The Lieutenant

What Defines a Lieutenant

You use AI as a thinking partner, not a vending machine. You give it a task, read the output, and then do something that most people never do: you push back. You ask follow-up questions. You stress-test ideas. You tell the AI to challenge its own assumptions. You push it to surface things you have not considered.

Most people quit when AI gives a bad answer. They close the tab, try a different tool, or decide AI "does not work." That reaction is understandable, but it is also the single biggest barrier to AI proficiency. The first answer is almost never the best answer. The value lives in the second, third, and fourth iteration. The gap between what AI gives you unprompted and what AI gives you after sustained pushback is enormous.

At Level 2, you learned to give AI clear instructions. Good prompts. Structured inputs. That matters, and it puts you ahead of most people. But Level 2 is still transactional. You put in a request, you get a response, you move on. Level 3 is where the interaction becomes a conversation. You treat AI output as a draft, not a deliverable. You evaluate it, challenge it, and demand better.

The human skill behind this level is self-management. That might sound generic. It is not. Self-management in the context of AI is a very specific thing: the ability to tolerate frustration, maintain persistence, and keep iterating when the easy path is to accept whatever you got and move on. It is the skill that separates the people who get mediocre results from AI from the people who get exceptional ones.

The Science of Self-Management

Daniel Goleman's emotional intelligence framework identifies four clusters. Self-awareness is the first. Self-management is the second. It is the internal regulation system that determines what you do with the information self-awareness gives you.

Goleman broke self-management into four competencies. Emotional self-control is the ability to keep disruptive emotions and impulses in check. Achievement orientation is the drive to improve performance and meet internal standards of excellence. Positive outlook is the ability to see the upside in events, situations, and other people. Adaptability is flexibility in handling change and managing multiple demands.

Goleman defined emotional intelligence as "being able to motivate oneself and persist in the face of frustrations; to control impulse and delay gratification." That definition maps directly to what happens when you work with AI. The frustration is real. The impulse to accept a mediocre answer is real. The ability to delay gratification, to keep pushing when the first three attempts were not good enough, is what produces results that most people never see.

Self-confidence is a critical enabler here, and it sits at the intersection of self-awareness and self-management. Goleman's research through the EI Courses framework showed that people with high self-confidence are more likely to persist through setbacks because they trust their own judgment. They know the AI output is not good enough because they have an internal standard to measure it against. Without that standard, you accept whatever AI gives you. With it, you push until the output meets your bar.

The Wind4Change analysis of Goleman's work highlights a pattern that applies directly to AI use: self-management is not about suppressing emotion. It is about channeling it. Frustration with AI is useful when it drives you to iterate. It is destructive when it drives you to quit. The difference is not the frustration itself. It is what you do with it.

Grit: The Research on Persistence

Angela Duckworth's research on grit, published in the Journal of Personality and Social Psychology in 2007, provides the most rigorous scientific framework for understanding why persistence matters more than talent in almost every domain. Including AI.

Duckworth defined grit as passion and perseverance for long-term goals. Not just effort. Not just endurance. Grit requires both a sustained interest in what you are pursuing and the willingness to keep working at it when the work gets hard, boring, or frustrating. Both components are essential. Perseverance without passion burns out. Passion without perseverance never finishes anything.

Duckworth tested grit across three demanding populations. At West Point, grit predicted which cadets would survive Beast Barracks, the brutal summer training program that eliminates roughly 5% of each class. It predicted retention better than the Whole Candidate Score, which is West Point's own composite measure of academic ability, physical fitness, and leadership potential. At an Ivy League university, grit predicted GPA, even after controlling for SAT scores. Among National Spelling Bee competitors, grittier contestants practiced more hours and advanced further in the competition.

The critical finding for AI proficiency is that grit can grow. It is not a fixed trait. Duckworth's later research, supported by the Digital Promise framework for educator development, showed that grit responds to deliberate practice, to environments that reward persistence, and to experiences where sustained effort produces visible results. You are not born persistent or not. You build persistence through repeated exposure to situations where quitting is easy and pushing through is rewarding.

This is exactly what happens when you use AI at Level 3. Every time you push back on a weak answer and get a better one, you are building grit in the context of AI use. Every time you resist the impulse to accept the first response and instead ask "what did you miss," you are strengthening the neural pathways that make persistence automatic rather than effortful.

The Hallucination Problem

Critical thinking is not optional with AI. It is a survival skill. The reason is hallucination, and the numbers should concern anyone who uses AI output without verification.

Vectara's hallucination leaderboard tracks how often leading models fabricate information. The best models hallucinate at roughly 0.7% on document summarization tasks, where the AI has the source material right in front of it. On open-ended general knowledge questions, the rate jumps to 9.2%. That is nearly one in ten responses containing fabricated information on straightforward factual questions.

In specialized domains, the rates are staggering. Medical AI systems hallucinate at rates between 64% and 68%, according to systematic reviews. Stanford's RegLab found that legal AI tools hallucinate between 69% and 88% of the time on legal research tasks. These are not fringe tools. These are commercial products marketed to professionals who make consequential decisions based on their output.

The business cost is not theoretical. AI hallucinations cost organizations an estimated $67.4 billion in 2024 through flawed decisions, recalled content, compliance violations, and wasted labor. A survey from the Columbia Journalism Review found that 47% of executives admitted to making business decisions based on AI-generated content they never independently verified.

This is the problem that Level 3 solves. Not by avoiding AI. Not by distrusting everything AI produces. But by developing the habit of verification, of asking follow-up questions, of cross-referencing claims, and of never treating AI output as ground truth without examination. The Lieutenant does not assume AI is wrong. The Lieutenant does not assume AI is right, either. The Lieutenant checks.

Critical Thinking Under Threat

There is a documented risk that AI use actively degrades your ability to think critically. This is not speculation. It is measured data from a major study published at CHI 2025, the top venue for human-computer interaction research.

Lee et al. (2025) from Microsoft Research and Carnegie Mellon University studied 319 knowledge workers using generative AI tools in their daily jobs. Their findings are direct and unsettling. The correlation between AI usage frequency and critical thinking engagement was r = -0.68. In behavioral science, that is a strongly negative correlation. The more people used AI, the less they engaged their own critical thinking faculties.

The mechanism is straightforward. Higher confidence in AI outputs correlated with less critical evaluation. When people trusted the AI, they stopped checking its work. Conversely, higher self-confidence in one's own abilities correlated with more critical thinking. People who trusted themselves more than the tool were more likely to verify, challenge, and improve what the AI produced.

The study documented a fundamental shift in how knowledge work happens with AI. Thinking moves from information gathering to verification. Instead of researching a topic yourself, you ask AI to research it and then verify what it found. Problem-solving shifts to response integration. Instead of working through a problem, you review the AI's proposed solution and decide whether it holds up.

Perhaps the most concerning finding: most users asked only one question per problem with no follow-up. One prompt, one response, done. No iteration. No pushback. No "what did you get wrong" or "what are you assuming." The vast majority of AI interactions are single-turn, accept-the-first-answer exchanges. That is Level 2 behavior. Level 3 is where you break that pattern.

The researchers were clear: the solution is not to use AI less. It is to use AI differently. With deliberate verification. With follow-up questions. With the self-management discipline to resist the path of least resistance and actually engage with the output before acting on it.

Productive Failure

Manu Kapur, a learning scientist at ETH Zurich, has spent over a decade researching a counterintuitive finding: struggling with difficult problems before receiving instruction produces better learning outcomes than receiving instruction first.

He calls this productive failure, and the evidence is robust. A meta-analysis covering more than 12,000 participants across multiple studies found that students in productive failure conditions significantly outperformed those in instruction-first classrooms on conceptual understanding and transfer tasks. They did not just learn the material better. They could apply it to new problems they had never seen before.

Kapur identified four mechanisms that explain why productive failure works. First, it activates prior knowledge. When you struggle with a problem, your brain pulls in everything it already knows about related topics, creating a richer foundation for new learning. Second, it directs attention to critical features. Failure highlights what matters and what does not, because you learn which aspects of the problem actually drove the outcome. Third, it promotes explanation and elaboration. When something goes wrong, you ask why, and that question drives deeper processing than passively receiving the right answer ever could. Fourth, it supports organization and assembly. Struggling forces you to build your own mental models rather than accepting someone else's framework wholesale.

The design principle Kapur emphasized is that the problems must "challenge but not frustrate." There is a threshold. If the challenge is too easy, there is no productive struggle. If it is too hard, people disengage entirely. The sweet spot is problems that are solvable with effort but not solvable without it.

This maps directly to Level 3 AI use. When AI gives you a bad answer and you push back, you are in a productive failure condition. You are struggling with the gap between what you got and what you needed. That struggle activates your critical thinking, forces you to articulate what "good" actually looks like, and builds the evaluation skills that make every future AI interaction better.

The research from Sinha and Kapur (2021) extended this further, showing that productive failure benefits are durable. They persist over time and transfer to new domains. The critical thinking skills you build by pushing back on AI today will serve you in contexts that have nothing to do with AI tomorrow.

Practical Exercise: The Push-Back Protocol

What Comes Next

Your thinking is strong. You push back, you verify, you iterate. But here is the problem you have not solved yet: your conversations degrade over time. You start a session with sharp, specific outputs. Twenty messages later, the AI is repeating itself, losing your constraints, and producing generic filler. You know the output has gotten worse, but you do not know why or what to do about it.

The next level is managing the AI conversation lifecycle. Not just what you say to the AI, but the environment your words exist in. Context management. Conversation architecture. Knowing when to start fresh and what to carry forward.

That is Level 4: The Commander.