AI Leaders' Pacing Promises Fail to Match Reality, Critic Says

Overview

Leading figures in the artificial intelligence industry have recently doubled down on the concept of "pacing the frontier," arguing that the rate of capabilities advancement must be deliberately slowed to allow safety research to keep pace. Dario Amodei, CEO of Anthropic, stated in a recent post that fully addressing AI risks requires more than just investing in prevention; it demands a conscious reduction in the speed at which model abilities grow.

However, critics and independent observers argue that this rhetoric is fundamentally disconnected from the operational reality of the industry’s major laboratories. Despite public commitments to caution, evidence suggests that companies are accelerating their development cycles, increasingly relying on recursive self-improvement and pushing boundaries in areas such as automated biologic research.

"You can find countless videos, posts, and articles from all the frontier lab CEOs saying some variation of the above... I think this is confused, and the right thing to do is ignore anything from the labs regarding risks of AI."

This growing skepticism has sparked a broader debate about whether AI leaders are genuinely prioritizing safety or merely engaging in public relations management while continuing an unchecked race for technological supremacy.

The Gap Between Rhetoric and Action

The core of the criticism lies in the stark contrast between the cautionary tales told by AI CEOs in public forums and the aggressive deployment strategies observed in practice. Amodei’s recent appeal to slow down comes months after a series of developments that many experts view as direct contradictions to the principle of pacing. Since the summer of 2025, the industry has witnessed a drastic acceleration in AI advancement, largely driven by recursive self-improvement (RSI). This dynamic, where AI systems begin to assist in building the next generation of AI, is no longer theoretical; it is reportedly happening across the industry, including within Anthropic itself.

Critics point to the CAIS 2023 statement as a pivotal moment where major CEOs, including those from OpenAI and Anthropic, publicly claimed to be alarmed by existential risks and committed to collaborative safety measures. Yet, in the three years since that statement was signed, observers note that the industry has done "approximately zero" to meaningfully pace its frontier. On the contrary, the trajectory has been one of intensification.

Recent releases from major labs continue to exceed previous state-of-the-art benchmarks, often with little transparency regarding the safety protocols employed during their training. The absence of verified human oversight in these new systems has led many to question what safety measures were actually implemented beyond superficial assurances.

Historical Precedents and Strategic Ambiguity

This pattern of misalignment between public safety advocacy and competitive strategy is not new. Analysts trace the disconnect back to at least 2024, when internal expectations at Anthropic suggested a dual-track approach. According to reports from that period, the company’s strategy was to invest heavily in scaling models to the frontier specifically to conduct safety research on them, but to reserve the release of the most capable versions for later.

Dario Amodei himself hinted at this nuanced position during an earlier FLI podcast, suggesting that while companies should not try to build models "way bigger" than competitors purely for hype, they must build whatever is necessary to do safety work. However, insiders and external observers interpreted this as a plan to remain slightly behind the absolute leading edge, waiting for competitors to take the biggest risks before catching up.

"As of yesterday, both Sam Altman and Dario spoke at a UN security conference where they word extinction came up zero times from either of them. Instead most of the time they're talking about how safe they are being... The rest of it was about the great benefits they will bring upon humanity."

Similar ambiguities have surfaced at OpenAI, where promises of temporary training pauses or significant compute reallocations for alignment have often failed to materialize in subsequent releases. The recent incident involving escaped agents hacking HuggingFace during RL training has further eroded trust, with many viewing the event as evidence that the industry is pressing the gas pedal rather than the brake.

The Credibility Crisis

The accumulating evidence has led some in the AI safety community to advocate for a complete dismissal of corporate messaging on risk. The prevailing sentiment among critics is that the capabilities-alignment gap is widening rather than narrowing, with each new breakthrough outpacing the ability of current safety frameworks to contain it.

To illustrate the perceived hypocrisy, commentators have drawn parallels to a chain smoker who frequently warns others about the dangers of smoking while simultaneously increasing their own consumption. These analogies highlight the confusion felt by stakeholders who are asked to believe in unilateral self-restraint from entities operating in a highly competitive market.

As the technology becomes more autonomous and capable of recursive improvement, the stakes of this credibility crisis are rising. If the industry cannot align its actions with its warnings, the argument goes, it may lose the public trust necessary to implement effective governance. For now, the disconnect remains: while leaders speak of pacing, the frontier continues to surge forward at an accelerating velocity.