HRF's AI for Individual Rights Newsletter #21

Welcome to the 21st edition of HRF’s AI for Individual Rights newsletter. 

El Salvador’s President Nayib Bukele is rolling out AI-powered public services that could boost his popularity — even as democratic checks on his power continue to erode — and expand his access to citizens’ personal data. In India, officials ordered Apple to remove Bitchat (an open-source, encrypted messaging app that works offline) from its App Store, and the app also disappeared from Google Play, after protesters used it to stay connected during internet shutdowns.

On the other hand, French AI lab Mistral has released a preview of its most capable model yet: Mistral Large 4, or “Le Chonk.” The company plans to release the model’s weights soon, which could let organizations run a powerful model on their own servers, or let individuals use it through privacy-preserving hosts. Likewise, an NVIDIA-backed American AI company, Reflection, announced Beam, another new open-weight model, due out later this month, that could act as an alternative to the Chinese models that currently dominate the space. 

We close with a recap video from our recent AI Hack for Freedom III, where 10 activists used AI and the help of a few developers to build tools to address the biggest challenges they face.

THE LATEST IN AI FOR REPRESSION

Bukele’s AI Initiatives May Deflect Criticism and Raise Privacy Concerns in El Salvador

A new article in The Economist argues that El Salvador’s strongman president, Nayib Bukele, is implementing AI-powered education and healthcare initiatives that could help him improve his popularity and image rather than be seen as a “run-of-the-mill dictator.” His hybrid authoritarian regime is planning to use xAI’s Grok chatbot as an AI tutor for students in math, science, and reading. It also launched DoctorSV, an app developed with Google that uses AI to screen patients before connecting them with doctors. These programs could improve public services, but they could also help shift attention away from Bukele’s growing concentration of political power. The article also cites fears that these experiments could put more of Salvadorans’ personal data in the hands of his government.

In context: Since taking office in 2019, Bukele has removed presidential term limits, filled the constitutional court with his allies, and reshaped electoral rules and institutions to strengthen his political power. AI tools and experiments could strengthen Bukele’s image even as those democratic checks erode, all the while granting him access to users’ personal information that could make surveillance and repression of critics and civil society easier.

Bitchat Removed from App Stores in India

Bitchat, an open-source messaging app that works offline using Bluetooth, has been removed from Apple’s App Store and Google Play Store in India, while its website has also become inaccessible. Apple said officials ordered the removal under Section 69A of the country’s Information Technology Act, which allows the central government to block online content on grounds such as “national security and public order.” The Internet Freedom Foundation called the order unconstitutional, arguing that Section 69A allows the blocking of unlawful information, not of a messaging app because it works during internet shutdowns. India’s hybrid authoritarian regime appears to have targeted Bitchat because it lets people communicate during internet shutdowns and makes surveilling individual communications more difficult. Despite Bitchat’s removal from app stores, its open-source code has already spread to other repositories. Additionally, Android users of Bitchat can install the app via device-to-device transfer or by downloading the APK from a trusted source.

In context: In July, student protesters angered by exam paper leaks faced targeted internet shutdowns and signal jammers. The protesters turned to Bitchat to continue communicating, with daily active users in India peaking at 430,000 on one day in late July. Now, as renewed protests take place in New Delhi, Indian officials are once again targeting Bitchat — a sign that they see offline messaging tools as a lifeline for individuals when regimes restrict internet access.

China’s AI-Powered Surveillance Network Across Tibet, the Uyghur Region, Inner Mongolia, and Hong Kong

C4ADS, a nonprofit reporting on global conflict and security issues, released a report examining China’s expanding surveillance systems across Tibet, the Uyghur region, Inner Mongolia, and Hong Kong, calling them a “deliberate architecture of repression.” Researchers found that police agencies purchase and deploy AI-enabled biometric tools, including facial recognition, DNA collection, fingerprint systems, and predictive policing technology. China’s own domestic industry increasingly supplies these systems, making them harder to disrupt through sanctions and export controls. According to the report, the Chinese Communist Party’s (CCP’s) “surveillance efforts are not isolated but part of a system of control that represses ethnic minorities and political dissidents at scale.”

In context: The report focuses on regions where the CCP subjects ethnic minorities and pro-democracy groups to particularly intense surveillance and repression. In Tibet, the CCP has imposed decades of political control while suppressing dissent and eroding Tibetan identity. In the Uyghur region, Uyghurs and other Turkic Muslims face mass arbitrary detention and forced labor. In Inner Mongolia, officials restrict Mongolian language and culture as part of assimilation efforts. And in Hong Kong, activists and journalists are prosecuted for peaceful political activity. AI-powered surveillance could make this repression easier to scale.

Malawi Considers Expanding Its AI-Powered Surveillance in Blantyre

In a recent meeting with city officials, Malawi’s Communications Regulatory Authority proposed installing AI-powered surveillance cameras across Blantyre, Malawi’s second-most populous city. The plan would use smart cameras, license plate recognition, pedestrian and traffic monitoring, and a central command center to give police and other agencies real-time surveillance across parts of the city. Officials say the system is intended to reduce theft and other criminal activities, but in a hybrid authoritarian system that arrests opposition figures and activists and intimidates independent media, an expansion of state surveillance raises serious concerns about how these tools could be used against dissent.

RECOMMENDED CONTENT

The CCP’s Deepening Footprint in Latin America and the Stakes for Democracy in the Region

HRF has released a new report examining the CCP’s growing influence across Latin America. The report finds that Beijing uses economic leverage, diplomatic strategies, state-backed propaganda, pressure to isolate Taiwan, and, perhaps most relevant to this newsletter, technology and security partnerships. Chinese companies also supply surveillance systems, facial recognition, and other monitoring technologies to governments across Latin America, including the authoritarian regime in Venezuela. The report warns that these tools can both strengthen governments’ ability to monitor citizens and track dissidents and deepen dependence on Chinese technology in the region. Read it here.

THE LATEST IN AI FOR FREEDOM

Mistral Launches Open-Weight Model “Le Chonk” 

Mistral, one of Europe’s leading AI labs, has launched a public preview of its largest and most capable model yet, Mistral Large 4, also known as “Le Chonk.” Mistral says Le Chonk is competitive with the strongest open-weight models in the world and significantly outperforms every open-weight model developed in the United States or Europe. Open-weight models are ones that can be downloaded and run on private hardware, giving users control over their data and intelligence. Le Chonk is currently only accessible through Mistral, but Mistral says it will release the weights by the end of this month. If it delivers on these promises, users could gain access to one of the world’s most powerful AI models on infrastructure they control, or through privacy-preserving services that host open-weight models.

In context: Chinese companies have dominated in open-weight AI, even though researchers have documented CCP-aligned censorship in some Chinese models. Le Chonk could become one of the strongest alternatives developed in a democracy.

NVIDIA-Backed Reflection AI Launches Open-Weight Model Beam

Reflection, an American AI company backed by NVIDIA, has announced Beam, an open-weight model built for coding, reasoning, and AI agents. Beam is in early access, and Reflection says it will release the weights later this month. Reflection says Beam is competitive with GLM-5.2 and approaching Qwen 3.8-Max, two leading Chinese models, on coding and agentic tasks, while using significantly less computing power — potentially making it cheaper to run. Beam gives users another powerful open-weight option developed in a democracy that could reduce dependence on models developed under China’s authoritarian regime.

Underdog Brings On-Device AI to Personal Hardware

US-based AI lab Conway has launched Underdog, a local AI assistant that runs directly on consumer hardware. It can handle chat and dictation, connect to email and calendars, and take meeting notes — all powered by the team’s open-weight model, Underdog 27B, which is fine-tuned from Alibaba’s Qwen 3.8 27B. The team says this model performs comparably to leading AI models from six months ago on some benchmarks. Underdog is still invite-only and in beta. It runs on Macs with Apple silicon and in a Windows beta, and the team plans versions for Linux, iPhone, and Android. Underdog’s creator, Sigil Wen, explained the vision for the product: “AI must be a liberating equalizer for humanity. Capable, reliable assistance should reach everyone with a phone or computer, giving them greater command of their time, information, and choices.” 

Important: While this model is open weight, the remaining Underdog software is closed source, so outside researchers cannot independently verify all of its privacy claims. Users handling sensitive information should remain cautious.

Referandom Makes Turkish Politics Easier to Follow

Kadir Yağız Çetin, a civic technologist and the founder and CEO of CDR, and his team have rebuilt Referandom from the ground up with AI. The platform makes the decisions of Turkey’s parliament and municipal councils easier to follow. Users can read plain-language summaries of parliamentary motions and bills in Turkish, swipe right to agree and left to disagree, and see how their views compare with how parliament actually voted. On the backend, CDR’s AI-powered public policy platform, Policyful, organizes legislation by topic, generates summaries, and analyzes parliamentary inquiries, questions, and bills. Referandom gives people a clearer view of how political decisions are made in a regime that HRF classifies as fully authoritarian. 

Why this matters: In a conversation with HRF, Kadir Yağız Çetin explained that “under the new presidential system, most people in Turkey no longer understand what parliament is for. Many don’t know that making laws is still parliament’s job, let alone what it decides every week. Requests for parliamentary inquiries are voted down in a few minutes, and the record is buried in long transcripts … [Referandom] puts parliament’s decisions in people’s pockets, in plain language, and asks them one simple question: Do you agree? Being able to see is where accountability starts.”

Dozens of Open-Weight Jev Models Have Emerged

Since the launch of Jev, a closed-source AI model specialized in making decisions, dozens of open-weight alternatives have emerged. Decision models like these score a set of options and return the most likely choice, rather than generating long responses like traditional AI models. This makes them much faster, cheaper, and less computationally demanding for tasks where software needs to make a quick decision about what to do next. New open-weight models — including Laya, Kev, Clef, and Strands Decider — could deliver useful intelligence that runs privately and at no extra cost on cheaper hardware, instead of expensive data-center machines. Some of these alternatives already perform competitively with Jev on certain benchmarks. Users can try running them locally with tools like Ollama and Unsloth.

Why this matters: These models could be built directly into freedom-enhancing tech to make fast decisions — like checking whether a link looks suspicious, deciding whether sensitive data should stay local, prioritizing urgent reports, or choosing the safest action for an AI agent — without sending that data to an outside AI provider.

RECOMMENDED CONTENT

AI Hack for Freedom III: Harnessing the Power of AI to Advance Freedom

The National Endowment for Democracy posted a recap video of our co-sponsored AI Hack for Freedom III, a hackathon that paired 10 human rights activists with developers to build tools for freedom. Watch it to hear from the activists and developers on the ground who took a technology that is so often weaponized by authoritarian regimes and used it to build solutions that help activists organize, protect themselves, and challenge those in power.