Ethical Tech

Ditching Perplexity & Comet: A Guide to Ethical AI Alternatives

A scored guide to ethical AI alternatives after leaving Perplexity and Comet. 14 research and in-browser tools ranked by ethics, capability, and real-world use.

11 Mar 2026 · 20 min read · By Sophie Kazandjian

Scored alternatives to Perplexity and Comet, written after I walked away from both. The trade-offs are real and I have listed them.

Over the past year, I have been steadily migrating my tech stack away from US-owned, billionaire-backed platforms. I closed my Amazon accounts. I moved from Google to Proton Mail. I left Facebook and Instagram for Mastodon and Pixelfed. And I started looking very hard at the AI tools I was relying on every day.

The more I looked, the less comfortable I became. AI-driven research assistants and in-browser tools have changed how we work, no question. But this progress comes with serious ethical baggage: data harvested for model training without meaningful consent, infrastructure subject to US surveillance laws, opaque ownership structures, and an environmental footprint that most providers would rather you did not think about.

I ditched ChatGPT first. OpenAI's trajectory towards aggressive commercialisation, its cosying up to an increasingly authoritarian and belligerent US administration, and its approach to training data made it an easy decision. I moved to Claude (Anthropic) as my primary AI, and it remains the most capable tool I have used for complex, multi-step work. But Claude is US-based, and I am not comfortable having all my eggs in one American basket.

Then came Perplexity. I had been using it as a research AI and its Comet browser for in-page analysis and summaries. Both are useful tools. I even wrote about how impressed I was with Comet: I published guides on its practical use cases and on using it specifically for Squarespace design work. I meant every word at the time. Comet is slick, capable, and well-suited to the kind of work I do.

But I could not keep ignoring what sits behind it. Perplexity has Jeff Bezos as an investor, operates under US jurisdiction, faces ongoing publisher lawsuits over copyright infringement, and its data practices do not meet the standard I hold myself and my clients to. No matter how good the tool is, if the ethics do not align, I need to walk away. You can read my original Comet articles here: In-Browser Comet AI Use Cases and Comet AI Browser for Squarespace Designers. I stand by the practical assessments in those pieces, but I can no longer recommend the platform itself. This guide is the result of finding what comes next.

So I went looking. This guide maps every viable ethical alternative I could find to Perplexity (research AI) and Comet (in-browser AI assistant), evaluated against four questions: where is your data going, who owns the company, what are the models trained on, and can you actually get your work done with this tool?

I also tested these tools against each other. I asked GreenPT and Mistral Le Chat to produce their own versions of this research, then cross-referenced the findings. The results were illuminating, and not always flattering for the ethical contenders. More on that later.

My setup now

My current setup: Claude (Anthropic) as primary AI for complex work, with full awareness of the US jurisdiction trade-off. Mistral Le Chat (France) as my ethical backup and go-to for EU-based research. Brave Leo for passive in-browser page analysis, already built into the Brave browser I was using anyway. Claude in Chrome for agentic tasks where I need Comet-style live page interaction, also running inside Brave. And GreenPT as an experiment I am watching with interest, though with some reservations about accuracy.

Between Brave Leo and Claude in Chrome, I have replaced everything I was using Comet for, all inside the same browser.

I wrote it up for anyone making similar choices, whether for personal use or, like me, for advising clients on ethical technology transitions.

The bigger picture

The choices we make about AI tools are not just personal preferences. They are decisions about where our data lives, which legal jurisdictions govern it, whose profits we support, and what kind of digital future we are collectively building. For small, values-led businesses especially, these decisions send a signal to clients and collaborators about what you stand for.

In 2026, European AI has grown up. Choosing ethical alternatives no longer means accepting poor tools. The trade-offs exist, and I have listed them. But the gap between ethical and capable has narrowed to the point where you can make principled choices without handicapping your workflow.

Part 1 covers research AI alternatives. Part 2 covers in-browser AI. I scored each tool against these criteria:

• Jurisdictional sovereignty: preference for EU/Swiss-based providers subject to GDPR and the EU AI Act, avoiding US CLOUD Act exposure
• Independent or open-source ownership: avoidance of Big Tech ownership, favouring community-driven or sovereign models
• Transparent data practices: zero retention architectures, opt-in policies, end-to-end encryption
• Environmental responsibility: renewable energy, carbon transparency, sustainable hosting
• Practical capability: because an ethical tool you cannot use is not an alternative at all

How I scored these tools

Each tool is scored on two axes using a 5-star system. I have tried to be accurate rather than generous. Marketing claims are not the same as actual practice, and I have prioritised what I could verify over what companies say about themselves.

Ethical credentials covers jurisdiction (EU/Swiss vs US vs other; subject to GDPR, EU AI Act, or US CLOUD Act?), ownership (independent, open-source, or tied to Big Tech?), data practices (training on user data? encryption? zero-retention?), transparency (open-source models? auditable infrastructure?), and environmental responsibility (renewable energy? carbon transparency?).

Power and efficacy covers research depth (multi-source synthesis, citation quality), model quality (reasoning, accuracy, multilingual support), features (web search, document analysis, code generation), and usability (interface, speed, accessibility for non-technical users).

Part 1: Research AI alternatives to Perplexity

Perplexity does one thing very well. You ask a question, it searches the web, synthesises multiple sources, and gives you a cited answer. Fast, clean, useful. The tools below aim to do the same thing without the ethical baggage. Some get close. Some exceed it in specific areas. None are perfect.

Tool by tool

Mistral Le Chat

Mistral Le Chat is the strongest EU-based research assistant I have tested. Headquartered in France and regulated under GDPR and the EU AI Act, its open-source models (including Mistral Large for Le Chat) are fully auditable. Deep Research mode produces structured, citation-rich reports comparable to human researcher output, with voice input, Projects for organising ongoing work, and multilingual reasoning across languages including French, Spanish, and Japanese.

The critical caveat: Mistral uses Microsoft Azure and Google Cloud as sub-processors, and its data training policy is opt-out rather than opt-in. This means data could pass through US-owned infrastructure despite Mistral itself being French. For clients with the strictest sovereignty requirements, this is important. Mistral also discloses environmental impact at 1.14g CO2 per query, which is transparent but not trivial.

GreenPT

GreenPT has arguably the strongest ethical credentials on this list. Dutch-headquartered, running entirely on renewable energy in EU data centres (Scaleway, France), with self-hosted open-source models and a true zero data leakage architecture. For values-led businesses, the environmental positioning is hard to argue with: real-time CO2 tracking per query, no external API calls, and transparent sustainability reporting.

The capability gap is real, though. GreenPT runs Mistral Small (24B parameters) and GPT-OSS 120B, much smaller than frontier models. In direct testing, GreenPT produced confident but incorrect factual claims about competing tools, including misidentifying the AI models used by other platforms. This pattern is consistent with established research on the relationship between model size and hallucination rates. Independent accuracy benchmarks for GreenPT are not yet available due to the platform's early stage (domain registered February 2025). The ethical credentials hold up. The accuracy gap is real. For general conversation and drafting, this may be acceptable. For research work where accuracy is the point, outputs should be verified.

CamoCopy

CamoCopy operates entirely within EU infrastructure, with encrypted chats and a firm no-training-on-user-data policy. Its integrated search engine provides anonymous, cited web answers. The platform is powered by open-source models and offers both chat and search functionality in one interface. Their privacy analysis of major AI platforms is worth reading, particularly their assessment of Mistral's sub-processor dependencies.

The trade-offs are slower response times and a less polished interface compared to commercial competitors. Research depth does not match Mistral or Claude for complex multi-source synthesis. If you are a European business handling sensitive data and GDPR compliance is more pressing for you than raw power, this is worth trying.

Okara AI

Okara takes an unusually strong approach to encryption: client-side key generation means even Okara's own systems cannot read chat content in secure mode. It offers 20+ open-source models switchable mid-conversation, integrated search across web, Reddit, X, and YouTube, and workspace collaboration for teams. At $12.50/month for Pro, it is well-priced for what you get.

The Singapore jurisdiction is neither EU nor US, which means it avoids CLOUD Act exposure but does not benefit from GDPR enforcement. Founded in 2025, it remains a young platform. Reviews note it can feel technical for casual users, and it lacks access to frontier proprietary models. If you are a privacy-focused professional comfortable with a learning curve, it could work well.

Duck.ai

DuckDuckGo's Duck.ai acts as an anonymous privacy proxy to multiple AI models including Claude, GPT, Llama, and Mistral. No account required, no data retained, no training on user conversations. The proxy architecture strips identifying information before queries reach model providers.

US-based, which limits its sovereignty credentials. And because it proxies to other models rather than running its own, it offers no deep research or multi-step synthesis. Think of it as a privacy-respecting gateway for quick lookups rather than a full Perplexity replacement.

Claude (Anthropic)

Full disclosure: Claude is my primary AI tool. Its deep research mode, web search, and complex reasoning capabilities are the strongest available for multi-step analysis. Anthropic's Constitutional AI framework aims to align outputs with ethical principles. Projects and memory features support ongoing work.

US jurisdiction and CLOUD Act exposure are the clear ethical limitations. Anthropic's privacy practices are more transparent than some US competitors, but data is processed on US infrastructure. For this guide, I treat Claude as the capability benchmark and map the ethical alternatives around it.

Part 2: In-browser AI alternatives to Perplexity Comet

Comet's appeal is simple: an AI assistant living inside your browser that can analyse the page you are looking at, summarise articles, help with code, and answer questions in context. I wrote about this in detail in my reviews of Comet's practical use cases and Comet for Squarespace designers. The tool itself is good. Really good. But I am no longer willing to hand my browsing data to a Bezos-backed US company. These alternatives aim to deliver the same experience without the ethical compromise.

Tool by tool

Brave Leo

Already built into the Brave browser, Leo requires no additional installation or account. It summarises webpages, analyses PDFs and documents, translates content, and provides contextual help via a sidebar. All models run in Brave's own secure environment with zero data retention. The Bring Your Own Model feature lets advanced users connect local models via Ollama.

There is an important distinction to make here: Leo reads and analyses the page you are on, but it cannot interact with it. It will not click buttons, fill forms, or navigate between pages on your behalf the way Comet does. If you used Comet mainly for summarisation and research, Leo covers that well. If you relied on Comet for agentic tasks (having the AI actually do things on webpages for you), Leo does not replace that part.

Privacy-wise, Brave Leo holds up well despite being US-based: no server-side logs, no training on conversations, and the recent migration to AWS Bedrock eliminated the previous 30-day data retention for Claude models. You can compare the available AI models in Leo here.

Claude in Chrome

There is one more option worth covering, and it comes from the AI I already use every day. Claude in Chrome is Anthropic's browser extension that turns Claude into an active agent inside your browser. Unlike Brave Leo, which reads and analyses pages passively, Claude in Chrome can click buttons, fill forms, navigate between pages, and complete multi-step tasks on your behalf. It is the closest thing I have found to a direct replacement for what Comet actually does, and it works inside the Brave browser despite some reviews suggesting otherwise.

You need any paid Claude plan to use it. Install the extension, sign in, open the sidebar, and describe what you need. You can also record workflows and replay them later, which is useful for repetitive tasks. It pairs with Cowork to turn web research into finished documents, spreadsheets, and reports without copy-pasting between tabs.

Two things to know before you get excited.

First, it is still in beta. Anthropic launched it as a research preview with 1,000 Max plan users initially and has since expanded to all paid plans, but this is not a finished product. It can be slow (the screenshot-analyse-act cycle takes time), it occasionally struggles with complex navigation, and Pro plan users are limited to the Haiku model, which is noticeably less capable than the Sonnet or Opus models available on higher-tier plans.

Second, and more importantly: agentic browsing carries real security risks that passive tools like Brave Leo do not. When an AI agent can see your pages and take actions in your browser, it becomes a target for prompt injection attacks, where malicious code hidden on a website tricks the agent into doing something you did not ask for. Anthropic is unusually open about this. Their own testing found a prompt injection success rate of 23.6%, which their safety mitigations reduced to 11.2%. On browser-specific attacks (hidden form fields, malicious URL text, injected tab titles), targeted defences brought the success rate down to 0% on their test set. They have also blocked Claude from accessing financial services, adult content, and pirated content by default, and the extension asks for your permission before taking high-risk actions.

That 11.2% figure should give you pause. It means roughly one in nine attack attempts could still succeed. Anthropic is working to bring that closer to zero, and the beta is partly designed to surface new attack types in real-world browsing. But if you are handling sensitive client data, treat Claude in Chrome as a supervised tool rather than something you leave running unattended on unfamiliar sites. Start with trusted websites. Review what it is doing before confirming actions. And be aware that this category of tool, from any provider, is still maturing.

For me, the combination of Brave Leo (for quick passive page analysis) and Claude in Chrome (for active tasks when I need Comet-style interaction) covers everything I was using Comet for, all inside the same browser.

BrowserOS

The most ethically pure browser option. Fully open-source under AGPL-3.0, all AI processing happens locally on your machine, and data never leaves your device. Compatible with Chrome extensions and supports bring-your-own-keys or local models via Ollama.

The downsides: maintained by a small team, security updates may lag behind Google's Chromium patches, and the experience is less polished than commercial browsers. Suitable for technically comfortable users who prioritise sovereignty above convenience.

So what should you actually use?

Not everyone has the same tolerance for trade-offs. A sole trader handling sensitive client data has different needs from a creative freelancer who just wants to search ethically. Here is how I would advise different situations.

What I use (and why)

Primary research AI: Mistral Le Chat (France) for research with citations, Claude (Anthropic) for complex multi-step analysis. Between them, I get EU-based research for most tasks and US-based capability for the hardest problems.

Primary in-browser AI: Brave Leo for passive page analysis (zero data retention, multiple model choices, no setup needed) and Claude in Chrome for agentic tasks when I need Comet-style live page interaction. Both run inside the Brave browser.

What I recommend to clients

Tier 1, maximum sovereignty. CamoCopy for research, BrowserOS for browsing. Pure EU/open-source, zero US exposure. BrowserOS handles both passive analysis and agentic tasks locally. GreenPT has the strongest environmental credentials but accuracy limitations mean outputs should be fact-checked. Capability gaps are real at this tier.

Tier 2, best balance. Mistral Le Chat for research, Brave Leo for passive browsing AI. Strong ethical credentials with minor compromises (Mistral's Azure sub-processors, Brave's US jurisdiction) but excellent capability and usability. If clients need agentic features (form-filling, page navigation), Claude in Chrome works inside Brave but adds US jurisdiction exposure and is still in beta. Okara AI is a good supplementary option for privacy-sensitive multi-model work.

Tier 3, maximum capability. Claude (Anthropic) for research, Claude in Chrome for agentic browsing. The full Claude ecosystem gives you the strongest research and the closest Comet replacement in one subscription. TextCortex for multilingual content work within EU infrastructure. Accept the sovereignty trade-off in exchange for the best available tools.

The complications

Mistral's sub-processors

Mistral uses Microsoft Azure and Google Cloud as sub-processors, and its data training policy is opt-out rather than opt-in. This means data could pass through US-owned infrastructure even though Mistral is French. CamoCopy's analysis of Mistral's privacy position provides useful detail on this. CamoCopy and GreenPT avoid the issue entirely by self-hosting open-source models on EU-only infrastructure.

GreenPT's accuracy

GreenPT disappointed me, because I want it to succeed. It runs smaller open-source models (Mistral Small 24B and GPT-OSS 120B) that are inherently more prone to factual errors on niche topics. When I asked GreenPT to research alternatives to Perplexity, it confidently misidentified the AI models used by Brave Leo and Okara AI, and recommended several tools that appear to be hallucinated or too obscure to verify. That pattern aligns with established research on the relationship between model size and accuracy. Independent benchmarks for GreenPT are not yet available because the platform is so new (domain registered February 2025). I will keep testing it, but for now, fact-check everything it tells you.

Agentic browsing risks

If you used Comet for its ability to autonomously navigate websites, fill forms, and complete multi-step tasks, Claude in Chrome is the closest ethical replacement and works inside Brave. But all agentic browsers, including Comet, carry prompt injection risks that passive tools do not. Anthropic has reduced attack success rates and built in permission controls and site-blocking safeguards, but the technology is still in beta across the industry. Use agentic features on trusted sites, supervise what the agent is doing, and keep sensitive workflows separate until these tools mature. BrowserOS remains the most privacy-pure option for agentic tasks, processing everything locally, but requires technical confidence.

Brave Leo's US jurisdiction

Despite zero-retention policies and local processing options, Brave's US base means potential CLOUD Act exposure. The privacy architecture goes a long way to mitigate this, but for clients requiring absolute EU sovereignty, BrowserOS or a European browser with manual AI integration may be preferable.

Self-hosted options

BrowserOS and Web AI Browser offer full local processing and data control but require users to manage security updates and maintenance. These are not suitable for non-technical small business clients without IT support.

What comes next

No single AI tool ticks every box. That is the truth, and pretending otherwise would not serve anyone reading this. My approach is to combine tools: Mistral Le Chat for EU-based research with citations, Claude for the complex multi-step analysis that nothing else handles as well (accepting the US jurisdiction trade-off with open eyes), and Brave Leo plus Claude in Chrome for in-browser work, both already running inside Brave.

I am also keeping an eye on GreenPT. The environmental positioning is exactly where I want the industry to go, and if the accuracy improves as the models mature, it could become a serious contender. For now, I treat its outputs as a starting point that needs verification rather than a finished answer.

The broader picture is important too. A year ago, recommending European AI alternatives meant recommending compromise. That is no longer true. Mistral's Deep Research mode can hold its own against Perplexity. Between Brave Leo and Claude in Chrome, I have replaced everything I was using Comet for. The tools are here, they work, and choosing them is a small act of digital sovereignty.

If you run a values-led business and you are still using tools that do not align with what you stand for, the excuse that there are no alternatives no longer holds. I have named them, scored them, and left you to decide where your own line falls.

For further reading on European privacy-first tools beyond AI, I recommend Plausible Analytics' European tools list, WauwAI's privacy-first alternatives directory, and EuroBoxx's 2026 European AI comparison.

The tech we choose shapes the future we get, so choose deliberately.

FAQs

Why did you leave Perplexity if it works so well?

Perplexity has Jeff Bezos as an investor, operates under US jurisdiction, and faces ongoing copyright lawsuits. More importantly, its data practices don't meet the ethical standards I hold myself and my clients to. No matter how capable a tool is, if the values don't align, it isn't a sustainable choice for a values-led business.

Is Brave Leo really free?

Yes, Brave Leo is free and built directly into the Brave browser. You get access to multiple AI models at no cost. There's an optional premium tier for faster responses and additional model choices, but the core functionality requires no payment or account creation.

What's the difference between Mistral Le Chat and Claude?

Both are excellent research AIs. Claude (US-based) has slightly stronger reasoning for complex multi-step analysis. Mistral Le Chat (France-based) offers EU jurisdiction and GDPR compliance, making it better for European sovereignty requirements. Mistral also has Deep Research mode comparable to Claude's capabilities.

Does GreenPT's accuracy problem make it unusable?

Not entirely, but it requires caution. GreenPT runs smaller open-source models that are more prone to hallucinations on niche topics. It works well for general conversation and drafting, but for research where accuracy is the point, you should fact-check outputs. The environmental credentials are real, so it's worth monitoring as models improve.

Can I use these tools if I'm not in Europe?

Absolutely. The sovereignty benefits matter most if you're handling sensitive client data or have strict compliance requirements. If you're a freelancer anywhere in the world prioritising privacy and ethics over US jurisdiction exposure, these tools still offer real advantages. Okara AI (Singapore) is a good option outside both US and EU jurisdictions.

What about data privacy with Brave Leo if it's US-based?

Brave Leo has zero data retention. Conversations aren't logged on Brave's servers. The recent migration to AWS Bedrock removed the previous 30-day retention for Claude models. While US jurisdiction creates theoretical CLOUD Act exposure, the privacy architecture is strong compared to other US-based AI tools.

Is BrowserOS suitable for non-technical users?

Not really. BrowserOS requires local AI processing setup, manual security updates, and comfort with technical configuration. It's ideal for privacy-focused professionals or developers, but small business owners without IT support should consider Brave Leo for page analysis or Claude in Chrome for agentic tasks, both of which run inside the Brave browser.

How often do these tools change?

AI tools evolve rapidly. The landscape in March 2026 is already different from a year ago. I recommend checking the privacy policies and ownership structures of any tool you adopt, and revisiting your choices annually. Things change quickly, so staying informed is important.

Can I use multiple AI tools together?

Yes, and I recommend it. My own setup combines Mistral Le Chat (EU research), Claude (complex analysis), and Brave Leo and Claude in Chrome (in-browser work). This approach lets you get EU sovereignty for most tasks while accessing the best capability when you need it, rather than forcing a single tool to do everything.

What if a tool I'm using gets acquired or changes its privacy policy?

That's a real risk with any tech platform. The best defence is diversification. Don't rely on a single tool, and stay informed about ownership changes. Open-source options like BrowserOS and GreenPT offer more stability since they're not subject to acquisition, though they come with different trade-offs.

Written March 2026 by Sophie at Sophie's Bureau. Ethical assessments are based on publicly available information and my own direct testing. Things change quickly in AI. If something here has changed since publication, please let me know.

Back to the Journal