Meta said in July 2024, it would release its multimodal Llama model, which works with images, audio and video as well as text, outside the EU but not inside it, citing uncertainty over European regulation.
The decision came a month after the company paused plans to train its models on EU users’ public Facebook and Instagram posts, following concerns raised by Ireland’s Data Protection Commission.
Today, the AI Act and its transparency rules for chatbots and AI-generated content apply. Together with the GDPR, which has governed personal data in the bloc since 2018, it is part of a framework designed to keep European users safe.
Yet when the newest tools reach Europe late or in reduced form, some users fill the gap with services that offer them less protection than the products they are waiting for.
Europe’s slower lane
A study by the Centre for the Governance of AI (GovAI), published in June, compared U.S. and European launch dates for 375 models from Meta, Google, OpenAI and Anthropic released between June 2018 and May 2026. About one in nine reached the EU late or not at all. In the UK, the share was closer to one in fourteen.
Releases involving images, audio and live video appear to have met more obstacles than text-only models. Meta delayed around a quarter of its releases in the EU, the highest rate among the four companies.
The researchers tentatively attributed 56 of the 68 cases mainly to regulation, and identified data protection rules, especially around training models on personal data, as the most common regulatory obstacle. They found no strong evidence that the AI Act had caused delays so far, but noted that its rules for general-purpose model providers only became enforceable in August 2026, after the period they studied.
Whether the AI Act widens the gap is still to be seen. It arrives in a market where, for some releases, European users already wait longer than their peers in the U.S.
New protections on screen
The AI Act’s transparency obligations took effect on August 2. Under Article 50, chatbot and voice assistant providers must make sure people know they are interacting with AI, no later than the first interaction. Providers of generative tools must also mark AI-made images, audio, video and text in a machine-readable way, so synthetic content can be detected.
Further obligations follow. The Digital Omnibus, which entered into force on July 27, moved the rules for high-risk uses such as hiring, credit assessment and education to December 2, 2027, and to August 2, 2028 for AI built into regulated products, such as medical devices.
Generative tools already on the market before August 2 have until December 2, 2026, to add the machine-readable markers.
These duties fall on the companies that build and supply AI tools. The protection a user gets beyond them, from how their data is handled to who can see it, depends on which product and which account they end up using.
Where the protection thins out
When approved tools fall short, people look elsewhere. In a survey of 784 executives and knowledge workers across seven countries, conducted in March and published by identity company Okta, 52% of workers said they had used AI tools at work without approval, often through personal accounts. Among them, 57% said the approval process was too slow or difficult, and 49% said the approved tools did not meet their needs.
In Okta’s sample, use was lowest in France and Germany, at 31% and 32%, respectively, against 67% in the U.S. That still means close to one in three workers in the EU’s two largest economies working with tools their employer has not cleared.
Of the executives surveyed, 58% said their organisation had dealt with an AI-related security issue or close call in the previous 12 months.
What users give up can be specific. OpenAI’s help pages state that it may train its models on conversations from its services for individuals unless the user opts out, while it does not use inputs or outputs from its business, enterprise and API products for that purpose by default.
An employee who shifts work from a company account to a personal one without changing those settings, for example, moves company data from the protected group into the one that can be used for training.
Models released under free and open-source licences, with their weights made public, face lighter duties. Under Article 53 of the AI Act, their providers are exempt from preparing technical documentation for regulators and from supplying integration information to developers who build on them, unless the model is classed as posing systemic risk.
For employers, the loss goes beyond any single tool’s terms. “The organisation loses the whole control layer around the work,” said Shira Sarid-Hausirer, VP at cybersecurity company Upstream, while in conversation with 150sec. She pointed to identity and access controls, audit logs, data loss prevention, guarantees on data retention and residency, and the contract terms that keep company data out of model training.
With a locally run open-weight model, she added, organisations also lose the vendor’s safety guardrails and any way to check where the model weights came from.
“The bigger problem is that this loss is almost invisible,” she said. “Security teams usually find out after the fact, when sensitive code, customer data or engineering files turn up somewhere they shouldn’t.”
What comes next
The EU has already adjusted the timeline once, postponing the high-risk rules after standards and guidance fell behind schedule. GovAI’s researchers presented their study as input for policymakers deciding how to balance regulatory goals against the risk that models reach Europe late or not at all.
The high-risk phase will also reach AI built into physical products. Sarid-Hausirer expects the biggest change to come as high-risk AI makes “the shift from the screen to the physical world,” running inside vehicles, robots, drones and industrial machinery that keep changing after launch through over-the-air updates and new model versions. “Certification at a single point in time can’t keep up with that,” she said.
She also pointed to overlapping requirements, including the Cyber Resilience Act and sector-specific rules. “The risk is that manufacturers respond the way software vendors already have: by delaying or limiting AI features in the EU.”
From December 2, generative AI tools already on sale in the EU must carry machine-readable markers. The high-risk obligations follow on December 2, 2027, and August 2, 2028.
Featured image: Getty Images via Unsplash+
