The word "voluntary" appears eight times in the executive order that Donald Trump signed on June 2, 2026 for AI safety. An earlier draft required 90-day mandatory review, then the tech lobby prevailed. OpenAI, Anthropic, and Google can now submit new AI models for security review or skip it entirely.
What Trump Signed
The executive order targets operators of frontier models, AI systems at the technological frontier. Companies can submit new models up to 30 days before public release to the national security advisor. Security authorities can review models for risks and issue recommendations. Blocking publication is explicitly excluded.
The document's wording is clear: "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models." The order directs federal agencies to develop benchmarks for assessing AI models based on cyber capabilities and establish an information channel for AI security gaps. From a security standpoint the document targets less the AI market than securing government systems against hostile use.
From 90-Day Draft to Voluntary Compromise
An earlier executive order draft required 90-day mandatory review. Government and security agencies would gain not just insight but intervention options for models deemed dangerous. This draft was dropped in May 2026 after Trump argued overly strict rules endangered American AI competitiveness against China. The compromise: 30 days, voluntary, no veto. OpenAI and Anthropic welcomed the result. Anthropic had previously stated the need for government review of large AI models, conditioning this on sufficient technical expertise on the government side and oversight through democratically accountable institutions.
EU Requirement: Fines Starting August 2
In the European Union, voluntary self-regulation for AI content is not the model. Article 50 of the AI Regulation takes effect August 2, 2026. It requires AI system operators and providers to label AI-produced content: texts, images, videos, and audio. Violations risk fines up to 15 million euros or three percent of global annual revenue. For companies the size of Google or Meta, the three-percent threshold applies.
The Bundesnetzagentur (Federal Network Agency) was designated Germany's supervisory authority and receives 49 million euros annually according to the federal government. The regulation includes an exception in Article 50, Paragraph 4: anyone assuming identifiable personal editorial responsibility for content is exempt from labeling requirements. According to the German Journalists Association, pro forma approval doesn't count.
The China Argument and Its Limits
Trump and American tech industry share a premise: overly strict regulation favors Chinese AI companies. The argument has a real basis. China has controlled AI model exports since 2023 while continuing to develop powerful systems without Western security standards.
Civil society critics of voluntary self-regulation point to platform regulation history. AlgorithmWatch in Berlin and Netzpolitik.org note voluntary self-commitments by digital platforms regularly fail: Meta pledged until 2018 not to misuse user data and lost 87 million user profiles to the Cambridge Analytica scandal. Without fear of sanctions, companies prioritize speed over safety, goes the structural argument against the Trump approach.
YouTube offers a counterexample from its own initiative: the platform introduced automatic AI detection for videos in May 2026 without legal compulsion. The label now appears on photorealistic AI content automatically, even if creators don't declare it. For videos made with Google's own Veo 3 model, the label is permanent and uncontestable. YouTube, one of three companies named in Trump's order, simultaneously operates under EU labeling requirements, existing in two different regulatory regimes.
Nine Weeks Until EU Fines
On August 2 the EU deadline expires. Media law experts expect first fine proceedings by fall 2026 when the Bundesnetzagentur opens sanction cases after initial violations. The trick for affected companies lies in the editorial exception: whoever verifies content with a named responsible person is exempt. How strictly authorities interpret this will emerge in initial test cases. The cybersecurity benchmarks Trump's order asks agencies to develop haven't even been defined; the White House set no deadline for completion. When first voluntary AI submissions arrive in Washington remains unclear.
The regulatory gap between Brussels and Washington thus grows wider for now, not narrower. For companies operating in both markets: in Europe, AI labeling requirements are legally enforceable starting in August. In the US a voluntary channel exists whose use no one monitors. Whether that suffices will emerge when a major new model is announced and the question arises whether the channel was used.