A report highlighted Sunday has renewed scrutiny of the United States’ fast-evolving system for overseeing frontier artificial intelligence, after Sixth Street Vice Chairman Martin Chavez warned that the current approach is inconsistent, opaque and vulnerable to unequal treatment. The central problem, in his view, is that federal officials appear to be handling major model releases one at a time without a sufficiently clear public framework explaining who decides, what standards apply and what consequences follow.

Chavez delivered the criticism on June 30 at the Reuters Momentum AI London conference. His comments have gained greater relevance as Washington and leading AI developers work through a new process for testing advanced models before broader deployment. Chavez, who also serves on Alphabet’s board, described the U.S. approach as “incredibly problematic” and “inconsistent,” according to Reuters. He focused less on whether the government should examine powerful systems than on the absence of a predictable, repeatable structure for doing so.

That distinction matters. A government can impose demanding safety expectations while still giving companies a clear rulebook. It can also pursue a formally light-touch policy that becomes unpredictable in practice because decisions are made through private negotiations, informal requests or overlapping agency reviews. Chavez’s warning is that the United States is drifting toward the second outcome: consequential intervention without the transparency normally associated with a mature regulatory regime.

The debate intensified after recent interactions involving OpenAI and Anthropic. Axios reported July 10 that both companies’ newest, most capable models received government attention before wide release, even though the administration does not characterize the process as formal approval. The White House said it did not approve or disapprove OpenAI’s release decision. That clarification underscored the ambiguity facing developers: federal participation may be described as voluntary, but companies building strategically important models still have strong incentives to satisfy officials responsible for cybersecurity, export controls and national security.

The legal foundation for the emerging process is President Donald Trump’s June 2 executive order on advanced AI innovation and security. The order directs multiple agencies to develop classified benchmarks for assessing the cyber capabilities of advanced models and deciding when a system qualifies as a “covered frontier model.” It also calls for a voluntary framework allowing developers to engage the government before launch.

Under that framework, participating companies may provide federal officials access to covered models for as long as 30 days before releasing them to other trusted partners. Developers can also work with the government to identify organizations that should receive early access. The order expressly states that those provisions do not authorize mandatory licensing, preclearance or permitting for new AI models. The administration’s challenge is ensuring that real-world implementation remains consistent with that limitation.

For technology companies, the difference between consultation and permission is commercially significant. A delayed model release can alter cloud-computing commitments, enterprise adoption schedules, developer road maps and expected revenue. Uncertainty can be particularly damaging to smaller laboratories that lack the government-relations teams, legal resources and security infrastructure available to the largest companies. An informal system may therefore strengthen incumbents even when no rule explicitly favors them.

It can also complicate investment decisions. Sixth Street is a major investment firm, and Chavez has spent his career at the intersection of finance, software and risk management. Before joining Sixth Street, he held senior positions at Goldman Sachs, including chief information officer and chief financial officer. His regulatory argument reflects a financial-system perspective: markets function more effectively when institutions know how regulators will test them, what data they must provide and how failures will be addressed.

Chavez pointed to the annual stress tests created for large financial institutions after the global financial crisis. Those examinations subject banks to standardized adverse scenarios and measure whether they have enough capital to withstand severe economic shocks. The system is not free from criticism, but its basic structure is legible. Covered institutions understand the process, regulators publish key assumptions, and results can be compared across firms and over time.

An AI equivalent could establish common evaluation categories for cyber offense, biological misuse, autonomous replication, deception, data security and resistance to safeguards. It could also define escalation levels, disclosure obligations and remediation procedures. A model that crosses a published risk threshold might receive additional testing or a staged release, while a system below that threshold could proceed without improvised negotiations. Such a framework would not eliminate judgment, but it could make government judgment more accountable.

Applying financial stress testing to AI would nevertheless be difficult. Bank balance sheets can be measured using established accounting concepts, while advanced models display capabilities that may change after fine-tuning, tool access or deployment. Safety evaluations can become outdated quickly, and public disclosure of certain tests could help adversaries evade them. Classified benchmarks may protect sensitive information, but excessive secrecy would reproduce the transparency problem Chavez identified.

Standardized testing could also be manipulated if companies optimize specifically for known evaluations. Regulators would need to update scenarios, use independent testing and examine how systems behave outside laboratory conditions. They would also need enough technical personnel to challenge claims made by developers. Axios reported that experts remain concerned about the federal government’s ability to recruit and retain AI specialists, a capacity problem that cannot be solved through executive orders alone.

The governance question extends beyond individual launches. Without a durable federal structure, companies must navigate executive actions, agency authorities, national-security reviews and state requirements simultaneously. Congress has not enacted a comprehensive national AI safety law, leaving much of the current system dependent on executive policy and voluntary cooperation. That makes procedures more vulnerable to changing administrations, emergencies and disputes among agencies.

The White House’s forthcoming voluntary framework will therefore be an important test. The June order set a 60-day timetable, placing the expected deadline around August 1. Developers will be watching for clear definitions of covered models, consistent evaluation methods, confidentiality protections, timelines for government feedback and a process for resolving disagreements. They will also want assurances that participation will not quietly become an unavoidable prerequisite for commercial release.

Chavez’s criticism does not amount to an argument against AI oversight. His stress-test proposal assumes that frontier systems can create risks serious enough to justify structured government examination. The objection is to governing through one-off interventions whose authority and standards are difficult for outsiders to evaluate. Predictable oversight could support both safety and competition by giving companies a common set of expectations rather than rewarding those best positioned to navigate informal channels.

The immediate policy challenge is to reconcile the administration’s stated opposition to mandatory preclearance with the practical demands of national-security review. If Washington can create transparent thresholds, technically credible evaluations and equal treatment for competing developers, voluntary cooperation may become a workable model. If it cannot, Chavez’s warning will remain unresolved: the United States may have significant power over frontier AI releases without a sufficiently coherent system for exercising it.