Skip to main content

← AI ethics and policy

Why I Use Claude

by Sergueï Spetschinsky
October 2026

When I teach researchers to use AI for their work, I show them how to do it with Claude, the AI environment made by the American company Anthropic. I am often asked whether there are not fundamental concerns about using such tools, for research, for work or for personal use. This note answers that question as best I can, as things stand in autumn 2026.

The issues with Claude

People who have reservations about using Claude (or ChatGPT, or any big private provider) are probably right to have them. There are undoubtedly risks attached to feeding it any kind of sensitive information. Data guarantees are lacklustre, on several counts.

Where your data goes. On a personal paying plan, which is the plan most researchers will have unless their institution has a special agreement with Anthropic, your conversations go to Anthropic's servers, mainly in the United States, and no personal plan lets you keep them in Europe.1

What Anthropic says it keeps. By Anthropic's own account, you can switch off the use of your conversations for training, deleted conversations leave its systems within thirty days, and conversations flagged for a safety review are kept for up to two years and may still be used to improve its models, even if you have opted out. These are the company's own statements. No regulator or court has examined them, its security certifications are stated only for its commercial products, and its own policy allows it to keep data longer for legal reasons. Courts can also override deletion: in 2025, an American judge ordered OpenAI to keep the ChatGPT conversations its users had deleted.2

No data-processing agreement. No personal plan comes with a data-processing agreement, the contract that binds a service to handle personal data only on your instructions and under European rules. Anthropic offers one only with its organisational plans and its developer API.3

The GDPR. This is where the problem is most serious for research. No European regulator or court has ruled on Claude so far, so whether a given use complies is, strictly speaking, undetermined. But the GDPR puts the burden of proof on whoever is responsible for the data, usually the university or the researcher: they must be able to demonstrate that they comply. For other people's personal data, that means showing a legal basis, informing the people concerned, a contract that fixes the provider's role, and a lawful ground for sending the data out of Europe. On a personal plan, Anthropic handles what you type under its own privacy policy and for its own purposes, and no contract binds it to your instructions. A researcher who puts interviews, patient information or any other personal data about other people into a personal Claude plan therefore cannot demonstrate compliance, and failing to demonstrate it is in itself a breach of the GDPR. Doing it properly needs at least an organisational contract with a data-processing agreement, or a model that runs on your own machine.4

American law and American intelligence. Anthropic is an American company, so American law (the CLOUD Act) can compel it to hand over data in its possession to American authorities, wherever that data is stored.5 American intelligence services have further powers, outside any criminal procedure: under Section 702 of the American foreign-intelligence law, which lapsed in June 2026 but is reportedly still applied under existing authorisations, the government can direct American providers to hand over the communications of non-Americans abroad without an individual warrant, and much collection abroad happens under executive rules, without any court order. What these powers reach in practice is not public. Published documents indicate that American agencies sought economic intelligence on French and German companies. That researchers' data could become the target of scientific or industrial intelligence is a reasonable hypothesis, though I know of no case involving an AI provider. The same concerns apply to any foreign provider, and more sharply to Chinese ones, whose national law requires organisations to support the state's intelligence work.6

None of this is particular to Claude: most of us already entrust our data to Gmail, Outlook, OneDrive or iCloud, whose consumer versions raise largely the same questions.7 And the legal ground for sending personal data from Europe to the United States at all remains, to say the least, blurry.8

Anthropic and the American state. Anthropic works with the American military and intelligence services. Since November 2024 it has supplied Claude to defence and intelligence agencies through Palantir and Amazon, it has built "Claude Gov" models for national-security customers, and in July 2025 it signed a Pentagon agreement worth up to 200 million dollars. Press reports based on anonymous sources say Claude, running inside Palantir's targeting system, was used in American operations in Venezuela and in Iran; neither Anthropic nor the Pentagon has confirmed this. In February 2026, by its own account, Anthropic refused Pentagon terms that would have allowed the mass surveillance of Americans and fully autonomous weapons, and the administration then declared it a risk to the military's supply chain; the case is still before the courts.9

Your rights, and their limits. As a European, you keep your rights under the GDPR against Anthropic: you can ask to see your data, have it deleted, or object to its use. But you have no control over where your data lives, what American law can compel an American company to hand over, and once your data has been used to train a model, it cannot simply be taken back out.1

For those reasons alone, I always advise my students and clients not to put into Claude (or into any AI tool, be it American, Chinese or European, unless it runs locally) any sensitive information they would not want to fall into the hands of governments or large corporations: personal information, sensitive research, or data protected by privacy laws, such as patient information or interviews.

Why Claude

Yet, in spite of my scepticism, I use Claude intensively for my own work, and teach other people how to use it. The reason is simply pragmatic: for advanced knowledge work and research, the frontier models (such as Anthropic's Opus and Fable, or OpenAI's GPT-6 Astra) and the software that comes with them are currently the only realistic solution for most researchers.

The models are still better than their alternatives, although these are starting to catch up. The bigger difference today lies in what surrounds the model, the harness: Claude's desktop app and Claude Code, or OpenAI's ChatGPT desktop app and Codex. This is the software that lets the model read and write the files on your computer, produce documents in different formats, connect to your applications and control them, run tools and search the web, with an ease of use that nothing available in Europe matches yet.10

Then there is the price. As I read it, these companies are trying to win dominant positions and to get customers used to their environment, so they price their subscriptions far below their own per-request rates. For about 20 euros a month, a heavy user gets what would cost up to hundreds of euros at those rates; for about 100 or 200 euros a month, several thousand euros' worth of computation.11

What I would rather use

If they were up to the task, I would advise Europeans to use sovereign European systems. Unfortunately, in autumn 2026, nothing European comes close for advanced, complex work. Mistral, a French company, makes the strongest European models, and its assistant, Vibe (formerly Le Chat), is a credible tool for everyday use. But its models trail far behind on long, multi-step work, and it offers no equivalent of an agent that works on the files on your own computer for non-programmers. The other European offers are either much less capable, such as Apertus, the Swiss public model, or run American and Chinese models on European servers. They serve well for many everyday tasks, but not yet for the kind of systems I teach.12

Better still would be open-source models that run locally, on your own computer or on your institution's servers. For researchers, they have real advantages:

  • Your data never leaves your machine. Interviews, field notes and unpublished work can be processed without passing through any company, which makes consent and data protection far simpler.
  • They do not need to be bought or rented: most can be downloaded and adapted freely, under open licences.
  • They cost nothing to run (besides electricity) once you have the hardware.
  • They stay the same. A model pinned to a version can be rerun on the same material years later, whereas a commercial model can be swapped or withdrawn from the chat apps with a few weeks' notice.
  • No company can change the terms, raise the price or switch them off. Having one installed is like having a solar panel on your own roof.

In many ways, I believe they are the future of AI. But they are still in their infancy, and will need some years at least before they can be used for advanced work on an ordinary computer. An ordinary laptop runs only small models, and the open models closest to the American frontier need a machine costing many thousands of euros.13

Teaching the skill, not the tool

This complex situation is why I try as much as possible to teach deep AI skills rather than tools, techniques or tricks: an understanding of how AI works, and of how to think and work with it. These skills carry over to any tool, now and in the future, with some adaptation to each one's technical specifics, and they do not go out of date as models and tools evolve. What I teach is, in a sense, AI metaknowledge: how to learn, on your own, to use AI.

This is also what makes the work portable. Researchers learn to structure their data and their AI systems so that, down the road, any model, tool or provider can use them: plain text and Markdown files on their own computer, and skills, instructions and systems written in open formats, which other companies' tools, OpenAI's and Mistral's among them, already read, as do open-source ones.14 Learning to use Claude properly means learning to be able to leave Claude at any time.

My goal is to make the people I teach into builders of their own AI systems, able to work with any tool without needing me, or anyone else, to do so. Using AI for research means using it critically: understanding not only how to use a given instrument for a given task, but how these systems work, why they work the way they do, and how to think about them with a specific goal in mind. I believe this is the only approach that may allow a use of AI that enables not only an ethical use of AI, but also, an expert one.

Using Claude carries risks. But not learning to work with AI also carries risks. To give up or delay the ability to understand and use these tools is to lose much of what they may bring to scientific research. It also runs the risk for scientists to be left behind in knowledge of what may be one of the most powerful scientific discoveries ever to come to light.

Notes (14)

  1. Anthropic's privacy policy (effective 10 September 2026) states that data goes "to our servers in the US, or to other countries outside the EEA". For European users the contracting company is Anthropic Ireland, Limited, and transfers rest on adequacy decisions and standard contractual clauses. European hosting is offered only to developers, through Amazon's and Google's cloud platforms, not on personal plans. The same policy lists European users' rights: access, portability, deletion, correction, objection and restriction, with complaints to the national data-protection authority. Anthropic states that these rights have limits for data already used in training (Anthropic, training-data requests, 8 July 2026). Anthropic, Privacy Policy; Anthropic, Regional compliance; Anthropic, Data residency. ↩︎a ↩︎b

  2. Anthropic's statements are in How long do you store my data? (1 July 2026) and its Privacy Policy:

    • deleted chats leave its back-end within 30 days;
    • with training on, de-identified data may be kept for up to five years;
    • chats flagged for safety review are kept for up to two years, and their classification scores for seven;
    • data may be kept longer "if required by law, to resolve disputes, or to combat Usage Policy violations";
    • flagged content can be used to improve its models even after an opt-out.

    Its SOC 2, ISO 27001 and ISO 42001 certifications are stated for its commercial products, and their reports are available on request (Anthropic, certifications, 16 March 2026).

    In New York Times v. Microsoft and OpenAI, an order of 13 May 2025 required OpenAI to preserve deleted conversations. The obligation was lifted for new chats from 26 September 2025 (Engadget, 11 October 2025).

    On Anthropic's own record with data: in 2025 it agreed to pay 1.5 billion dollars to settle authors' claims over pirated books it had downloaded to build its library, after a judge ruled that this part of its conduct was not fair use. The settlement was finally approved on 20 July 2026 (Authors Guild, 21 July 2026). ↩︎

  3. Anthropic's Data Processing Addendum comes with its Commercial Terms, which cover the Team, Enterprise and Education plans and the API and exclude consumer plans. Anthropic, Data Processing Addendum (16 March 2026); Anthropic, Commercial Terms. ↩︎

  4. GDPR, Articles 5(2) and 24: the controller must be able to demonstrate compliance, and a breach of Article 5 can be fined under Article 83(5)(a).

    • The Court of Justice of the EU confirmed that this burden of proof lies on the controller (C-60/22, 4 May 2023).
    • The organisation, not its employee, is normally the controller (EDPB Guidelines 07/2020, version 2.0, 7 July 2021).
    • The one comparable fine, the Italian authority's 15 million euros against OpenAI (December 2024), was annulled by a Rome court on 18 March 2026, on jurisdiction rather than on the merits (Reuters via Freevacy, 20 March 2026).
    ↩︎

  5. 18 U.S.C. § 2713 (the CLOUD Act, 2018): providers must disclose data "within such provider's possession, custody, or control … regardless of whether" it is stored inside or outside the United States. The same law reaches Google, Microsoft and Apple. On 10 June 2025, a French Senate commission of inquiry asked Anton Carniaux, director of public and legal affairs at Microsoft France, whether he could guarantee that French public-sector data would never be handed to American authorities without French consent. He answered: "Non, je ne peux pas le garantir, mais, encore une fois, cela ne s'est encore jamais produit" (No, I cannot guarantee it, but, once again, it has never happened yet). French Senate, record of the hearing of 10 June 2025. ↩︎

  6. Intelligence access, espionage and the Chinese case.

    • Section 702. Section 702 of the Foreign Intelligence Surveillance Act (50 U.S.C. § 1881a) lets the Attorney General and the Director of National Intelligence authorise the targeting of non-Americans believed to be abroad and direct providers to assist. A court reviews the procedures, not the individual targets. The statute lapsed on 12 June 2026, but collection reportedly continues under existing authorisations until about March 2027 (NPR, 12 June 2026). Collection abroad under Executive Order 12333 (1981) rests on executive-branch approval; a court order is required only for collection inside the United States and for targeting Americans abroad (Privacy and Civil Liberties Oversight Board, Executive Order 12333 report, 2021).
    • What providers report. The 2013 PRISM revelations named Microsoft, Google, Apple and others among the providers compelled under Section 702. Anthropic reports receiving "0–99" national-security requests for 2024, a legal reporting band that includes zero (Anthropic, government requests report).
    • Economic intelligence. Documents published by WikiLeaks in June 2015 set out NSA collection requirements on the French economy and on contracts above 200 million dollars (WikiLeaks, "Espionnage Élysée"). In April 2015 Der Spiegel reported that the NSA had asked Germany's foreign intelligence service to monitor targets including EADS and Eurocopter (The Register, 24 April 2015). The NSA confirmed neither. The US government states that it does not collect such information to give American companies a competitive advantage (Executive Order 14086, 2022). No case involving an AI provider is documented.
    • China. China's National Intelligence Law (2017), Article 7, requires organisations and citizens to "support, assist, and cooperate" with state intelligence work; its reach outside China is disputed (National Intelligence Law, Wikipedia; Canadian Security Intelligence Service). Italy's data-protection authority blocked DeepSeek on 30 January 2025, and several European administrations, Belgium's federal administration among them, banned it from official devices (ai-regulation.com, 28 January 2026).
    ↩︎

  7. For purely private use, the GDPR does not bind the user (Article 2(2)(c)). For research data about other people, the consumer versions of Gmail, Outlook, OneDrive and iCloud have the same flaw as a personal Claude plan: their providers act as controllers under their own privacy policies and offer no data-processing agreement. Their institutional versions (Google Workspace, Microsoft 365) do offer one, and Google and Microsoft are certified under the EU-US Data Privacy Framework (official list, read 19 October 2026). Regulators still disagree about them:

    • Germany's data-protection conference (DSK) found on 24 November 2022 that compliance with Microsoft 365 could not be demonstrated (DSK). The Hessen authority departed from that view in November 2025 (HBDI report).
    • The European Data Protection Supervisor found on 8 March 2024 that the European Commission's own use of Microsoft 365 breached data-protection rules (EDPS, 11 March 2024), and closed the case on 11 July 2025 after changes (EDPS, 28 July 2025).
    • Denmark's data-protection authority seriously criticised 51 municipalities over Google's services in their schools on 2 February 2026 (Datatilsynet).
    ↩︎

  8. The EU-US Data Privacy Framework was upheld by the EU General Court on 3 September 2025 (Latombe v Commission, T-553/23; press release), and that judgment is under appeal (C-703/25 P). On 29 June 2026, the US Supreme Court allowed the President to dismiss members of the Federal Trade Commission (Trump v. Slaughter); the commission enforces American companies' commitments under the framework. On 31 July 2026, the European Data Protection Board asked the European Commission to assess whether the framework still functions (EDPB letter). The two previous agreements were struck down by the Court of Justice, in 2015 and in 2020. ↩︎

  9. Anthropic's work for the American state:

    The reported uses, both resting on anonymous sources, with no confirmation from Anthropic or the Pentagon: Reuters on Venezuela (13 February 2026, reporting the Wall Street Journal and Axios, and not independently verified by Reuters), and the Washington Post on Iran, via Stars and Stripes (4 March 2026), which describes Claude running inside Palantir's Maven targeting system.

    The dispute: Anthropic's statement of 26 February 2026. A federal judge in California found the measures against Anthropic to be unlawful retaliation on 27 August 2026 (TechCrunch, 28 August 2026), while a divided appeals court in Washington upheld the Pentagon's designation on 25 September 2026 (ABC News). Anthropic has asked for a rehearing. ↩︎

  10. The models: on the Artificial Analysis Intelligence Index, read on 19 October 2026, Claude Opus 5.5 scores about 58, OpenAI's GPT-6 Astra 53, and the best open-weight model, Xiaomi's MiMo-V2.6-Pro, 46. Epoch AI estimates that the best open models trail the best closed ones by about four months (29 May 2026). The software: Anthropic merged its Cowork agent into the main Claude app on 16 September 2026 (Anthropic), and OpenAI's Codex app became its ChatGPT desktop app on 9 July 2026. A comparison of the four main work agents found that only Claude and ChatGPT offer a documented mode working on local files (Nextstart, 5 September 2026). ↩︎

  11. Claude Pro costs 20 dollars a month before tax, or 17 dollars billed annually (Anthropic, pricing), about 21 to 22 euros with VAT. Independent measurements reported in October 2026 put the full monthly allowance of that plan at about 740 to 1,180 dollars at Anthropic's per-request prices (Wokey; SemiAnalysis, as reported by explainx.ai, 6 October 2026). The Max plan, from 100 dollars a month, offers five times the usage of Pro, which puts its full allowance at several thousand dollars at the same prices; that figure is derived from the measurements above, not measured itself. Whether any of this amounts to a loss for Anthropic rests on analysts' models, not on published figures. ↩︎

  12. The European offers in October 2026.

    • Mistral's Vibe (formerly Le Chat, renamed 28 May 2026; Pro at 14.99 euros a month before VAT) has a Work mode, with connectors to Google Workspace, Outlook, Slack and others and scheduled tasks, and a Code mode (Mistral, 28 May 2026; heise, 30 May 2026). It has no official desktop app. Its Work mode handles uploaded files and cloud connectors, and only its coding tools reach files on your own computer.
    • Mistral's models. Mistral Large 4, its strongest, previewed on 6 October 2026, scores about 38 on the Artificial Analysis index. Medium 3.5, the model behind Vibe, scores 14. On the Vals benchmark of long terminal tasks, Large 4 scores 27.8 per cent against Claude Opus 5.5's 65.2, though it does better on some specialised tests (Vals AI).
    • Apertus, the Swiss public model (EPFL, ETH Zurich, CSCS; version 1.5, July 2026), is trained only on openly available data and publishes its weights, code and recipe; its earlier version scores 5 on the same index.
    • Services running models from elsewhere. Proton's Lumo runs Qwen, GLM and Apertus. Dust runs OpenAI, Anthropic, Google and Mistral models. The German academic service Chat AI (Göttingen) runs American and Chinese open models.
    • The rest. OpenEuroLLM has not yet released a flagship model, and Germany's Aleph Alpha is being combined with Canada's Cohere.
    ↩︎

  13. Most "open" models are open-weight rather than fully open source: they publish their weights, usually under permissive licences (Apache 2.0, MIT), sometimes with conditions (Meta's), but rarely their training data and code; Apertus is an exception. On notice: Anthropic gives at least 60 days before retiring a model from its developer platform and OpenAI six months, but in the chat apps GPT-4o was withdrawn from ChatGPT on 13 February 2026 after about two weeks' notice, and Claude Haiku 4.5 leaves the Claude apps on 29 October 2026. On hardware: a 16 GB laptop runs models of about 7 to 13 billion parameters; a good mid-size model needs 24 to 32 GB of memory; the leading open models need hundreds of gigabytes (DeepSeek V4 Flash about 151 GB in compressed form), on machines such as a Mac Studio with 256 GB, at about 9,500 dollars. ↩︎

  14. Agent Skills, the format for packaged instructions published by Anthropic as an open standard on 18 December 2025, is now read by ChatGPT and Codex, Mistral's Vibe, Cursor, GitHub Copilot, and the open-source OpenCode and Goose, among others (agentskills.io). Claude Code can already be pointed at a model running on your own computer through Ollama or LM Studio, although Anthropic does not support it and small models drive it unreliably. ↩︎

Have a question about Sergueï's work? Ask me.

Ask about my services

Answers come from an AI model. Please don't include personal or confidential information.

Hello! I can help you learn about Sergueï's AI services for academics. What would you like to know?