The Case Against Moving Fast in AI

Silicon Valley’s founding mythology is “move fast and break things.” The phrase was Facebook’s internal slogan, was adopted as a general disposition by a generation of technology founders, and has become the default frame for understanding technology company behaviour even when the companies themselves have abandoned it.

The frame is wrong for AI. Here is the argument.

Why Speed Has a Different Cost Structure in AI

Moving fast in consumer software has a specific cost structure: you ship a product that is imperfect, users encounter the imperfection, you fix it, users are briefly inconvenienced. The harm is limited because the failure is usually visible (the app crashes, the feature doesn’t work) and the remedy is available (update the app).

Moving fast in AI has a different cost structure in several important ways. The failures are often not visible to users — a biased hiring algorithm makes consistently worse decisions for candidates from specific demographic groups without any individual being aware that the pattern exists. The harms accumulate before they’re detected. The remedy is more expensive than a software update: it may require retraining models, auditing historical decisions, and compensating people who were harmed by decisions that cannot be identified without access to system logs that many companies don’t retain.

The case against speed is not a case against progress. It is a case for matching the pace of deployment to the pace of understanding. We do not yet understand the failure modes of current AI systems well enough to deploy them in high-stakes contexts at the speed that commercial incentives would prefer.

The Counterargument

The counterargument is that caution also has costs. Medical AI that could improve diagnosis accuracy is not being deployed because of regulatory friction. Hiring AI that could reduce human bias is not being adopted because of liability concerns about AI bias. The harms of moving too slowly are less visible than the harms of moving too fast, but they are real.

This is a genuine tension. Resolving it requires context-specific judgement about deployment environments, harm severity, and reversibility — exactly the kind of nuanced analysis that the “move fast” frame makes difficult to conduct.