The Metrics Mirage: Can We Trust Industry’s Self-Regulation?

The tech world has been abuzz with calls for a slowdown in AI development, sparked by Dario Amodei’s three-step plan to temper model capabilities without sacrificing commercial advantage or US dominance. Amidst this commotion, Anthropic’s release of three metrics to monitor pace of development seems more like a Band-Aid on a bullet wound than a meaningful attempt at self-regulation.

These metrics can provide a starting point for assessing AI development, as Anthropic claims. The company measured its own Claude models’ autonomy, the number of AI agents overseeing internal actions, and compute allocation toward safety. Transparency is indeed heartening from an industry notorious for opacity. However, upon closer inspection, it becomes clear that these metrics are more akin to window dressing than a genuine attempt at accountability.

Determining what constitutes autonomy in this context raises as many questions as it answers. Is autonomy merely the absence of human oversight or something more nuanced? Without clear definitions, these metrics risk becoming exercises in semantics rather than genuine attempts at measurement. For instance, Anthropic’s claim that “no subset” of research and development work is done autonomously by Claude models leaves much to be desired.

Anthropic’s second metric – building a system to oversee and intervene in AI agent actions – is similarly opaque. The company proudly declares that approximately 30,000 agents are doing research and engineering work across its internal platform. However, what does this number really tell us? Are these agents truly autonomous or merely programmatic entities designed to execute tasks without human oversight? Without more context, it’s impossible to say.

The most egregious example of industry self-regulation is Anthropic’s third metric: measuring compute allocation toward safety. The company claims that roughly 6% of its compute went toward AI research and development was allocated toward safety – a number that seems suspiciously low given the existential risks associated with uncontrolled AI growth. What does this number really say about Anthropic’s commitment to safety? Is it a genuine attempt at mitigating risk or merely a public relations exercise?

Anthropic’s metrics are not without precedent, however. In 2018, Google introduced its “DeepMind principles,” which aimed to address concerns around AI development and use. These principles – including transparency, accountability, and fairness – seemed like a step in the right direction at the time. However, as we’ve seen with subsequent industry developments, these promises have largely gone unfulfilled.

In light of Anthropic’s metrics, it’s clear that the tech industry is still struggling to balance its pursuit of innovation with its responsibility to society. Rather than relying on self-regulation and the goodwill of industry leaders, we need more concrete measures – like regulatory oversight, public accountability, and genuine transparency – to ensure that AI development aligns with human values.

Anthropic’s metrics may provide a starting point for assessing AI development, but they are hardly a panacea for the problems plaguing this industry. As the tech world hurtles toward increasingly complex and potentially catastrophic applications of AI, it’s time to demand more from our leaders – not just empty promises or metrics that mask deeper issues.

The stakes are too high to rely on self-regulation alone. We need real action, not window dressing.