The Transparency Window Is Closing Fast

The 2026 Stanford AI Index has surfaced an alarming trend: the Foundation Model Transparency Index collapsed from 58 points last year to just 40 points in 2026. This 31% drop signals a fundamental shift in how the world’s largest AI companies operate—and it’s happening precisely as European regulators are trying to enforce new oversight rules.

The pattern is clear: Anthropic, Google, OpenAI, and other frontier labs are increasingly withholding critical information about model training, including dataset sizes, parameter counts, and computational costs. The very transparency mechanisms that underpin responsible AI governance are being systematically dismantled.

Why This Matters for Ireland and Europe

This timing couldn’t be worse for Ireland’s incoming EU Presidency and the newly established Irish AI Office. The EU AI Act explicitly requires transparency around model capabilities, limitations, and training data. Yet the entities building the most powerful systems are moving in the opposite direction.

The Foundation Model Transparency Index measures public disclosures across 10 dimensions—from model documentation to risk assessment details. When leaders like Anthropic and Google stop publishing this information, it creates an asymmetric information problem: regulators lack the data needed to enforce compliance, while competitive pressure incentivizes other labs to follow suit and go dark themselves.

For Irish tech companies and builders, this creates immediate friction. Companies deploying foundation models for regulated use cases (recruitment, financial services, customer service) will struggle to document their systems’ behaviour in ways the AI Act requires—precisely because upstream model developers aren’t disclosing the technical foundations.

The Competitive Race to Obscurity

The irony is bitter: as AI capabilities accelerate, the companies building them are becoming less accountable. The Stanford report shows that the best-performing models (Claude Opus 4.6, Gemini 3.1 Pro) are also the least transparent. With competitive margins razor-thin and regulatory oversight tightening, there’s a perverse incentive to keep architectural details proprietary.

This race to obscurity has real consequences. Without access to training methodologies and dataset composition, independent researchers can’t audit for bias, detect capability risks, or verify safety claims. Regulators lose their ability to distinguish between genuine innovation and marketing hype.

What Irish Builders Need to Do Now

Document everything upstream. If you’re building on foundation models, keep detailed records of model versions, capabilities tested, and limitations discovered. This creates an audit trail when regulators inevitably demand transparency from your organisation.

Demand disclosures from vendors. Contractual clauses requiring foundation model providers to disclose training methodology and risk assessments will become table stakes. Start asking for these now.

Prepare for regulatory asymmetry. Ireland’s distributed enforcement model (15 sectoral regulators) means compliance expectations may vary. Build systems flexible enough to satisfy multiple interpretations of the AI Act.

Open Questions

Will the August 2026 EU AI Act transparency deadline force labs back into the light, or will the first major enforcement actions reveal that regulators have no teeth? And will the intelligence advantage gained by keeping models opaque outweigh the reputational cost when that opacity is eventually exposed?

The Stanford data suggests we’re about to find out.

Source: Stanford AI Index 2026