Four AI Labs Ask Washington to Watch AI That Improves Itself

Senior executives from Anthropic PBC, OpenAI, Meta Platforms Inc. and Microsoft Corp. called on policymakers on Sept. 28, 2026 to scrutinize how far AI systems are improving themselves and to take steps to safeguard the technology, Bloomberg Technology reported.

Executives at four of the companies building the most capable artificial intelligence systems in the world have asked policymakers to look closely at something the industry has mostly discussed among itself: the degree to which AI models are now helping to improve the next generation of AI models.

Senior figures from Anthropic PBC, OpenAI, Meta Platforms Inc. and Microsoft Corp. are urging regulators to scrutinize the extent of that self-improvement and to take steps to safeguard the technology, Bloomberg Technology reported on Sept. 28, 2026. The appeal adds to a widening set of calls from inside the industry for greater outside oversight.

Why self-improvement is the line the industry keeps pointing at

"Self-improving AI" is shorthand for systems that contribute materially to their own development — writing or optimizing the code for training runs, generating the synthetic data used to teach successor models, designing evaluations, or automating the research work that used to require human engineers. In its mild form, it is already ordinary practice: engineers at every large lab use models to accelerate model building. In its extreme form — what safety researchers call recursive self-improvement — capability gains compound faster than the humans supervising them can measure.

The distinction matters for policy because oversight regimes are usually built around snapshots. A regulator certifies a product, an auditor reviews a model before release, a lab publishes a system card. A system that meaningfully improves itself between those checkpoints breaks the snapshot. That is the specific gap the executives are pointing at: not whether AI is dangerous in the abstract, but whether anyone outside the labs can see how quickly the frontier is moving and by what mechanism.

It is also a rare instance of direct competitors converging on the same request. Anthropic and OpenAI compete for the same enterprise contracts; Meta has pushed open-weight releases that undercut the closed-model business; Microsoft is simultaneously a cloud supplier, an investor and a product rival across the field. Getting all four to agree that a category of capability deserves outside scrutiny is a signal about how the labs read their own trajectory.

What an oversight mechanism would actually have to do

The lead facts do not spell out which specific instruments the executives want, and no one should assume a particular bill or agency is behind the push. But the practical requirements of monitoring self-improvement are not mysterious, and they are worth setting out because they define the cost to the companies if anything is enacted.

• Disclosure of training inputs. Any regime would need to know what share of a model's training data, code or evaluation design was machine-generated rather than human-authored.
• Compute reporting. Visibility into the scale of training runs is the closest thing the field has to a measurable proxy for capability jumps.
• Pre-deployment testing with teeth. Evaluations that a third party can run, rather than results a lab reports about itself.
• Continuous rather than one-shot review. A model that is being updated by automated pipelines is a moving target, and a single certification at launch says little about what it can do months later.
• Incident reporting. A channel for labs to flag unexpected capability gains, analogous to adverse-event reporting in pharmaceuticals.

Each of those imposes real operating friction. Compute disclosure touches competitive secrets. Third-party testing means handing model access to outsiders. Continuous review would change release cadence at companies whose product cycles are currently measured in weeks.

The market barely rewarded the gesture

The call landed on a soft session for technology stocks. As of the last trade at 20:00 GMT on Sept. 28, 2026, the Nasdaq 100 tracker (QQQ) was at $736.53, down 1.07% on the day from a previous close of $744.50, with an intraday range of $731.63 to $741.42. The S&P 500 proxy (SPY) was at $765.61, off 0.74%, and the Dow tracker (DIA) at $514.02, down 0.67%. Tech, in other words, was leading the market lower rather than following it.

Two of the four companies named are publicly traded. META was quoted at 715.62 against a previous close of 751.66, a fall of 4.79% on the day, with a session range of 713.19 to 750.58 — a decline roughly 3.7 percentage points steeper than the Nasdaq 100's, on an illustrative comparison of the two moves. MSFT stood at 509.22 versus a prior close of 516.17, down 1.35%, having traded between 502.22 and 513.33. Anthropic and OpenAI are privately held and have no listed shares to react.

Meta's move was the outlier of the day, and nothing in the lead ties it to the oversight statement. Investors reading a governance headline from a company whose stock is already down nearly 5% are entitled to treat the two as separate events until evidence connects them.

What to watch from here

The useful test of an industry-led call for oversight is whether it converts into anything a company can be held to. Three things would distinguish this from previous rounds of AI-safety rhetoric.

First, whether the executives name a measurable threshold — a compute level, a benchmark score, a share of automated research — above which extra scrutiny attaches. Vague appeals invite vague rules; specific ones create enforceable obligations.

Second, whether the companies accept binding third-party access to frontier models, or limit themselves to voluntary reporting. Voluntary regimes tend to hold until the moment a competitive deadline collides with a disclosure obligation.

Third, whether the ask survives contact with jurisdictional politics. A federal framework, a state patchwork and overseas regimes pull in different directions, and companies asking for oversight in general are frequently asking for one particular regime in place of several inconsistent ones. That preference is not disreputable — it is simply a different request from the one the headline describes, and worth separating out as any legislative text appears.

For investors, the near-term financial exposure is modest: compliance and testing costs are small next to the capital going into data centers and model training. The longer-term question is whether limits on automated self-improvement would slow the capability curve that current AI valuations are built on. That is the number nobody has yet put a figure to.

Key facts

• Companies whose executives made the call: Anthropic PBC, OpenAI, Meta Platforms Inc., Microsoft Corp.
• META share price: 715.62, down 4.79% on the day, as of 20:00 GMT Sept. 28, 2026
• MSFT share price: 509.22, down 1.35% on the day, as of 20:00 GMT Sept. 28, 2026
• Nasdaq 100 (QQQ): $736.53, down 1.07%, as of 20:00 GMT Sept. 28, 2026

Frequently asked questions

What did the AI executives actually ask for?

Senior executives from Anthropic PBC, OpenAI, Meta Platforms Inc. and Microsoft Corp. urged policymakers to scrutinize the extent to which artificial intelligence systems are improving themselves, and to take steps to safeguard the technology. Bloomberg Technology reported the appeal on Sept. 28, 2026. It adds to a growing set of calls from within the industry for greater external oversight of frontier AI.

What does 'self-improving AI' mean in practice?

It describes AI systems that contribute to building their own successors — generating training data, writing or optimizing training code, designing evaluations, or automating research that engineers previously did by hand. Mild versions are already routine at large labs. The concern is the compounding case, where capability gains accelerate faster than human supervisors can measure or review them.

Which of the four companies are publicly traded?

Meta Platforms Inc. and Microsoft Corp. have listed shares; Anthropic PBC and OpenAI are privately held and have no public stock. As of the last trade at 20:00 GMT on Sept. 28, 2026, META was quoted at 715.62, down 4.79% on the day, and MSFT at 509.22, down 1.35%.

Did the announcement move the market?

There is no evidence linking the oversight call to share price moves. Technology stocks were broadly lower that session: the Nasdaq 100 tracker QQQ finished at $736.53, down 1.07%, the S&P 500 proxy SPY at $765.61, down 0.74%, and the Dow tracker DIA at $514.02, down 0.67%. Meta's decline was notably steeper than the index.

Why would competing AI companies jointly ask for regulation?

Companies that expect rules regardless often prefer to shape them, and a single federal or international framework is cheaper to comply with than a patchwork of conflicting regimes. Labs also face genuine internal safety concerns about capability jumps. The commercial and the sincere motives are not mutually exclusive, and the published details do not separate them.

What would monitoring self-improvement require?

Practical mechanisms would likely include disclosure of how much of a model's training data and code is machine-generated, reporting on compute used in training runs, third-party pre-deployment testing, continuous rather than one-time review, and incident reporting when unexpected capability gains appear. The executives' specific proposed instruments were not detailed in the initial report.

Sources

• Anthropic, OpenAI Executives Urge Oversight of Self-Improving AI — Bloomberg Technology