Is it Safe to Use AI?
A Submarine Commander vs. YOLO Mode and Loss of Skill

A growing percentage of work is being delegated to coding tools and agents, throughout product development and deployment pipelines. It seems more and more time-consuming to verify technical understanding, controls and guardrails, and the business intent behind development goals. These three areas are crucial, and it is more important than ever to check human understanding regularly, throughout your product operations and deployment pipelines.

People are using YOLO mode in tools like Cursor² — "You Only Live/Look Once" — an unhinged level of enthusiasm for relinquishing control. Meanwhile, agents are going too far and deleting databases.³ Even regular ChatGPT has way too much of a go-getter "let me just take care of that for you" attitude. Adding to these concerns, there are a growing number of studies and articles detailing the loss of human expertise in areas where AI does most of the work.⁴

David Marquet transformed his submarine crew from the worst-performing in the US Navy into the highest-performing, without changing crew members.¹ He recognized that people need to be empowered to make consequential decisions with minimal oversight — but empowerment only works when the people making decisions have genuine mastery of the domain.

Regarding AI-native workflows and the use of agents: "Is it safe?"

We want frequent reassurance, with credible and sufficient confidence from an actual person. How can we build in trustworthy human oversight as more development responsibilities are delegated to AI — and de-skilling is a common risk?

Do you have enough expertise to judge safety? Are you losing it?

Marquet's Leader-Leader model has given me some useful angles on human oversight in AI-native software production. To be confident in a "yes, it's safe" answer, Marquet would certify people by asking key questions to confirm ownership of context. His certification questions focused on three pillars:

He wanted skilled leaders under his command, not followers who needed explicit rulesets to dictate each course of action. That distinction matters enormously when AI is executing consequential actions on behalf of a team.

On a related note, the loss of human expertise in AI-native environments has been flagged as a serious danger. A January 2026 Randomized Controlled Trial conducted by Anthropic revealed that while AI assistance dramatically speeds up software development, it can lead to a 17% decline in developer mastery and significantly weaken debugging skills.⁵

A culture of continuous learning has been championed at every tech company I've encountered, frequently by me. Sadly, there is a vast chasm between principled discussion and meaningful implementation. A learning platform isn't a set of embedded practices, and the licenses won't get used by the majority of people who have access. We need to verify learning and understanding, with checkpoints built into our daily workflows. The need is becoming more obvious and urgent.

A Certification Checklist for Safety and Learning

Here is an approach to ensure that a human has engaged responsibly with what an agent is producing. I recommend that managers or leads use two or three randomly selected questions from this list during any AI-assisted code review, deployment, or consequential workflow decision. Real live human people credibly answer the question, "Is it safe?" — with credible mastery, control, and clarity. Are they learning through the use of sophisticated tools, or letting the tools get out of control?

Fair warning: This will save people from de-skilling, long-term redundancy, and technical liabilities — but it will also feel strange to be quizzed this way until you've gone through it several times. There's a good justification for that strangeness. Studies have shown that taking tests, or just being informed about upcoming tests, improves learning and recall. This is the "forward testing effect."⁶

The Bottom Line

The most important safeguard is a team that understands the domain well enough to spot when the agent has gone wrong. Marquet called this mastery. It's still the prerequisite for safe delegation. Confirm human mastery by certifying people respectfully and as a matter of course, at each gate.

• When AI is doing a significant amount of heavy lifting, check for people's continuous learning and reliable grasp of context.
• Design all agentic workflows with explicit gates where the agent surfaces its intended next action, then waits for human confirmation before executing anything with real-world consequences.
• Cognitive engagement with AI solutions, reinforced in collaboration with other people, will lead to greater skills mastery rather than a reduction.