Welcome to the 13th edition of HRF’s AI for Individual Rights newsletter.
Across this week’s news, authoritarian regimes continue deploying AI-powered surveillance to better identify and track critics. In China, an unsecured police database revealed how facial recognition and personal data can be combined to monitor foreigners. Meanwhile, police in Pakistan’s capital are discussing ways to expand technology-powered policing and surveillance.
This week also brought an encouraging shift toward user-controlled AI. Meta committed to developing AI that empowers individuals and released a new open-weight model. Just one day later, NVIDIA unveiled an open-weight model of its own. These are models that individuals anywhere, in any country, under any government, can download, modify, and run privately. For dissidents, that means having access to intelligence without exposing their activity to state-controlled companies or authoritarian regimes. These developments suggest that major AI companies in democracies may be starting to invest more seriously in open models, giving users more alternatives to the Chinese models that currently dominate much of the open-weight landscape.
We end with a practical tutorial from freedom tech educator, Sovereign Sessions, that shows beginners how to use LM Studio, an application that lets users run AI models locally (and, therefore, privately) on their own computers. If you have a few minutes today, this is worth your time.
The Latest in AI for Repression
China Uses AI and Big Data to Track Foreigners
A New York Times article uncovered an unsecured Chinese police dashboard, apparently built for officials in Zhangjiakou, that aggregated personal data and facial recognition on foreigners. The platform held nearly 12,000 entries, including foreign journalists (some of whom had not been to the city), people labeled as fugitives, and people from Hong Kong and Taiwan. It combined passport details, phone numbers, addresses, relationship networks, financial data, and camera footage from China’s expansive AI-powered facial recognition network. While it remains unclear how the police have used these databases, it shows how facial recognition feeds and data collection put foreign dissidents, journalists, and critics under an increasingly watchful eye.
Why this matters: AI could turn massive surveillance databases like this into ones that are far easier to search and analyze. What was once an array of personal data could become a system that Chinese officials use to quickly identify, connect, and track dissidents.
Police in Pakistan Meet to Discuss Strengthening Surveillance and Smart Cities
Police met in Pakistan’s capital city, Islamabad, to discuss accelerating the city’s use of technology and Safe City technologies for policing and surveillance. While officials did not specify which technologies would be expanded, Islamabad’s existing Safe City network already uses AI-powered facial recognition cameras, vehicle tracking, and other surveillance tools. Expansion of such systems makes it easier for police to identify people in crowds, track dissidents’ movements, and investigate them after they attend protests.
In context: This surveillant expansion comes amid a deadly crackdown on protests and independent reporting. In Pakistan-administered Kashmir, demonstrators protested election rules, while dozens of protesters have reportedly been killed and officials have imposed restrictions on journalists. This already repressive and violent hybrid authoritarian regime could use AI-powered surveillance to further crush its critics.
Chinese Communist Party (CCP) Sanctions HRF Grantee Human Rights in China
Amid an escalating trade dispute with the United States, the Chinese regime barred HRF’s AI for Individual Rights Fund grantee, Human Rights in China (HRIC), from engaging in trade or other activities with Chinese entities. The sanctions came in retaliation for the US government’s recent restrictions on drone imports and Chinese companies linked to Uyghur forced labor. China’s sanctions were intended to pressure Washington to reverse those measures. But HRIC is not a commercial company or an economic competitor. HRIC is a New York-based human rights organization, founded by Chinese activists in exile, that documents CCP abuses and supports Chinese activists. Beijing framed the sanctions as part of the broader trade dispute, but it appears to provide a convenient excuse to target those working to expose its abuses.
Why this matters: The sanctions on HRIC are widely viewed as Beijing using economic pressure as a tool of transnational repression to target critics abroad.
Nigerian Lecturer Urges Regime to Adopt AI-Powered Surveillance System
Adetoro Banwo, head of the Chinese Unit at the University of Lagos, urged Nigeria to adopt China’s AI-powered surveillance model to combat terrorism and insecurity. He argues that AI could “spot early signs of unrest and identify people of interest through big data, computer modelling and machine learning.” But such tools could be turned against peaceful dissent. Authoritarian regimes like Nigeria often invoke security threats to target opposition figures, protesters, and critics. An AI-powered surveillance system could make it easier to identify and track them at scale.
Hong Kong Looking to Add AI Facial Recognition to Drones Enforcing a Construction-Site Smoking Ban
Officials in Hong Kong are reportedly looking to integrate AI-powered facial recognition into drones being tested to detect illegal activity. Currently, these drones use infrared cameras to identify people violating smoking laws — a minor offense. But adding facial recognition could allow the authoritarian regime to better identify and track protesters, journalists, and critics remotely and from the air.
In context: Hong Kong has increasingly treated peaceful political activity as a crime since the imposition of a national security law in 2020. Officials have arrested opposition candidates for organizing an unofficial primary, imprisoned journalists over their reporting, and sentenced a man to 14 months in prison for wearing a T-shirt bearing a protest slogan.
Recommended Content
A Roadmap for Confronting the Chilling Effects of Censorship, Surveillance, and New Technology
A recent episode of the Tech Policy Press Podcast features an interview with Jonathon W. Penney, author of Chilling Effects: Repression, Conformity, and Power in the Digital Age. Penney explores how digital surveillance can encourage people to self-censor, conform, or disengage from political life even when they are not directly threatened or punished. He also warns that increasingly powerful data analytics and AI systems could make these chilling effects more targeted and easier to scale.
The Latest in AI for Freedom
Meta Says It Commits to AI that Empowers Individuals, Releases Open-Weight Model
Meta CEO Mark Zuckerberg posted a 6,500-word essay explaining that the company’s approach to AI will center around empowering individuals. He argues that concentrating powerful AI in the hands of a few companies could leave individuals with less control and privacy. To prevent Meta’s own technology from contributing to those risks, he writes, the company plans to build private personal AI agents and continue releasing open-weight AI models. An open-weight model is one whose trained parameters (weights) are publicly available. Anyone with suitable hardware can download it and run it locally, including offline.
On the same day as this announcement, Meta released Muse Glimmer, an open-weight model optimized to power private AI agents running on a Mac or PC. Hopefully, this is just the beginning of a new wave of Meta tools that dissidents can run privately, on their own terms, to supercharge their work.
NVIDIA Releases Open-Weight Nemotron 3.5 Lightning
Just one day after Meta’s statement and release of Muse Glimmer, another US technology giant, NVIDIA, released its own open-weight model, Nemotron 3.5 Lightning. Similar to Meta’s model, it can run locally (therefore, privately) on personal computers and was designed to power AI agents. The model is optimized for common agent tasks, such as reviewing code, using external tools, and answering routine questions. Its release gives individuals and dissidents another powerful model developed in a democracy that they can run privately to supercharge their work. Try it out on Ollama, an application for running AI locally.
Why this matters: Chinese models like Kimi K3 and GLM-5.2 are among the leading open-weight systems. Meanwhile, many leading AI companies in democracies often keep their top models closed. These back-to-back releases signal that top tech companies in democracies may be breaking this pattern and moving toward a future where open AI is not shaped by authoritarian regimes alone.
Kimi K3 Live on Privacy-Protecting AI Sites
The Chinese AI model, Kimi K3, is now available through several privacy-protecting AI services, including Tinfoil and PayPerQ. Kimi K3 is one of the most capable open-weight AI models to date, but it requires expensive hardware to run. That’s why Tinfoil, a privacy-focused AI company, runs the model on its own powerful hardware and lets users access it privately in a simple chat interface. Tinfoil says prompts are encrypted to a hardware enclave so that Tinfoil itself cannot read them. That is the company’s claim; it depends on the security of those enclaves, and it is not the same as running the model on your own device.
Another privacy-protecting AI service, PayPerQ, also offers this end-to-end encrypted access to Kimi K3, while adding an additional layer of privacy. Users do not need to create an account, and can pay in Bitcoin to protect their financial privacy, rather than with identity-linked payment methods like credit cards. Activists and dissidents can use Kimi K3 through these services without buying costly hardware. The privacy benefit is only as strong as the provider’s encryption and enclave design, and the model can still reproduce CCP censorship and propaganda.
Important: Although “best in class,” this open-weight model can still reproduce censorship and propaganda embedded during its development in China. Tools like Tinfoil and PayPerQ do not address this, but they do address a separate risk — sensitive prompts being sent directly to a Chinese provider.
Maple AI Announces Plans for Privacy-Protecting AI Agent
Privacy-focused AI company and HRF grantee Maple is preparing to release a new privacy-protecting agent. In Maple’s app, users simply switch to “Agent Mode” and watch the AI work its magic. It will be able to work directly with local files, build websites and apps, and connect to external tools, without sending that data to a conventional closed AI lab. For small tasks, the agent can use models that run locally on a user’s device. For heavier tasks, Maple says it can route work through privacy-focused providers such as Tinfoil, which is still a third party. Maple says the goal is to offer a simple way to use agents without requiring users to hand over sensitive work, personal information, or company data to closed AI labs.
Why this matters: AI agents can dramatically expand what activists can accomplish. A private, easy-to-use agent like Maple’s could put that power in activists’ hands without requiring them to hand sensitive files, strategy, or personal data to a conventional closed AI lab.
SayIt Offers Private Text-to-Speech
Pioneering freedom tech developer Calle released a free, open-source Mac app, SayIt, that uses local, private AI to transform text into speech. Users can download the app and choose a small, open-weight model that runs locally on their Mac and can read documents, articles, or other text aloud. Because everything runs locally, activists can listen to sensitive documents or catch up on articles while multitasking — without paying a cloud provider or sending that text off the device.
Why this matters: Not every AI task requires sending data to powerful cloud models like ChatGPT or Claude. Small, specialized models can increasingly handle niche work entirely on our own devices. SayIt is a simple example of how small, local AI models can perform useful, everyday tasks without sacrificing user data.
Recommended Content
LM Studio: Private AI on Your Own PC
In this video, freedom tech educator Sovereign Sessions, also known as BTC Sessions, teaches beginners how to run AI locally on their computers with LM Studio. He walks through how to install the application, choose the right model for your devices, and understand some basic AI terminology. Give it a watch to learn why keeping AI under your control matters, and how to get started today.
