How can AI be used to repress?
Mass surveillance
Regimes can use facial recognition cameras, biometric databases, and online activity to monitor their populations.
Censorship and information control
AI can scan massive amounts of online content, flag criticism of regimes, remove posts, and block dissent faster than human censors ever could.
Propaganda and manipulation
AI can generate pro-regime content, coordinate bot campaigns, flood platforms with fake comments, and create deepfakes targeting dissidents — all at low cost and massive scale.
Predictive policing
Regimes can use AI tools to predict who may be considered a “threat” or potential “criminal,” including individuals who criticize authoritarian governments or participate in pro-democracy movements.
How can AI be used to liberate?
Capacity building
AI helps dissidents and small nonprofits conduct research, translate content, draft campaigns and communications, summarize reports, and increase their reach and impact. When using privacy-protecting AI tools, they can also work more safely with sensitive information.
Digital tool creation
Activists with limited resources or technical experience can now use AI and plain-language prompts to quickly build websites, applications, and other digital tools that support their movements.
Personal AI agents
AI agents work like personal assistants capable of helping activists execute tasks and work more efficiently. When they are open-source, activists can control how agents operate and connect them to private models that help keep sensitive information secure.
Freedom tech acceleration
Advanced software developers can now combine their coding and digital security expertise with AI to build more sophisticated, privacy-protecting tools for activists living under repression.
Program Initiatives
AI for Individual Rights Toolkit
Avoid Giving Centralized AI Providers Sensitive Information
Many AI services operate through centralized infrastructure, meaning users send prompts and data to company-managed servers. Examples include ChatGPT, Gemini, Claude, or any model accessed through a third party provider.
This design can introduce privacy risks. Depending on the provider, user inputs may be stored, logged, or used to improve services and models. Providers may also be subject to legal requests by authoritarian regimes or government regulations in the jurisdictions where they operate. For dissidents, activists, and nonprofits handling sensitive information, sharing confidential details through centralized AI services can create unnecessary risk.
Suggestion: A good rule of thumb when using centralized platforms like ChatGPT is: “Don’t type any queries that you don’t want to show up in the newspaper.”
Use Permissionless Proxies
A proxy is an intermediary that sits between users and the AI models. Instead of sending a question directly to a model provider, users send it to a separate service (proxy). The proxy mixes these questions with questions from other users, then sends them all to the AI company. This makes it difficult for companies to trace any single query back to an individual user.
Some services also accept Bitcoin payments, which can reduce reliance on identity-linked payment methods (such as credit cards) and provide an additional layer of privacy.
Recommended tools:
PayPerQ: A proxy service that accepts bitcoin and does not require an account. It grants access to leading closed-source models like ChatGPT, Gemini, and Claude without revealing user information.
Routstr: Another proxy service that accepts bitcoin. Routstr is also built on a censorship-resistant protocol called Nostr. This means that no government or company can shut down users’ or providers’ access to the platform.
Use Encrypted AI Services
Some AI services encrypt users’ chats with AI from end to end so that only the user can read them. Users interact with these services like any other AI chatbot with the benefit that their messages stay private, even from the provider.
To do this, they use open-weight models, which are AI systems whose inner workings are publicly available and can run in secure, controlled environments without sending user data to a central company server. This allows users to engage with AI tools and keep interactions private.
Recommended tools:
Maple AI: An open-source, end-to-end encrypted AI chat interface that enables private access to top open AI models. Users can even create completely anonymous accounts if they pay in bitcoin.
Tinfoil: An encrypted AI assistant providing private AI interactions with top open models.
Run AI Models Locally on Personal Devices
Open-weight AI models can run directly on personal devices without requiring a connection to external servers or, once downloaded, the internet. Users can download an AI model and interact with it locally, meaning their prompts and data remain on their device rather than being sent to a third party. Running a model locally on a computer can significantly improve privacy and reduce reliance on centralized infrastructure.
The tradeoff is performance: personal devices aren’t able to run models that are as intelligent as top-tier ChatGPT or Claude models. But for many tasks, smaller, local models are capable enough. And for privacy, nothing beats running a model locally on a computer that is not connected to the internet.
Plus, the performance gap is closing quickly. Open-weight models continue to improve and are becoming increasingly capable of running efficiently on laptops, phones, and other personal devices.
Recommended tools:
Ollama: An application that lets users run open-weight AI models locally on their own computer. Simply install the app on a computer to download and interact with different AI models privately and offline.
LMStudio: Another application that lets users download, run, and chat with open-weight AI models locally on their personal devices.
Use AI To Vibe Code
Vibe coding is an AI-assisted approach to building software using natural language. Instead of writing every line of code manually, users describe what they want and AI helps generate code and applications. This can lower the barrier to creating digital tools.
For activists, nonprofits, and nontechnical users, AI can make it easier to build websites, apps, workflows, and other tools that support their work. While some technical understanding is still often helpful, AI increasingly allows people to create and prototype ideas that previously required specialized software development skills.
Recommended tools:
Replit: A simple vibe-coding platform that quickly builds basic websites and mobile apps with a plain language prompt.
Shakespeare: A vibe-coding platform built on Nostr, a censorship-resistant protocol, so that users can build sites no one can shut down. Users can choose Maple AI or PayPerQ as the AI model powering the site-building process to enhance privacy.
OpenCode: An AI coding agent that lets users develop advanced software like full digital applications, tools, and infrastructure (opposed to just simple static websites). Because it is open-source, users can choose which AI model powers development, including privacy-focused options like Maple AI or local models so their interactions remain private.
Important Consideration:
Some AI agents are closed-source, meaning they are controlled by companies. While these tools can have powerful vibe-coding capabilities, they typically lock users into one model provider and limit the user’s sovereignty. For example, users cannot connect their agent to a local (and therefore private) model or create privacy-preserving setups using services like Maple. For activists, dissidents, and organizations handling sensitive information, this can create privacy risks. Prompts, files, code, and other data from the computer or workspace where the agent runs may pass through third-party systems and be subject to storage or retention policies.
For sensitive work, users should use tools like OpenCode and connect their agent to a local or private model.
Experiment With Open-Source Personal Agents
Open-source personal agents are powerful AI assistants that users can control and customize. Unlike ordinary chatbots that simply answer one question at a time, these agents are systems that can remember context, learn user preferences, and execute complex tasks on the user’s behalf. Give them instructions using plain, everyday language, and they deploy these requests — analyzing documents, writing code, creating and editing files, building applications, or automating parts of a workflow.
These agents connect to the user’s choice of AI models, including local models for greater privacy and control. Users can then interact with them through apps like Telegram, WhatsApp, Slack, or Discord. Used well, they can learn user’s preferences, interests, and workflows and function as 24/7 assistants tailored to the user’s work and needs.
Recommended tools:
OpenClaw: An open-source AI personal agent. Users can set it up through a separate personal device and watch it transform into a fully customizable AI assistant that works on their behalf. Its invention was a breakthrough enabling individuals to control powerful AI assistants without relying on closed, centralized platforms.
Hermes: An open-source agent that features a continuous learning loop and automatically documents its projects. This means it learns and improves from past experiences to become a better assistant.
Important Consideration:
These tools are still experimental and should be used carefully. Agents are not privacy tools. Users should not give them any sensitive information, credentials, private files, or accounts.
For more safety, use a dedicated device (like an old laptop or Mac Mini) that does not contain any sensitive information. Limit agents’ access and treat them like colleagues, giving them only the permissions needed for a specific task (not full access to an entire system).
If agents are connected to centralized services like Telegram, OpenAI, or Anthropic, users’ plaintext commands, queries, and files may be visible to those platforms and exposed to corporate or state oversight.
Set Up an Agent Without a Second Device:
Clawi.ai is a service that lets users set up AI agents (like OpenClaw and Hermes) through the cloud instead of a personal device. This means that the AI agents run in an isolated environment rather than on a personal machine. It reduces the risk of an agent accessing sensitive local files. This can be a less risky alternative to running an agent on a personal machine. However, running agents in the cloud also poses risks of potential exposure of user information if the cloud environment is misconfigured or compromised. Do not give it any sensitive information.
Additional Privacy Tips
For everyday browsing, consider using Brave Browser. It is designed with privacy in mind and blocks many ads and trackers by default to help reduce the amount of data websites collect.
Also consider a virtual private network (VPN). It is a tool that helps protect privacy online by masking users’ Internet Protocol (IP) address, a unique identifier used by devices communicating over the internet. VPNs encrypt users’ internet traffic, making it more difficult for external parties, such as public Wi-Fi networks or internet service providers, to monitor online activity.
Recommended VPNs:
Proton VPN: A user-friendly VPN that offers a no-logs policy, meaning it does not track or store users’ browsing activity.
Obscura: A VPN service that uses a special technique (called a 2-party relay) to ensure that no single party can see both who the user is and what websites they access. Users can pay with bitcoin to avoid payment methods like credit cards that can be tied to their identities.
Mullvad: A VPN service that does not require an email address to sign up. Users can create anonymous accounts using randomly generated account numbers. Mullvad also accepts bitcoin for increased privacy.
For more sensitive activity, consider using Tor Browser. It is a privacy network that routes internet traffic through multiple encrypted layers to make online activity more difficult to trace or link to a specific user.
For messaging and calls, consider using Signal. It uses end-to-end encryption by default, meaning only the sender and the intended recipient can access the contents of a conversation.
