, this is who owns them, and this is when they were created. Then we encrypt all that data with a key that only the partner holds, and that's one commit. Now we repeat that, I think we're up to about 18,000 times across each of the five groups. Sorry, 18,000 times, that's just Invisible Institute. Sorry, it's 40,000 times for the whole group.

Invisible Institute's got the biggest store, but the rest of us contribute quite a bit to it. So now the commits, we have each commit, each commit lives in the server that org owns. Now the commit gets distributed across the entire network. So we have about 20 servers now, and they're distributed across Mexico, the United States, Canada, and a European country that I'm not going to name here. And each of those servers is independent. Nobody tells any server what to do. Every server has shared a rule set, but the idea of decentralization is nobody's saying, "Oh, you over there, make these copies." Instead, the one over there says, "Hey, I see that the Invisible Institute server has published a new commit and they say that they'd like to have five copies of that commit. Am I the right server to make that copy?" And it uses an algorithm that was developed way back, more than 20 years ago by a project called BitTorrent that decides where the files should go. And using that decentralization algorithm, it says, "Oh, yeah, I'm one of the servers that should take a copy of that."

So it reaches out to the Invisible Institute server and grabs that commit and pulls it down. For people who want to know, all the computers are inside what's called a tail scale network or a tail net, which we self-host on a headscale server that we manage in one of our co-location facilities. Each server exposes a specific read-only location to other machines that are doing this replication so they can pull files from each other. So now our other server has pulled that file across and at least three others have as well. So now there's five copies. So there's five copies that are distributed. There's some other rules about jurisdiction routing and location routing. Because my office has six different servers in it, so it doesn't do us any good if all the copies are in one office. So there's all sorts of these little more subtle rules to make sure that a commit gets to, well, there's one in Chicago and there's one in Puerto Rico and there's one in Canada and there's one in Europe.

So let's just... Maybe to make it real though, someone comes into your office there, maybe they knock down the door or maybe something more, I don't know, normal happens, like the apartment floods, what have you, those six servers you're running in the place are gone or otherwise disconnected. You're saying that effectively you have liberated this information to parts unknown and that won't matter?

Well, it'll be a hassle, but it'll be a manageable hassle. It'll be a recoverable hassle. In fact, just this week I successfully... I did a drill in which I replicated the entire store of one of the organizations with just the key. So all you need is the decryption key from the organization who was lost, in this sense, and you need another place to put the data. And so I had those things and I put the key and the other place to put the data and I was able to recover from machines that are not the one that was, in the drill sense, lost. So if somebody comes to our office and all the computers there for whatever reason are gone or unavailable or destroyed, then I could rock up in another city and with a machine that was adequate to get all our data and with all our keys, which are very carefully protected and backed up in encrypted and distributed ways, recover all the information to this new server and restart our project.

, he talked about it as a kind of preparation for a post-American internet, I think is the way he kind of characterized this. How do you think about that? You're effectively trying to make sure that you can work across borders, that you're not reliant on infrastructure that could obviously be subject to the long arm of the American government or perhaps other governments. Where do you situate this in that thinking about United States and internet freedom within it?

I think for me, part of this project I think is also fighting against how technology and new tech tools or tech companies function under capitalism. I've been really interested in technology for a long time. I studied computer science in college and I think I've just thought the idea of coding and software and hardware has just been very interesting. But I think I have also been pretty turned off by I think the fact that it feels like if you're working in tech, the whole thing is like, "Oh, yes, what's the newest thing? What is the thing that's going to make us even more efficient? And what's going to help these companies make more money as we do these things with these tech tools?"

And I think that this cooperative is kind of pushing back against that, and I think instead is bringing together people who value technology and really see that it can help us create and build new things that we couldn't have thought of. But that it's not for the purpose of just scaling or getting big or making more money, instead, this is to push back against that and be like, "We can build these really interesting new technologies and these infrastructures, but for a bigger purpose, for preserving these human rights violations that we're all documenting. And it's not just about becoming more efficient and being able to scale up." So I really see it as kind of pushing back against that.

And I think especially in the current context with how AI is being pushed everywhere, I'm someone who is very interested in technology and these, but I'm also extremely critical and extremely skeptical. I switched from Google search to DuckDuckGo a couple months ago because I really didn't like that Google Search will not let me turn off AI search results. And I don't want to be forced to opt into all of these AI things. I want to be able to be critical and think about, "Oh, what are the use cases in a contained way where it actually makes sense to be using this specific tool?" And I feel like the TFC is a space to do that and is a space to show people technology doesn't just have to be, for these efficiencies, in order to make these companies richer, it can actually be for people and controlled by people. So that's how I've seen it as. And I feel like that's what I was looking for so long when I was taking computer science classes or trying to learn more about technology, but it's all been from such, oftentimes, an apolitical lens.

I want to really agree with Maheen that I object to the conflation of technology and capitalism. They're not the same. Now, most technology gets built by big capitalist companies, for sure, but there's plenty of tech that is built in other ways. And I have always been a huge fan, going back to the mid '90s, of technology that's built for people who are trying to advance human freedom. That's the goal that I think we should be having. How can we increase human dignity and human freedom with our tools and with our work? And that's what the technology freedom cooperative is about. It's about using technology to advance freedom rather than... We're not being free of technology, to the contrary, we're embracing technology to build tools that'll lead us in this direction.

So to Corey's point that you asked about a post-American internet, I would ask us, "Why do we..." I'm not objecting to America or objecting to China or objecting to Europe, I'm objecting to government. I'm objecting to corporations. I'm saying, "Hey, why don't we build stuff for us, for people? Those other big structures are just problems to route around. They're not helping us now." And indeed, many are working, not maybe deliberately to harm us, but they're working in ways that other actors can turn to harm us. I don't think that, for example, big technology companies that built surveillance into every single piece of their tooling did so because they intended to destroy our ability to organize politically, but that's the effect.

That may not be the intention, but that's the effect because that's what is being weaponized against us by the people who say, "Hey, gosh, you've collected all this information on citizens. Why don't we use that to surveil them for political reasons?" Well, that's terrible, and we need to recognize the unintended consequences, or whether they're intended or not is irrelevant to me really, the consequences in any case of technology that, just like Maheen said, is driven by scaling and profit and surveillance. That's not good for us.

Part of the announcement of this initiative is that you intend to build independent capacity to work with large language models to help develop local AI infrastructure for organizations that are working on human rights issues. What does that entail? How are you planning to do that in a way that stays within the constraints of some of the things that you just said?

Well, we certainly do not have the capacity to train our own models. That's too expensive. What we have is a bunch of GPUs, a bunch of machines that are capable of running large language models, and we're up to, I guess three machines, and I hope to be up to five by the end of the year, of machines that are in independent locations and jurisdictions, again, that provide us, with everybody in the cooperative, a way to put queries through. And those queries can do all kinds of really valuable and useful things. I think some of the ones that I've found most powerful scaling are when we get, for example, hundreds or thousands of hours of body-worn camera footage in court disclosed... Court orders the police to disclose this footage to different kinds of projects that we're participating in.

We get the footage and we can use LLMs to watch the footage and identify officers and specific kinds of events and link them across a whole set of videos, and then later across the underlying documents for use of force or for the stop records or whatever it is that is the useful information in order to enable defense lawyers to demonstrate that what the police are saying is untrue in crucial ways. And that's really important to be able to do, but that there's just not enough human attention on the planet to be able to read all that stuff or watch all that video and do all that classification by hand. So an LLM can enable us to do that. Like I said, we decentralize that, but we also then enable our partners, like Invisible Institute or the other partners, to be able to use those models in their workflow to enable their work. And Invisible has incredibly exciting projects doing this kind of stuff. So I would like to stop, and Maheen, I'd love to hear your descriptions of some of that stuff.

Another example of, I think one of the services that I think will be really useful on these servers is the transcription service. Right now, we pay for transcription services for when we interview sources or something and need to transcribe those interviews. If we have our own closed server that can do that for us, that's huge because then also you're not uploading audio to this company that you pay a service for, instead, it's in this closed system. So I think that feels like a really interesting use case for me. And then we've previously used machine learning tools for really digging into complaints, like a trove of complaints that we have against police officers to really understand what is actually happening in those complaints. Because complaints get one category basically, but a lot of times that category doesn't tell the full picture at all of everything going on under the hood of that complaint.

So I think using LLMs have been really helpful in the past of actually figuring out what's actually going on in these complaints against CPD officers. It's actually not telling you the whole story when it says, "Oh, this was an illegal search case or operations violation." So I think being able to really extract stuff that's deep in the files of those types of complaints is really helpful to see what is actually the magnitude of what is going on with people feeling violated or hurt by the police. I think it's really exciting that we might be able to experiment with different ways that AI can support us in doing that work, but in very closed environments, where we're really deciding how we're using it and also just what types of models we're using.

There is so much discussion about artificial intelligence, particularly around the dangers and the fears that many groups have about ways in which the technology will negatively impact questions around social justice and equity, but it seems like you're really trying here to look at a way to take advantage, I assume, of open way models and other open source technologies to see if you can go the other direction.

The cool thing is that for most kinds of technology, if big powerful players can use it to take away people's freedom and dignity, the rest of us can use that technology or a version of it that's maybe freer to reaffirm and to insist on justice and dignity. And I think database technology gives us a real historical memory of the kinds of usage that we can put to machine learning back until a few years ago, and now LLMs.

I think all of these technologies do both, and a lot of times we forget about the way that we, as people who want to advance justice and dignity, can use these tools. And this project is, it's not really our intent, but one of the things that will come out of it I think is these examples of what we're doing with these technologies, which certainly can be used to do horrible things and will be used to do horrible things. Let's not be naive. Palantir is nobody's friend. They're going to do horrible things. They're already doing horrible things. I have actually a list of horrible things they've already done, but we can use similar technology to do good things too.

I think just to go off what Patrick is saying, like I said earlier, I'm super critical and skeptical of when we're all talking about AI. And I think even just saying AI, it's like even that is such a broad term that can mean so many different things. Are you talking about generative AI? Are you talking about the chatbots or are you talking about a transcription service? And I do think that bringing it back to capitalism, that under a capitalist society, I do think any sort of new technology, I do think the focus, for one, the companies that are the ones behind building, it's like, yeah, their priority is how can we keep growing it, scaling it out so then we can make more money and be more efficient and cut costs in these way? I think that's unavoidable under the type of world we live in at this time.

And I definitely think a lot of the ways that AI is being used currently is really detrimental to society. I think AI is being marketed as solutions for I think very real problems that exist. For example, I do think people are very lonely. Because I think we live in a very individualistic society where it's really hard to build community and just have a lot of people around you who live near you and can offer support to one another. I don't think the solution to that is for people to have a chatbot that becomes their best friend and the only person they talk to. So I definitely have a lot of issues with how AI is trying to be marketed as that solution to a lot of real problems that exist. And I think that's why this cooperative is really great because I think it is able to show these very concrete ways of using AI tools that are genuinely to help us get away from those big tech companies that are actually pushing it onto us in these harmful ways.

So yeah, I think I have huge problems with how it's being pushed on all of us. And also I feel like at the same time, it is really important to understand, here are these very explicit ways where it can be useful and you don't need to scale it out so much so that it's really harmful for the environment as well. Part of why AI... Everyone's talking about how harmful these hyperscale data centers are for the environment is because of how big they're getting and how much they're expanding, because these companies keep trying to make AI bigger and bigger. And it's like, that's not what we're trying to do. The whole point is we're trying to keep it very contained and very much fulfilling these certain needs that can help us do our work better, but we're not just trying to use AI to use AI because it's the new thing.

I understand the cooperative is intended to serve its partners, to start. That's the clear goal of it. So I guess on one level, why publicize it? I do know that you were also planning to, as you say, Patrick, release findings and share with the community, but what's the goal? What's the goal of even coming on a podcast like this or telling others about what you're up to?

So we very self-consciously are organizing as a cooperative. And the cooperative has members. We're not intending, as Maheen has emphasized, we're not intending to get any bigger, but I sure would love it if other groups who all trust each other want to create their own cooperatives. That's the scaling model if there is a scaling model here. It's that if other groups want to put it together, well, gosh, we'd love to help. And I've written all this technology and it's all open source and I'd be really delighted to share it with them. And I'll even help them get it started and I'll help them figure out what hardware they need to buy and how to provision it and how to think through where you put it and stuff like that. That would be really fun. And I think the big reason to go public is to try to spur a conversation, a horizontal conversation among groups that are working to advance human freedom and dignity and justice to think about, "Well, okay, what can we do with other groups?"

So maybe there's groups working on climate change or there's groups working on reproductive justice or there's groups working on LGBT issues or there's groups working on, fill in the blank, whatever that groups are. And they're groups that have affinity, that trust each other, that are connected to each other, that say, "Okay, well, we're going to throw in together a little bit here and build a technology infrastructure for ourselves that we control. Oh, sure, those folks at TFC will help us out. They'll share some tech with us and we'll have meetings with them." And I'll be happy to maybe show up and help them assemble some computers. That's one of my favorite things, but it's theirs and they control it. And whether it succeeds or fails, is up to them. And their focus is... Ours is up to us. And so I think that's the big reason to make it public, is to inspire this idea.

I have a kind of higher order hope for it. I'm not sure. And that's kind of my hope for this podcast, is to say, "Why do we let big companies be the people who define what technology is and what it's for?" It's not. It's ours. We're human beings. We make tools. It's what we do. Tool making is one of the sort of fundamental things that makes us people. I'm not going to let some big company say that they're making tools and I'm just a consumer of those tools. No, to hell with them. I'm going to build the tools.

So that's the other piece of it is to give some sense that this is not just possible... It's not even that hard. This stuff's all out there, I'm just plugging it together. I'm writing a ton of software, but when I say writing software, I'm gluing together all this other stuff that all these other people have written. That's what the free software movement does is it puts these pieces out there and then we can tie them all together for other purposes. And so I think that's the reason that we want to go public.

Yeah. I think people just need to really see that there are these ways of using technology. I feel like technology for a lot of people feels like this weird black box of just like, yes, the internet just exists, these things exist. And I think it can feel very much like we have absolutely no control of this just because it feels so technical, which it is. And I think it is really important to show people there are these creative and radical ways to be using this technology and taking control of it and deciding for ourselves what we want to do with it.

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe