Can Overlong Lectures Become a Talent in the AI Age?

Can Overlong Lectures Become a Talent in the AI Age?

A Hypothesis About Oververbalization, Management Experience, and AI Agent Aptitude

1. The Problem

The central hypothesis of this essay may sound unsettling: people with tendencies that can appear as power-harassment traits may be unusually suited to using AI and AI agents well.

This is not an attempt to justify violence, personal attacks, intimidation, insults, ignoring people, or forcing excessive work on others. Japan's Ministry of Health, Labour and Welfare defines workplace power harassment as conduct based on a superior relationship that exceeds the necessary and reasonable scope of work and harms the working environment. Since April 2022, prevention measures have been mandatory for all employers in Japan, including small and medium-sized enterprises, and workplace harassment remains a major labor issue. Harassment itself is clearly a social loss and is undesirable for both companies and individuals.

However, a trait that appears troublesome in human relationships can turn into a different capability when the recipient is AI. Some people continue lecturing for more than 10 minutes. Some turn a five-line email into 30 lines. Some cannot stop until they feel the other person has understood; they explain the same point from another angle, add examples, wander into side paths, and eventually try to push their thinking into the other person.

If this is directed at a human subordinate, it is likely to lower psychological safety, damage trust at work, and reduce productivity. But AI does not get tired. AI is not personally wounded by long text. Even when the speaker wanders, AI can extract goals, constraints, background circumstances, priorities, and implicit assumptions. Modern large language models often perform better when they receive context, examples, constraints, evaluation criteria, and failure cases rather than a bare command. In other words, the volume of language that becomes an excessive lecture toward a person may become a rich prompt toward AI.

To examine this hypothesis, we have to break down the rough phrase "power-harassment tendency." In this essay, I divide it into several components. First, low resistance to speaking for a long time. Second, a strong desire to make the other party understand one's thinking. Third, a tendency to explain goals, background, steps, prohibitions, and evaluation criteria in detail. Fourth, persistence in giving additional explanations and correction instructions when the other party does not act as expected. Fifth, experience as a manager or leader who has delegated work to others and inspected results.

These traits create friction in human relationships. But in AI use, they look quite important. For Pocho Lab, the point is not that an unpleasant boss is right. The point is whether part of the pressure that should never be directed at a human being may be reevaluated as work-design ability when directed at AI.

2. Look at Oververbalization, Not Harassment

The first important step is not to connect harassment directly to AI aptitude. Power harassment injures another person's dignity, and there is no need to find value in it. Instead, we should remove the harmful components that make behavior harassment and examine whether the remaining behavioral traits help with AI use.

For example, a long lecture contains multiple ingredients. If it includes anger, a desire to dominate, self-justification, and contempt for the other person, it is harmful. But it may also contain the ability to decompose a situation, explain background, give concrete examples to correct behavior, and translate abstract principles into practical judgment.

Imagine a manager saying this to a subordinate:

"This document does not work. The numbers are correct, but the order is not the order the audience needs. This audience is not the finance department; it is the sales department, so leading with market size will not land. You should first show the customer problem, then the adoption effect, and finally the cost effectiveness. Also, the competitive comparison should not only be a table. You need one sentence that says where we can win. I said the same thing last time, but a document is not a warehouse of information. It is a tool for decision-making."

If this is poured over a human being for 10 or 20 minutes, the person will be exhausted. With a strong voice, facial expression, and hierarchy in the room, explanation turns into pressure. But if the same content is entered into AI, the situation changes. AI receives the goal, audience, problem, evaluation criteria, past failure, and improvement direction all at once. This is a fairly high-quality instruction.

In prompt design, specificity, role, goal, constraints, examples, and output format matter. OpenAI's practical guidance recommends clear and specific instructions, sufficient context, and iterative refinement. Anthropic's guidance on prompt engineering and context engineering also emphasizes success criteria, examples, constraints, and long-term context management rather than short commands alone. Saying "fix this deck" is weaker than explaining who it is for, why it matters, what is wrong, and what standards should guide the revision.

What appears here is an oververbalization tendency. Toward humans, it is excessive. Toward AI, excess itself is not necessarily a weakness. A person who can think while talking, remember the goal halfway through, add constraints, supply exceptions, and finally arrive at "what I really want is this" may draw better output from AI than a person who gives a short command while still unclear about the goal.

3. AI Skill Is Less Question Skill Than Context-Supply Skill

People sometimes call good use of generative AI "question skill." In practice, it is closer to context-supply skill. AI does not read the human mind. It estimates the next useful response from the words provided, prior conversation, supplied materials, and constraints. Therefore, a strong AI user is not simply someone who writes short and beautiful questions. It is someone who supplies the necessary context without leaving too much out.

In that sense, people who write long emails are not automatically disadvantaged. Long text alone has no value. If signal and noise are too mixed, AI can misunderstand. But if the long text contains the goal, audience, background, failure examples, evaluation criteria, desired output, and expressions to avoid, it becomes useful material.

In a 2023 experiment by Noy and Zhang, 453 college-educated professionals, including marketers, consultants, data analysts, and human-resource workers, used ChatGPT for writing tasks. The reported result was that time spent fell by about 40% and quality ratings rose by about 18%. This suggests that AI is not only a faster writing tool. It externalizes human thinking, creates drafts, and accelerates revision.

But the same experiment also shows that AI is not magic. If the human side has no clear idea of what it wants to achieve, output becomes average and thin. Conversely, people who have a large amount of material, can state judgment criteria, and can look at the output and say "No, not that" are more likely to improve quality through repeated interaction with AI.

Consider again the structure of a long lecture. It often proceeds in this order: point out a problem, explain why it is a problem, recall past cases, infer the other person's thinking pattern, explain the same point through another situation, and finally state what should be done next. Toward a person, this is tiresome. Toward AI, it is rich context for an improvement task.

For example, there is likely to be a large difference between asking AI "improve this sales deck" and asking:

"Improve this sales deck. The audience is the information systems department of a large enterprise. They have budget, but they are cautious about new tools. The goal this time is not to close adoption immediately, but to move them to the next technical validation meeting. The current deck explains too many features and does not answer their concerns. They care especially about security, integration with existing systems, and operational burden after adoption. Our advantages over competitors are shorter initial setup and easier-to-read audit logs. Please use a tone that reduces the customer's risks one by one rather than pushing the sale aggressively."

This is long, but it is a good instruction. People who can naturally input this way are strong in the AI age. The question is whether that ability is aimed at humans or entrusted to AI.

4. Why Management Experience Resembles AI Agent Operation

An AI agent is not merely an AI that returns text. It is a system that works toward a goal through multiple steps, gathers information, makes judgments, uses tools when needed, and creates deliverables. In this setting, the human role becomes closer to manager than questioner. The human sets the goal, decomposes work, checks interim output, corrects errors, and remains responsible for final quality. This is very close to management.

Microsoft's 2025 Work Trend Index describes the rise of the "agent boss," someone who builds, delegates to, and manages AI agents. The report says that 28% of managers are considering hiring AI workforce managers to lead hybrid teams of people and agents, and 32% plan to hire AI agent specialists within the next 12 to 18 months. It also suggests that teams will increasingly build, train, and manage agent-based systems over the coming years.

The important point is that AI skill is not purely technical. In Microsoft's data, leaders were ahead of employees in familiarity with agents, regular AI use, trust in AI for high-stakes work, and expectation of managing agents. The issue is not only whether someone touches AI every day. It is whether they can decompose, delegate, and review work.

This is where management experience becomes an advantage. Managers routinely ask other people to do work. They also know that if the request is vague, the deliverable will drift. A person who has failed after telling a subordinate "just make it good" learns that goals, deadlines, priorities, evaluation criteria, and prohibitions must be made explicit. The same applies to AI agents.

When asking a junior employee to research a market, a capable manager does not stop at "look up competitors." The manager specifies which market, what comparison axes matter, who the decision-maker is, how to treat numerical reliability, how to reconcile past documents, and how many final slides are needed. For an AI agent too, "research competitors" is not enough. The person must specify research scope, source priority, comparison items, output format, separation of facts and inference, and risks to check.

In this respect, people who usually give too many detailed instructions to subordinates may be advantaged when dealing with AI. Of course, toward human subordinates, this can deprive autonomy and impede growth. But current AI agents do not have human dignity or career development needs. Detailed instructions, repeated revisions, interim checks, and rework can function as quality control.

What managers need in the AI age is not only kindness. They need the ability to decompose work into language. People who explain too much may already be doing this decomposition unconsciously.

5. Indirect Evidence From Research and Statistics

There is still limited large-scale research directly proving this hypothesis. We do not yet have an established paper showing that higher power-harassment-tendency scores predict better AI-agent performance. But related findings provide several pieces of indirect evidence.

First, generative AI has already spread quickly into workplaces. Stanford HAI's 2025 AI Index reports that 78% of surveyed organizations used AI in 2024, up from 55% the previous year. McKinsey's 2025 work on AI in the workplace reports that 92% of executives expect to increase AI investment over the next three years, while only 1% of leaders describe their organizations as mature in AI deployment. In other words, many organizations are investing in AI, but operational mastery remains early.

Second, AI works well when tasks contain transferable context. Brynjolfsson, Li, and Raymond studied customer-support agents at a scale of roughly 5,000 people and reported that generative AI assistance increased productivity by about 14% to 15% on average, with more than 30% improvement among novice or lower-skilled workers. One interesting implication is that AI acted as a medium for transferring experts' tacit knowledge to less experienced workers. Phrasing, response order, and judgment patterns from skilled agents became available to newer agents.

Third, AI effects depend on task boundaries. In an experiment involving 758 knowledge workers, Dell'Acqua and colleagues found that, for tasks inside AI's capability frontier, participants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and improved quality. But on tasks outside the frontier, AI users performed worse. The study describes AI capability as a jagged frontier: tasks that appear similarly difficult to humans may differ sharply in whether AI is good or bad at them.

To identify this frontier, one needs a manager's eye that does not swallow output whole. The over-involved person returns here. Someone who can say to AI, "this premise is wrong," "this comparison axis is missing," or "this wording will not reach the audience" is better positioned than someone who simply accepts the output. If the essence of the long lecture is not domination but detailed verbalization of work standards, that trait suits AI review.

Fourth, thinking while speaking has cognitive meaning. Psychology has a research area called self-explanation. Chi and colleagues' work in the 1990s showed that learners deepen understanding by explaining while solving problems. A 2018 meta-analysis by Bisra and colleagues examined 69 effect sizes from 64 reports and found that self-explanation prompts had a moderate effect on memory, comprehension, and application. Explanation is not only for the other person. It also reorganizes one's own thinking.

People who think while speaking wander. But in those detours, the true purpose may emerge. Someone may start by saying "I want to fix the deck," and while talking discover that the real issue is "I do not want to create anxiety for the customer," "we lack material to persuade the internal approver," or "the goal is not the contract but the next meeting." This process is too long for a human listener, but toward AI it can become excavation of thought.

Fifth, leadership research provides a supporting line. A 2002 meta-analysis by Judge and colleagues reported that extraversion is a relatively consistent correlate of leadership. Extraversion includes not only sociability but assertiveness, energy, and talkativeness. However, extraversion does not automatically mean empathy or listening ability. This is important. Talkative people may be strong with AI, but not necessarily strong with humans. Speaking skill and listening skill are different, and AI-age aptitude depends less on loudness than on whether explanation can be structured.

6. Revaluing the Long-Email Person

Every workplace has long-email people. The subject line is simple, but once you open the message, the scrolling does not stop. You think the conclusion appeared in the first three lines, but then come the background, past history, stakeholder names, risks, alternatives, and extra notes. For the reader, this is honestly difficult. On a busy day, one may want to close the message the moment it arrives.

But when this long text is thrown at AI, the evaluation can change. AI can extract issues, summarize them, break them into tasks, and turn them into a short message for a decision-maker. Even if the person writing the long message has not organized everything, AI can act as organizer if the raw material is present.

Here lies a strange reversal of the AI age. People who were told "just get to the point" at work may become strong when dealing with AI because they can supply a large amount of material. Being concise is excellent. But a person who can only be concise may give AI too little material. If AI has to infer too much from silence, it returns generalities. When input is thin, output tends to be average.

A long-email person can work with AI in the following way. First, write out every frustration, concern, and goal in the mind. Then ask AI to organize it by issue. Next, soften the parts that are too harsh for a human recipient and separate content that should be communicated from emotional expression. Finally, send only a short message to the human. In this way, AI becomes a buffer that compresses excessive verbalization into a socially usable form.

This can also reduce harassment risk. A person who gives long lectures should talk to AI before talking to a subordinate. Explain to AI for 10 minutes, including anger, background, and what seems wrong. Then ask AI: "From this content, create a message that does not harm the other person's dignity and communicates only the work-related improvement points." The heat of emotion may be cooled by AI, and the human recipient may receive only organized feedback.

Of course, this is not a cure-all. If one asks AI to amplify anger, it may produce aggressive wording. That is why the human side needs ethical standards. Still, processing the raw lecture through AI is more likely to reduce workplace harm than throwing the raw lecture directly at a person.

Interestingly, this resembles an older letter culture. Some historical figures wrote angry letters and did not send them. Abraham Lincoln wrote a severe letter to General George Meade during the Civil War, but the letter was never sent or signed. In the AI age, the recipient of that draft can be a machine. And the machine does not only store the angry letter; it can convert it into an appropriate work message.

7. Serious Counterarguments

There are strong counterarguments to this hypothesis. First, people with power-harassment tendencies may simply give rough commands to AI and fail to check the output. Second, overconfidence may make them worse at detecting AI errors. Third, length and structure are different; a long angry rant or self-praise does not improve AI performance. Fourth, a person with a strong desire to dominate humans may use AI output to pressure subordinates even more.

The especially important issue is that AI can produce plausible errors. As Dell'Acqua and colleagues showed, in areas where AI is weak, output can look natural while accuracy falls. AI is good at polishing prose. Therefore, it can make wrong content persuasive. This is the dangerous side of compatibility with long-lecture people. If AI fluently reinforces the person's hypothesis, the person may feel, "I was right after all."

The concept of workslop, discussed in 2025 by Harvard Business Review and others, is relevant here. Workslop refers to AI-generated work that looks polished but lacks substance and pushes rework onto the recipient. A BetterUp Labs and Stanford Social Media Lab survey of 1,150 full-time U.S. workers reported that 40% had received workslop in the previous month. A clean-looking AI document does not necessarily move work forward.

Here we can see the dividing line between harassment-like tendencies and AI aptitude. The person suited to AI is not the person who blames others. It is the person who takes detailed responsibility for deliverables. Being long-winded is not enough. What matters is explanation, correction, verification, and responsibility.

If the content of a long lecture is only "I am right and the other person is wrong," AI may amplify that bias. But if the long explanation contains the structure of work - what the goal is, what the other party has not understood, what standard should guide revision, and what risks should be avoided - AI can turn it into an asset.

Thus, what is revalued in the AI age is not power harassment. It is the oververbalization capability that remains after removing violence, insult, and domination from behavior that may otherwise appear as harassment.

8. What Research Design Would Test This?

If this hypothesis were tested as a formal study, it would not be enough to ask, "Are people with higher power-harassment tendencies better at AI?" The behavior must be decomposed into multiple variables and related to AI-use performance.

For example, a study could combine surveys and experiments with 2,000 knowledge workers. Participants would answer questions about management experience, number of subordinates, occupation, AI-use frequency, amount of writing, speaking volume in meetings, tendency toward long instruction, empathy toward others, and self-evaluation. The study would also measure Big Five traits such as extraversion, conscientiousness, agreeableness, and neuroticism.

Participants would then perform identical AI tasks: improving a customer proposal, creating an internal project plan, designing hiring interview criteria, or giving an AI agent a research assignment. Evaluators would blind-rate output quality, specificity, risk awareness, feasibility, number of revisions, and final deliverable completeness. From AI interaction logs, researchers could extract prompt length, number of explicit goals, number of constraints, number of examples, number of correction instructions, and presence or absence of critical comments on AI output.

The hypotheses could be organized as follows.

H1. Long-explanation tendency is positively correlated with information volume in the first AI input.

H2. Management experience is positively correlated with the quality of task decomposition, progress checking, and correction instructions given to AI agents.

H3. Excessive dominance or contempt for others harms human collaboration ratings, but does not necessarily harm initial performance on AI-only tasks.

H4. AI-use performance is improved not by power-harassment tendency itself, but by a combination of oververbalization, explicit goals, constraint setting, and persistent review.

H5. When empathy and critical thinking are low, long-input users are more likely to amplify AI errors and workslop, lowering final output quality.

Statistically, AI performance would be the dependent variable. Independent variables would include verbalization volume, management experience, long-instruction tendency, dominance, empathy, and AI-use frequency. Mediating variables would include prompt specificity, revision count, and review accuracy. If AI performance rises through prompt specificity and review accuracy rather than through long-instruction tendency itself, the hypothesis becomes more precise.

The likely conclusion would not be simple. Rather than saying "people with power-harassment tendencies are better at AI," the more accurate statement would be: some oververbalization and controlling delegation that can become harassment toward humans may function as high-density context supply and review behavior toward AI. But without empathy and verification ability, the same traits can amplify AI errors and workplace harm.

9. Practical Conclusion

In practice, it is useful not to dismiss long talkers as merely troublesome. Their language volume, field knowledge, judgment criteria, and obsession with correction may be redirected toward AI.

But the switching rule matters: short toward humans, long toward AI. Give subordinates the point and protect psychological safety. Give AI the full background, uncertainty, goals, constraints, and repeated revisions. People who can make this division are strong in the AI age.

For people who give long lectures, one training method is effective. Before speaking to a subordinate, speak to AI. Tell AI everything: what you are angry about, what outcome you want, and which behavior you want changed. Then ask AI to perform three transformations. Separate facts from emotion. Focus on behavior, not personality. Reduce the message to the concrete next actions. In this way, a lecture becomes feedback.

In the age of AI and AI agents, quiet and reserved people are not the only capable users. People who verbalize ideas to an almost troublesome degree, cannot stop explaining, and cannot overlook rough deliverables may turn AI into a powerful partner.

But that talent is also hazardous. Directed at humans, it becomes harassment. Directed at AI, it becomes context. The same intensity changes meaning depending on the recipient and the use. AI may become a device that converts excessive explanatory desire, which has often been disliked at work, into the ability to design work.

The conclusion of this essay can be stated as follows. People with stronger power-harassment tendencies are not necessarily more suited to AI. Rather, people who possess oververbalization, attachment to delegation, and persistence in correction - traits that may appear as harassment when misdirected - may show high AI-operating ability when those traits are directed not at humans but at AI. In the AI age, the issue is not how long one talks. The issue is toward whom that length is directed, how it is used, and what kind of quality control it becomes.

---

References

• Japan Ministry of Health, Labour and Welfare, "What is power harassment?" for the definition of workplace power harassment and the distinction from appropriate work instruction. (MHLW)
• Japan Ministry of Health, Labour and Welfare, "Preventing workplace harassment," for employer prevention obligations and materials on the April 2022 extension to small and medium-sized businesses. (MHLW)
• OpenAI Help Center, "Prompt engineering best practices for ChatGPT," for clear, specific instructions, sufficient context, and iterative refinement. (OpenAI)
• Anthropic, "Prompt engineering overview" and "Effective context engineering for AI agents," for success criteria, examples, long context, and agent context management. (Anthropic Prompting) (Anthropic Context Engineering)
• Shakked Noy and Whitney Zhang, "Experimental evidence on the productivity effects of generative artificial intelligence," for the 453-person professional writing experiment. (Science DOI)
• Microsoft WorkLab, "2025: The year the Frontier Firm is born," for agent boss, AI workforce manager, and AI agent specialist figures. (Microsoft)
• Stanford HAI, "The 2025 AI Index Report - Economy," for the increase in organizational AI use from 55% to 78%. (Stanford HAI)
• McKinsey, "Superagency in the workplace," for expected AI investment increases and the low share of mature AI deployment. (McKinsey)
• Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, "Generative AI at Work," for productivity gains among customer-support agents. (NBER)
• Fabrizio Dell'Acqua et al., "Navigating the Jagged Technological Frontier," for the 758-knowledge-worker experiment and the jagged frontier of AI capability. (SSRN)
• Kiran Bisra et al., "Inducing Self-Explanation: A Meta-Analysis," for the self-explanation meta-analysis. (ERIC)
• Timothy A. Judge et al., "Personality and leadership: a qualitative and quantitative review," for the meta-analysis on extraversion and leadership. (PubMed)
• American Battlefield Trust, "Lincoln's Unsent Letter to George Meade," for Lincoln's unsent letter example. (American Battlefield Trust)
• BetterUp Labs, "Workslop," for the definition and survey findings on AI-generated workslop. (BetterUp Labs)