A report by Corporate Europe Observatory (COE) describes how technology companies and industry groups sought to shape the European Union’s proposed AI Act. Its central concern is that general purpose AI providers could be exempted from direct obligations, leaving much of the responsibility for safety and compliance to companies that use their models.
Why the scope of the AI Act matters
The AI Act is designed around risk. Under the approach described in the report, applications in areas such as justice, education, employment and immigration would face the strictest requirements. Other uses would face fewer obligations, while low-risk applications could follow a code of conduct.
General purpose AI models can be adapted for many different uses. COE says the debate is whether the companies that make these models should have obligations of their own, or whether rules should apply only to organizations that put them to use in specific settings.
That distinction affects who is expected to address risks such as biased training data and safety problems. If model makers have no firm requirements, the report argues, responsibility could fall on organizations further down the chain, even though they may have fewer resources to manage model-level risks.
What COE says companies argued
The report names Google and Microsoft among the companies seeking to limit the AI Act’s reach over general purpose AI. It says both argued that requirements should focus on how models are deployed, rather than applying directly to the providers of the underlying systems.
COE recounts that Google opposed proposals to place requirements on model makers, saying they would shift too much responsibility to providers. A paper from Google, obtained through freedom of information requests, reportedly argued that general purpose AI systems are not themselves high-risk and that some compliance requirements would be difficult or impossible to meet in practice.
Microsoft expressed a similar position in an open letter to the Czech Presidency of the Council. The company said it saw no need for a specific section on general purpose AI and argued that providers could not meet high-risk requirements without knowing a tool’s purpose. COE also describes private discussions in which Microsoft raised possible effects on startups and small and medium-sized enterprises.
Lobbying through industry groups
COE’s account includes both direct lobbying and efforts through third-party organizations. It says a September 2022 letter from BSA | The Software Alliance urged EU institutions to reject proposals on general purpose AI, arguing that they could burden developers and hamper innovation.
The report notes that BSA was created by Microsoft in 1988. It also says some members of the European Digital SME Alliance had been approached about signing the letter, but were advised against doing so because they saw no benefit for smaller companies.
COE questions why Allied for Startups signed the letter, given that the organization describes itself as a network focused on startup policy. The report says its sponsors include Google, Apple, Microsoft, Amazon and Meta. It adds that although the sponsors are said to have no voting rights, observers have noted that the group’s positions closely aligned with those of Big Tech.
The report also points to a Commission expert group that advises on AI policy, saying industry representatives—including Google—are prominent in it. Together, these examples illustrate COE’s concern that large technology companies can influence the debate both directly and through organizations presented as representing broader interests.
Who carries the risk if providers are excluded?
COE warns that exempting general purpose AI providers could create a gap in the AI Act. Companies that adapt a model for a particular service might have to answer for its effects, while the providers that developed the underlying system would face fewer direct duties.
That could be difficult for smaller organizations. They may not have the resources to address problems embedded in a model, yet could still face compliance costs or liability tied to how they use it. The report argues that this arrangement could leave people exposed to biased or unsafe systems without clear accountability from the companies best positioned to address risks at the model level.
The AI Act was still under negotiation when the report was published. COE said the issue of general purpose AI had been pushed into later discussions, as the legislation approached closed-door trilogue negotiations involving the Council, Parliament and Commission. Microsoft told TechCrunch it sought to be a constructive and transparent partner to European policymakers. Google had not responded at the time of publication.