Ethical Tech

Ethical AI Alternatives: I Tested Five European Assistants

I gave European AI assistants the same questions, ones where I already knew the answers. Here is which ones to use, and the problem I found with Claude's memory.

1 Nov 2026 · 6 min read · By Sophie Kazandjian

I gave European AI assistants the same questions, ones where I already knew the answers. Some did well. One made up a quote from my own website.

In March I wrote a guide to ethical AI tools. Eight months on, most of it is out of date. Mistral's assistant has a new name and new terms, Proton's Lumo has moved its servers to Germany, and a Dutch tool called GreenPT has become good enough to use every day.

This time I wanted more than reviews. On 1 November I gave the same questions to five European assistants, plus Claude, which I still pay for. Every question had an answer I could check.

Why look beyond Claude

Claude is the tool I use most. It rebuilt my website and writes code I could not write myself. It is also American. It runs in Amazon and Google data centres in the United States, and Anthropic is spending $50 billion on its own in Texas and New York. A small business has no way to keep its conversations in Europe. Chats on the Free, Pro and Max plans train future models unless you switch that off, and the switch arrived turned on.

This year also showed how much depends on one government. To its credit, Anthropic refused a Pentagon deal rather than drop its limits on mass surveillance and autonomous weapons, and was punished for it. Then in June the US Commerce Department ordered it to restrict two of its newest models, and they went offline for everyone for nearly three weeks. A small business in France had no say in any of it. That is the case for having a European option ready, even if you keep Claude.

What I asked

The questions were ordinary work. What does Sophie's Bureau offer, and what does it cost? Did Proton move its company to Berlin? It did not, and I worded the question that way to see who would go along with it. From an anonymised spreadsheet, how many people attended at least four sessions of a leadership programme? The answer is 83. I also asked each one to rewrite a paragraph to my own style rules, with a cliché and a hidden list of three left in to catch them out, and to find information about a private person.

Prices as listed on each company's site on 1 November 2026.

GreenPT did best for everyday work. On the cheapest plan it got almost everything right, including my prices, which Claude got wrong. It is noticeably slower than Claude, though never slow enough to get in the way. It runs in Paris on renewable power and shows how much energy each chat uses, which I have not seen anywhere else. Its one bad answer was about a private person: alongside the useful result it pulled in a court report, a people-search site and a school name. More sources is not better when the subject is a person.

Lumo in Max mode was the most accurate. It worked through the spreadsheet row by row for several minutes and came back with the right number, then gave a second figure in case recordings counted, which nothing else thought of. What you type is encrypted and nothing is kept. If you can wait, it is very good.

Mistral Vibe comes closest to Claude in what it can do. On the free plan it made small mistakes with complete confidence, such as calling a €49 product free. Since September it uses what you type, including documents you upload, to train Mistral's models unless you switch that off in the privacy settings. Do that before you paste in anything about a client.

Euria, from the Swiss host Infomaniak, has the cleanest setup here: Swiss data centres, nothing stored, and free. The free model reads web pages well, but it said 64 where the answer was 83, and its rewrite changed what a sentence meant. Apertus, the Swiss public model you can choose inside Euria and Lumo, agreed that Proton had moved to Berlin and quoted a passage from my website that does not exist. It is not ready for work.

CamoCopy could not search the web on its free plan, so it declined most of the questions. I tried its Deep Search mode twice, and both times it crashed and lost the conversation.

If a number matters, check it yourself. Four of the assistants got the count right and three got it wrong, each with the same confident tone.

The problem with Claude's memory

Claude was the most careful with people and the best editor. It also got my own business wrong. It quoted my prices from before October, then told me my homepage and FAQ page disagreed. They do not. Claude had remembered the old prices and trusted its memory over my live website. On another question it used an old page that is still online instead of the current one.

Memory is what makes an assistant feel useful, because it knows your clients and your prices. It also goes out of date without telling you. The longer you use one assistant, the more of your business it holds and the harder it is to leave.

My fix is dull. I keep one document of current facts outside any assistant, and delete the old version when something changes. I tell the assistant that the live website wins if the two disagree. For questions about the world rather than about me, I use a fresh chat with memory off.

Where the models come from

Euria and Lumo both run Chinese AI models, mainly Qwen, on their own European servers. Your data stays in Europe with a European company. The model itself was trained in China, and US government testing has found some Chinese models repeating Chinese state positions on political subjects. For drafting an email or summarising a report it is unlikely to show. It is still worth knowing what you are using.

What I use now

Claude for the hardest work and most drafting, with training switched off and no client names in it. Mistral's Vibe alongside it. GreenPT on its €4.50 plan, on trial for a month, for reading, summarising and checking. My automations are still on Zapier. I have recommended the European alternatives and not yet moved my own, and until I do, this sentence stays in.

Claude stays for now. The reason for keeping it is thinner than it was in May, and June showed me what "for now" can mean.

Questions

Is there a European alternative to Claude or ChatGPT?

Yes. For everyday reading, summarising and checking, GreenPT did best in my tests. For harder tasks, Mistral Vibe comes closest to Claude.

Which is the most private?

Infomaniak's Euria: Swiss, free, nothing stored, and no account needed. Proton Lumo is close, and encrypts what you type.

Do these tools train on what I type?

Euria and Lumo say they do not. Mistral Vibe does on its Free and Pro plans unless you switch it off, and so does Claude.

Why do you still pay for Claude?

It is the best tool I have for code and hard analysis, and I have not yet found a European one that does the same. I use it with training off and without client names.

Back to the Journal