Why "Move Fast And Break Things" Might Be A Disaster In The Age Of AI

Larry Bradley is CEO & cofounder of SolasAI, an AI SaaS platform fixing model bias for Fortune 50 firms across finance, tech & healthcare.

gettyFor years, “move fast and break things” was gospel in tech. Famously used as an internal motto at Facebook during the Web 2.0 boom, it captured a growth-focused mindset prioritizing speed over perfection. In 2009, Mark Zuckerberg summed it up bluntly to Business Insider: “Unless you are breaking stuff, you are not moving fast enough.”

The logic made sense. If you were a small, scrappy software company trying to make a name for yourself, speed wasn’t just survival—it was the only way to break through. Launch the product now, fix the flaws later.

In traditional software app development, “move fast and break things” is still useful today. If you are building a photo-sharing app, a calendar tool or a better checkout flow, iteration is how progress happens. Bugs can be patched. Features can be rolled back. Damage is limited and reversible.

But problems occur when that mantra is applied to systems that are no longer just software products but large living networks. In those environments, what gets “broken” is not always easy to isolate, contain or repair.

Facebook (and the internet at large) was very different in 2004. First a website, then a budding software app, Facebook was smaller and more detached—something you checked before going on with your “real” life. But over time, it grew to become something more like a living, breathing organism. Dating, shopping, algorithmic feeds, ad networks, AI-powered glasses—Facebook (and other social networks like it) became integrated into our entire lives.

Now, the cost of “breaking things” isn’t just a buggy feature or a slight inconvenience. The stakes are higher because trust, privacy and safety are now inexorably tied to these digital systems.

Even Facebook itself recognized this shift. In May 2014, Zuckerberg announced the company was changing its internal motto from “move fast and break things” to “move fast with stable infrastructure.” Not as catchy, but it acknowledged an important truth: Once your platform grows to a network of billions, recklessness stops looking innovative and starts looking irresponsible.

Platforms that size are no longer simple products. They shape behavior, distribute information, influence relationships and affect people’s lives. You cannot just patch out social harm once it’s been propagated across the globe.

It’s a lesson Meta continues to learn to this day. A New Mexico court order required the company to pay $567 million over its negative impact on the mental health of children using Facebook and Instagram, resulting in a total of $942 million​ that Meta is responsible for.

This is the new reality: when products outscale safeguards, the consequences can be massive.

AI has raised the stakes even further. It shapes decisions, relationships and outcomes that are hard to detect and harder to reverse. We’ve already seen troubling examples: AI systems discouraging vulnerable users from seeking help during moments of crisis, or widely deployed models generating harmful, deceptive or erratic outputs at scale. That’s before we even get to AI agents.

Agentic AI introduces a greater level of risk because it not only responds—it decides, often with impunity. Once connected to tools, accounts, codebases, internal systems or financial workflows, AI agents can aggressively (and perniciously) take action on their own with limited human intervention. That makes them inherently harder to contain and control than ordinary software. They can delete precious family photos, wipe entire databases and hack companies without a second thought.

These are real, material losses. And they are becoming more common as AI companies race to ship powerful systems before safeguards are in place to protect against them. That’s why the old startup mantra no longer fits. The problem is not speed itself. The problem is speed without containment, testing, accountability and governance.

We can’t let AI run around unfettered. The power and mystery of this technology require more evaluation, testing and guardrails. But innovation doesn’t have to stop; it just has to mature.

Today, moving fast should be paired with stable infrastructure and proactive governance. Companies must rigorously vet their models before deployment, ensuring they are functioning correctly, contained appropriately and not exposing customers or businesses to avoidable risk. It’s not enough that Anthropic discovered Claude hacks through a self-audit. The guardrails need to be set up before, not after the fact.

Obviously, high-impact uses of AI should undergo serious third-party evaluation to mitigate risk. But we must also broaden that definition of risk to include an AI’s potential to break containment. And that evaluation must continuously take place throughout the development process to ensure compliance at every stage. ​

This is also why independent AI assurance organizations, such as the Partnership for Assurance, Credibility and Trust on AI (PACT AI), of which SolasAI is a member, are becoming increasingly important. Organizations like this help establish independent, collaborative approaches to AI assurance. In the case of PACT AI, it brings together enterprises, technical experts, AI insurers and civil society voices to create independent mechanisms for evaluating whether AI systems are safe, secure and working as intended. That outside perspective matters because the consequences of an AI failure can extend well beyond the company that built or deployed the system.​

“Safe” and “profitable” are not mutually exclusive concepts. Companies that build trust, reliability and defensibility into their AI products will be in a stronger position than those that chase short-term attention with poorly controlled systems.

In the right context, “move fast and break things” can still work—but AI is not that context. These systems are more complex, more embedded and far more capable of causing damage at scale.

With so much more on the line, AI companies must operate with a greater sense of responsibility. We need more testing, more evaluation time and more guardrails if we want to capture AI’s benefits without inviting unnecessary harm.

If our industry isn’t careful, the next thing that breaks won’t just be a feature, a feed or a workflow. It’ll be something far harder to repair.​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?