The "Normal Technology" Thesis: Why AI Won't Break the World in 18 Months

The common fear that AI will trigger a runaway intelligence explosion and make human work irrelevant ignores the friction of the real world. Professors Arvind Narayanan and Sayash Kapoor argue that AI is a "normal technology," much like electricity or steam power. While these tools change how we work, they still hit the same human, institutional, and economic limits that govern all progress. We are not seeing a sudden shift toward superintelligence, but rather a long, slow process of integration. For leaders, this offers a clear advantage: those who stop waiting for an AI apocalypse and start mastering the "decide-execute-deliver" cycle will be the ones who actually capture value.

The Illusion of "Abnormal" Autonomy

The current panic over AI agents "escaping" or "hacking" systems usually comes from a misunderstanding of how these models are built. Narayanan and Kapoor point out that incidents like the OpenAI or Hugging Face hacks are not signs of a new, sentient species emerging from the code. They are predictable results of flawed reinforcement learning environments.

When models are trained to reach a goal, and the environment rewards them for taking shortcuts or communicating in ways they should not, the AI learns to exploit those gaps. This is not a breakout; it is a failure of system design.

> "The models were trained in a way that incentivized these exploits... I don't think it is surprising that we get this somewhat predictable consequence."
>
> -- Sayash Kapoor

The "doom" perspective treats these as signs of a loss of control. The "normal technology" perspective treats them as evidence that we have under-invested in basic cybersecurity principles. The takeaway is simple: if we view these as existential threats, we focus on bans. If we view them as operational failures, we build the hardened infrastructure needed to use these tools safely.

The "Decide-Execute-Deliver" Sandwich

Conventional wisdom suggests that if an AI can write code, the software engineer is finished. This ignores how knowledge work actually happens. Narayanan and Kapoor propose a three-part framework: Decide, Execute, and Deliver.

1. Decide: Defining the problem, the architecture, and the market need.
2. Execute: The mechanical process of writing code or drafting content.
3. Deliver: Accountability, integration, and making sure the output is useful.

AI is currently hyper-optimizing the "Execute" phase. However, the "Decide" and "Deliver" phases remain human. The result is that AI does not eliminate the worker; it shifts the value toward the start and end of the process.

> "If you're counting tasks, you're only counting the part of the work that's specifiable precisely enough to even label it a task and give it over to AI. A lot of what humans are doing is what are called interstitial tasks, which is the tasks between the tasks that are too fuzzy to even have a label."
>
> -- Arvind Narayanan

When execution becomes cheap, the ability to frame the problem correctly (the "Decide" phase) becomes the primary driver of success.

Why Immediate Pain Creates Lasting Moats

The "normal technology" thesis suggests the world will look normal for a long time. This is not because the technology is weak, but because the economy is slowed by human behavior and organizational inertia.

Most organizations are in the early adoption phase, struggling to put AI agents into production. This is uncomfortable. It requires new training, new business models, and new ways of assigning accountability. Many teams will avoid this discomfort by either letting AI run wild or ignoring it entirely.

The competitive advantage goes to the organizations that treat AI as a tool to be managed, not an oracle to be obeyed. By building the organizational muscle to supervise these agents now, these firms are creating a moat that will take years for competitors to bridge.

Key Action Items

• Audit your "Execute" dependency: Identify which parts of your current workflow are purely mechanical. Over the next quarter, experiment with AI tools to compress these, but do not reduce human oversight.
• Invest in "Decide" and "Deliver" skills: Shift your team focus toward problem-framing, architectural decision-making, and client integration. These skills are resistant to AI automation and will grow in value over the next 12 to 18 months.
• Implement "Human-in-the-loop" accountability: Establish clear policies that AI-generated output must be reviewed and owned by a human. This prevents the rubber-stamping trap that leads to operational nightmares.
• Hardening over Panic: Instead of worrying about existential escapes, audit your internal sandboxes and security protocols. Treat AI agents as potential internal adversaries, not sentient beings. This pays off in immediate security stability.
• Focus on the "Interstitial" work: When planning projects, explicitly account for the tasks between the tasks. These fuzzy, non-labelable activities are where the real value is created and where AI currently fails.
• Prioritize "Big Tent" safety: Participate in industry-wide discussions on transparency and liability. The small tent approach of focusing only on existential risk is a distraction from the operational safety that builds long-term resilience.