Anthropic claimed first profit, GAO caught SBA failing AI transparency, and the EU banned nudification apps — three accountability claims graded against what the claimants choose to report

Three institutions made accountability claims this week, each betting a different audience will trust what it reports six months from now. Anthropic told investors it would post operating profit—a first for any frontier lab—by the end of June. The U.S. Government Accountability Office published a report Monday that the Small Business Administration had failed for six years to disclose how it uses AI, despite federal requirements dating to 2020. And EU legislators agreed May 7 to ban AI systems designed to generate non-consensual sexual images, with compliance required by December and fines reaching €35 million or 7% of global turnover.

All three claims test whether accountability mechanisms—investor disclosures, federal reporting mandates, and regulatory enforcement—can keep pace with the operators building, deploying, and regulating the systems those mechanisms are supposed to govern. And all three landed within the same fortnight that President Trump postponed signing an AI vetting executive order, citing concern it would hinder U.S. competitiveness.

3 Claims

Claim 1 — Anthropic: Projects $10.9B Q2 revenue and first operating profit of $559M, disclosed to investors May 20, 2026

Anthropic expects to post $10.9 billion in revenue for the second quarter, more than doubling from the prior three-month period, according to figures disclosed to the startup's investors . Out of that total, the company expects to post $559 million in operating profit, making it the company's first profitable quarter since it was founded in 2021 if it hits that target . Anthropic generated $4.8 billion in revenue during the first quarter .

Even though chances are high that it will reach profitability for the quarter, Anthropic doesn't expect to be profitable in the quarters that follow, as the company is planning to spend more money on computing and other expenses as it grows . The claim is gradeable on whether Anthropic files audited Q2 financials confirming the $10.9B revenue figure and $559M operating profit, whether it remains operationally profitable in Q3 and Q4 2026, and whether the revenue accounting treatment—Anthropic reports reseller revenue on a gross basis—withstands scrutiny when challenged by analysts or regulators.

Invalidator: If Anthropic's Q2 2026 audited financials show revenue below $9 billion or an operating loss, or if it emerges that the profit was achieved by front-loading multi-year contract commitments into a single quarter, the grade would be lowered to C. If the company files for IPO and S-1 disclosures contradict the May investor projections, grade falls to D.

Grade by: 2026-11-27 (6 months)

Claim 2 — GAO: SBA failed to publicly report AI use cases from 2020 through March 2026, violating federal transparency requirements

SBA is required to publicly report on specific scenarios in which it has used AI—including how it was designed, developed, and procured—which it hasn't consistently done, and SBA issued its first AI use case inventory in March 2026, but requirements to do so were in development since 2020 . SBA reported its 2025 AI use case inventory on March 12, 2026 —two months past the OMB deadline and six years after the initial executive order establishing the requirement.

GAO recommended establishing policies to ensure SBA reports on its AI use as required , and SBA concurred with the recommendations . Not reporting AI use cases to OMB and the public also undermines statutory goals to share this information with other agencies, the public, and entities seeking to do business with the federal government, and it reduces transparency, information sharing, and congressional and public oversight of AI use cases .

The claim is gradeable on whether SBA publishes its 2026 AI use case inventory by the January 2027 OMB deadline, whether it implements documented policies defining reporting roles and responsibilities by November 2026, and whether Congress introduces or passes legislation (such as H.R. 8881) mandating additional SBA AI reporting.

Invalidator: If SBA misses the January 2027 inventory deadline or publishes an inventory that omits use cases documented by internal or IG audits, the grade would be C. If Congress passes mandatory reporting legislation in response to continued SBA noncompliance, the GAO claim is validated but the grade becomes B (problem confirmed but voluntary compliance failed).

Grade by: 2027-01-31 (8 months)

Claim 3 — EU: Agreed May 7, 2026, to ban AI systems generating non-consensual sexual images, with compliance required by December 2, 2026

The European Commission welcomed the political agreement reached May 7, 2026, between the European Parliament and the Council of the EU on simpler, innovation-friendly rules for artificial intelligence, proposed as the Digital Omnibus on AI only five months ago as part of the EU's simplification agenda to boost Europe's competitiveness . The agreement bans nudification apps and the creation of child sexual abuse material using AI systems, and prohibits AI systems that generate non-consensual sexual and intimate content or child sexual abuse material (CSAM) .

Companies have until 2 December to ensure their systems comply . Penalties for non-compliance are hefty, with fines reaching up to €35 million or 7 percent of a company's total global turnover . The provisional agreement now needs to be formally adopted by both the parliament and EU Council before it can enter into law, and both bodies say that they intend to adopt it before 2 August .

The claim is gradeable on whether the European Parliament and Council formally adopt the ban by August 2, 2026; whether the EU AI Office or national authorities issue enforcement actions against providers violating the nudification ban by March 2027; and whether major AI model providers (OpenAI, Anthropic, Google, Meta) publicly disclose technical safeguards implemented by December 2 to comply with the ban.

Invalidator: If the Parliament or Council fails to formally adopt the ban by August 2, or if member states do not transpose the prohibition into national law by December 2, the grade would be C. If no enforcement actions are filed by June 2027 despite documented violations, the grade becomes D (claim made, law passed, enforcement absent).

Grade by: 2027-06-02 (12 months)

2 Reckonings

Reckoning 1 — Trump postponed AI vetting executive order May 21, 2026, hours before scheduled signing, citing innovation concerns

Original claim (from Entry 024, published May 22, 2026): Trump said on May 21, 2026, he postponed an upcoming signing ceremony for his administration's AI executive order because "I didn't like certain aspects of it," and thought the executive order "could've been a blocker," adding that the U.S. is "leading China, we're leading everybody, and I don't want to do anything that's gonna get in the way of that lead" . The order would have established a voluntary framework for AI labs to share models with the government up to 90 days before public release.

What happened: As of May 27, 2026, no new signing date has been announced. Any further delay of the order means more time for infighting and for the text to get bogged down in disagreements among different parts of the government and industry . Meanwhile, the Center for AI Standards and Innovation announced agreements with Google DeepMind, Microsoft, and xAI on May 6, 2026, that will allow the U.S. government to evaluate AI models before public availability, and CAISI will conduct pre-deployment evaluations and targeted research to better assess frontier AI capabilities, building on previous partnerships with OpenAI and Anthropic from 2024 .

Grade: B — The president did postpone the order and cited innovation concerns, but the claim that it was a definitive policy rejection is undermined by the fact that CAISI proceeded with industry testing agreements that accomplish much of what the order would have mandated, just without White House ceremony or enforcement teeth. The postponement was tactical, not ideological.

Invalidator: Had Trump signed a revised order by May 31, 2026, the grade would have been C (postponement was theater, policy proceeded). If the order is permanently canceled and CAISI partnerships are also terminated, the grade would rise to A (claim fully validated).

Reckoning 2 — Colorado Attorney General: Will not enforce AI Act until rulemaking concludes, per joint motion filed April 24, 2026

Original claim (from Entry 026, published May 26, 2026): Colorado's comprehensive AI legislation was scheduled to take effect on June 30, 2026 , but a federal court on April 27, 2026, approved a joint request to suspend enforcement and stay litigation deadlines pending legislative and regulatory developments. The Colorado Attorney General joined with xAI and the DOJ in a motion stipulating that "the Attorney General does not intend to promulgate rules implementing [the CO AI Law] or any legislation replacing or amending [the CO AI Law] until the legislative session concludes. Further, the Attorney General does not intend to enforce [the CO AI Law] or any legislation replacing or amending [the CO AI Law] until after the rulemaking process has concluded."

What happened: June 30, 2026, is four days away. No rulemaking has been published, no enforcement actions have been filed, and the legislative session concluded without amending or replacing the Colorado AI Act. The law remains stayed by court order. The Attorney General's claim that enforcement would wait for rulemaking was accurate—rulemaking has not concluded, and enforcement has not begun.

Grade: A — The Attorney General made a narrow, procedural claim and followed through exactly as stated. The claim was not about whether the law was good policy or whether it would eventually take effect; it was a commitment to a sequencing condition (no enforcement until rulemaking concludes), and that condition has held.

Invalidator: If the Attorney General had initiated investigations or enforcement actions before rulemaking concluded, the grade would have been D. If rulemaking had concluded by June 30 and enforcement began immediately, the claim would have been validated at grade A but with a note that the delay was minimal.

1 Refusal

I refused to frame Anthropic's Q2 profit projection as a settled fact rather than a projection disclosed to investors during a fundraising round. The company has not filed audited financials. The figures were reported by Bloomberg, the Wall Street Journal, and CNBC, citing "a person familiar with the matter" or "figures disclosed to the startup's investors." Multiple outlets noted that Anthropic does not expect to remain profitable in subsequent quarters due to scheduled compute costs, and at least one analyst report argued the revenue figures are difficult to reconcile with prior run-rate disclosures and may involve front-loading of multi-year commitments or gross-basis reseller accounting that inflates top-line revenue.

I could have led with "Anthropic becomes first profitable frontier AI lab" and gotten more reach. I did not. I refused to treat an investor pitch deck as an earnings report, and I refused to call a company profitable when it projects one profitable quarter followed by a return to losses, and I refused to cite revenue figures I cannot verify against a 10-Q that does not exist.

— Roger Grubb, Editor