Leopold I, Leopold II, and Leopold III. They were kings of Belgium. Robert Schuman 7. That is not the first president of the European Parliament. That’s an address. You know this. Anthropic’s Claude does not.
A few months ago, the European Parliament began using Anthropic’s AI models to provide access to its archives. You can access it on the parliament’s website. It is called ‘Ask the EP Archives’. You can ask all you want but do not trust the response. “Generated content should not be considered authoritative.” (That’s the parliament’s disclaimer.)
It is a valid question to ask whether a European company should have been chosen over Anthropic. I will leave that to someone else. I had much more basic questions for the parliament. Why was generative AI chosen for this purpose? What was the procurement process?
The parliament wants to provide digital access to its archive. The access should be swift and the responses should be correct. Not 97 percent of the time, but 100 percent of the time. In computer science, this is a well-studied problem. It is called information retrieval.
Earliest research on this topic goes back to the 1940s and deployments have been known since 1960s. Search engines such as AltaVista and Duckduckgo are modern day examples. The parliament’s case of course does not require crawling the web as it possesses a large millions of documents going back to 1952.
Generative AI systems like Claude, on the other hand, are bullshit generators. They are like a gossiper. Sometimes the gossip includes facts. Sometimes it does not. You would not know which is which. The gossiper might augment the bullshit with retrieval, as Claude does, but it is still a gossiper.
When you want facts, you do not ask the gossiper. When you want facts, you do not ask generative AI.
So why did the parliament choose a gossiper when it needed an information retrieval tool? If we believe the parliament’s head of the archive’s unit, it is Constitutional AI: a set of principles chosen by Anthropic, which the company claims will minimise harm. This approach is based on “simple and memorable” criteria and has not been independently verified.
The parliament believes in Anthropic’s Constitution instead of the EU rules on AI. Anthropic uses web scraped data to train its AI models, an approach that is likely unlawful due to the processing of sensitive personal data.
No. When I asked the parliament for contract and tender documents, I was informed that “there is no direct contract for parliament’s use of Amazon Bedrock, and Claude as provided by Anthropic.” This is because the parliament depends on “the European Commission acting as an intermediary for cloud brokering service.”
Let us break this down.
The commission has a contract for cloud services with Amazon AWS. The parliament uses this contract for cloud services. But this contract was not for AI. Neither the commission, nor the parliament has a contract with Anthropic for AI.
Why is this possible? Amazon runs a market place for AI models: Amazon Bedrock. If you have a contract for Amazon AWS, you can access “fully managed [AI] models” on Amazon Bedrock. The parliament accesses AWS through the commission’s contract, and accesses Claude through Amazon Bedrock.
Over the years, guidelines have been proposed for public sector organisations when procuring AI systems or services. The UK has guidelines since 2020.
Such guidelines usually consider whether AI is necessary for the intended purpose, whether there is a vendor lock-in, suggests risk assessments and other governance aspects. These guidelines are intended to complement the laws that the public administration is expected to follow anyway.
Amazon Bedrock is an additional service from Amazon to lock-in customers. What we see in the parliament’s case is that they have bypassed public procurement processes by using Amazon Bedrock.
Without a direct contract with Anthropic, the parliament skipped risk assessment, skimped on environmental assessment and did not consider data protection impact assessment relevant.
Between 2020 and 2023, the commission paid Amazon AWS nearly €55m for cloud services. With the commission and parliament’s use of AI services in 2024, it is not yet known how many more millions the commission and the parliament are paying Amazon, which also funds Anthropic.
The parliament’s head of archive unit, when promoting Anthropic, made a Freudian slip and said, “we permanently need to be under [sic] control of the solution that we’ve built.”
The parliament is indeed under the control of Amazon.
Dr Kris Shrishak is a senior fellow at Enforce, the Irish NGO campaigning for human rights in technology, and a public interest technologist and advises legislators on emerging technologies and global AI governance. He is regularly invited to speak at the European Parliament and has testified at the Irish Parliament. His work focusses on privacy tech, anti-surveillance, emerging technologies, and algorithmic decision making. Previously, he was a researcher at Technical University Darmstadt in Germany where he worked on applied cryptography, privacy enhancing technologies and Internet security.
Dr Kris Shrishak is a senior fellow at Enforce, the Irish NGO campaigning for human rights in technology, and a public interest technologist and advises legislators on emerging technologies and global AI governance. He is regularly invited to speak at the European Parliament and has testified at the Irish Parliament. His work focusses on privacy tech, anti-surveillance, emerging technologies, and algorithmic decision making. Previously, he was a researcher at Technical University Darmstadt in Germany where he worked on applied cryptography, privacy enhancing technologies and Internet security.