The European Union has published a fresh draft of rules for companies using "general-purpose" artificial intelligence models such as OpenAI's GPT-4 or Mistral's Large — but critics warn the proposals still raise "significant concerns" for tech companies operating on the continent.
By far the most comprehensive set of laws governing the technology in the world, the AI Act introduces a long list of new rules for organisations using the tech in the EU, with an emphasis on ethics, safety and transparency.
Since late last year, experts from industry, academia and nonprofits have been wrangling over a key document: the code of practice for general-purpose AI (GPAI).
Behind the scenes, lobbyists from all sides have been hashing out how large language models (LLMs) like those deployed by OpenAI, Google and Mistral will be regulated.
A matter of urgency
On Tuesday, the European Commission published the third draft of the code of practice. Proposals include prying open AI companies' training data to look for potential copyright breaches and allowing external experts to inspect their models for safety purposes.
Anselm Küsters, divisional head of digitalisation at leading think tank the Centre for European Policy, said the latest draft contains "serious ambiguity" around the obligations of small-and-medium-sized businesses.
"If a startup fine-tunes a GPT-4.5-level model after the obligations come into force, it risks inheriting compliance obligations without clarity on scope," he said.
Küsters also flagged that models already available on the market will not be retroactively bound by transparency rules, meaning startups tweaking them downstream will go without access to training data or copyright policies. "The AI Office needs to clarify these issues as a matter of urgency."
Serious issues
With the third draft of the GPAI code of practice published, the EU now enters the final stage of negotiations, where final tweaks will be made to the document based on stakeholder feedback.
Boniface de Champris, a senior policy manager at CCIA (Computer & Communications Industry Association) Europe — an industry body which counts Amazon, Google and Meta among its members — said the draft still raised "significant concerns" for tech companies.
“Serious issues remain, including far-ranging obligations regarding copyright and transparency, which would threaten trade secrets, as well as burdensome external risk assessments that are still part of this latest iteration," he said.
“The new draft makes limited progress from its highly problematic predecessor, yet the GPAI code continues to fall short of providing companies with the legal certainty that’s needed to drive AI innovation in Europe. In its current form, the code still risks directly undermining the EU’s digital competitiveness.”
Sifted approached the European Commission for comment.
