I Rated The Most Privacy- & Eco-Friendly AI Chatbots — The Results Surprised Me

Summary

The video reviews privacy-conscious and lower-impact alternatives to mainstream AI chatbots. The host rates each service on privacy and data use, openness and transparency, model variety, environmental impact, price and accessibility, and features and user experience.

Why the Host Is Concerned About AI

The video outlines several risks associated with AI systems:

• Poorly specified goals: Systems optimized for a narrow objective may produce harmful results if they fail to account for human context. The host cites examples such as an overly cautious spam filter deleting legitimate emails or an inflexible system denying someone access to a service.
• Bias and social impacts: Training data can carry biases, while automation may displace workers, enable fake news and deepfakes, and weaken human judgment when people rely on AI to make decisions.
• Environmental costs: The host discusses energy use for training and everyday chatbot use, as well as water for cooling, backup diesel generators, and the carbon footprint of hardware. She argues that routine AI use—“inference”—could become a major source of energy demand.
• Privacy risks: Prompts may reveal sensitive information, and AI services may collect or infer personal details. The host also raises concerns about models trained on data without people’s consent and about companies’ control over data storage and privacy rules.

The host says she still uses AI but tries to do so selectively—for example, choosing a regular web search when it can answer her question.

Chatbot Reviews

Scores are presented in this order: privacy, transparency, model variety, environmental impact, price and accessibility, features and user experience.

Service Scores Review highlights NanoGPT 5, 4, 5, 2, 4, 5 A cloud-based platform offering access to a wide range of models, including GPT-4, Claude, Gemini, Mistral, and DeepSeek. The host praises its anonymous use, optional registration, locally stored chats, pay-per-prompt pricing, crypto payment option, free model, and polished interface. Its cloud-based operation limits its environmental score. Jan AI 5, 4, 4, 5, 5, 4 Runs models locally, with optional encrypted cloud use, giving users control over their data. It is free and has a modern interface, though the host reports lag on her MacBook. Users need to download the models they want. Ollama 5, 5, 4, 5, 5, 3 A free, local, open-source option with support for models such as DeepSeek, Mistral, Llama, and Phi. The host praises its transparency and privacy, but gives its interface a lower score because it is command-line-based and may require using a terminal. Venice AI 4, 4, 4, Not stated, 5, 5 Browser-accessible and free, with a user-friendly design and access to models including DeepSeek. The host says conversations are handled through decentralized nodes and that its backend is based on open-source software, though the service is not fully open. DuckDuckGo AI Chat 4, 4, 3, 2, 5, 4 Offers access to models such as GPT and Claude through DuckDuckGo’s privacy layer, with no account required. The host notes that prompts still reach model providers, model availability may vary by region, and the cloud-based service has a lower environmental score. It includes search-integrated chat but, in the host’s review, lacks image and video search. Naga AI 4, 2, 4, 2, 4, 3 Aggregates models from providers such as OpenAI and Google and is accessible through Discord. The host says it adds measures such as IP masking but still sends prompts to the underlying providers. She notes limited information about how its proxy works and describes the interface as basic. LM Studio 5, 5, 3, 5, 3, 3 A local option that the host considers private and open-source, with reduced dependence on cloud services. She found its setup and downloads slow and more difficult than those of the other tools she tried. Its interface is functional but lacks extra features.

Honorable mentions: Private AI and Local AI are briefly recommended as additional options, but the video does not review or score them in detail.

Overall Takeaway

The host’s strongest personal discoveries were NanoGPT, for its breadth of models, and Ollama, which she enjoyed using locally. The review emphasizes that there is no single best choice for everyone: local tools favor privacy and lower cloud dependence, while cloud aggregators can offer greater model variety and convenience.

Main Speakers and Sources

• Elvi, the video’s host and reviewer.
• The host references computer scientist Stuart Russell on AI objectives and safeguards, MIT on energy use from AI inference, and author Cathy O’Neil on regulation and accountability.
• The host also credits Patreon supporter Mindas for sharing information that helped inspire the video.