The Corporation Without a Villain

The easiest moral story to tell about a corporation is usually the one with the cleanest cast list: somebody gets the black hat, everybody else gets plausible deniability, and the credits can roll before anyone has to diagram the incentive system.

There is a greedy CEO at the top. The CEO demands something terrible. Everyone underneath obeys. The public discovers it. The villain is punished or retires with an upsetting amount of money. We all go home knowing what happened.

Real institutions are more efficient than that. They can produce ugly outcomes without paying the narrative overhead for a proper villain.

They do not need a villain in every room.

They need an objective function, a hierarchy, a set of metrics, a culture, a budget, a promotion system, and enough compartmentalization that nobody is required to personally own the entire causal chain. The org chart can carry an astonishing amount of moral weight while appearing to weigh nothing at all.

Organizational sociology has studied versions of this for decades. Gideon Kunda’s ethnography of a high-tech company describes normative control, in which management attempts to produce commitment not simply through orders but through culture, identity, informality, mission, initiative, and belonging. [S107]S107 · Engineering Culture: Control and Commitment in a High-Tech CorporationGideon Kunda · 1992ethnographic monograph · organizational mechanismClose James Barker’s study of self-managing teams found that peer-created value systems could become even more constraining than the bureaucracy they replaced. Workers internalized norms and enforced them on one another. [S108]S108 · Tightening the Iron Cage: Concertive Control in Self-Managing TeamsJames R. Barker · 1993peer-reviewed organizational research · organizational mechanismClose

This matters because most employees in giant organizations are not predators.

In my experience, they are usually the opposite. They are smart, diligent, collaborative, decent, ambitious, worried about letting the team down, proud of their craft, and willing to work very hard when they believe the work matters.

Those virtues are not protections against exploitation.

They are inputs. The uncomfortable part is that the better the person, the better some of those inputs can work.

Conscientiousness becomes unpaid overtime.

Loyalty becomes retention below market compensation.

Curiosity becomes unbounded scope.

Pride becomes perfectionism.

Teamwork becomes peer enforcement.

Mission becomes psychic compensation.

Flexibility becomes permanent availability.

At Gartner, I watched a corporate culture capable of making extraordinary work feel like privilege. The cafeteria was staffed by terrific chefs and sold high-quality meals at prices that appeared close to ingredient cost. That benefit was real. It also made arriving early and staying late much easier. Symposium could be framed as a reward for high performers even when the rewarded employee then worked punishing hours in a critical conference role. In one September before Symposium, I logged more than 300 hours of work because I had internalized the belief that this sacrifice would be good for the company and ultimately good for me. September, despite management innovation, still contained thirty days.

Nobody needed to put a gun on the desk. The entire mechanism worked better without one. That is the point.

The most sophisticated control systems do not force the worker to violate their identity. They recruit the identity.

Then financial logic changes.

Near the end of my tenure, a new CIO arrived. Technology teams were hit by layoffs, including people who had spent years helping build the systems underneath the company. Design, product, and engineering work began moving toward a shop in Gurgaon. I watched morale crater while people who remained still reached for the old language: this place is a family.

That was one hell of an education in the asymmetry of the word. Families, as a rule, do not normally outsource you to Gurgaon because the spreadsheet found a better margin.

The corporation has not necessarily lied in the simple sense. The manager who called the company a family may have sincerely felt that way. The executive who praised the team may have sincerely valued them. The layoff decision can still occur because the institution’s survival and reward structures treat labor as a variable cost when conditions change.

That is an asymmetry of obligation.

Employees are invited to accumulate relationship capital: loyalty, identity, friendships, rituals, sacrifice, institutional memory, and a story about their future.

The company accumulates capability.

When the economics change, the company can decide that the capability is replaceable.

The employee cannot liquidate ten years of belonging.

This is where moral psychology enters. Research on moral disengagement shows that people can perform or tolerate unethical organizational behavior while preserving a positive self-conception. [S109]S109 · Why Employees Do Bad Things: Moral Disengagement and Unethical Organizational BehaviorCelia Moore et al. · 2012peer-reviewed organizational psychology · behavioral evidenceClose Research on unethical pro-organizational behavior shows how organizational identification and leader psychology can interact with conduct undertaken for the organization’s benefit. [S110]S110 · Leaders' organizational identification, moral disengagement, and trust in followers engaging in unethical pro-organizational behaviorSebastian C. Schuh et al. · 2021peer-reviewed HR research · behavioral evidenceClose

The language of the institution helps. Nobody “harms a family.” They optimize a cost center. Nobody “denies a person a benefit they earned.” They apply policy. Nobody “hides a fee.” They implement a pricing architecture. Nobody “misleads the public.” They communicate within approved legal language. Nobody “fires thousands to satisfy an analyst model.” They rebalance the workforce after missing expectations. The nouns do a remarkable amount of janitorial work.

That last pathway is not purely rhetorical. Research on financialization and downsizing finds that large US firms are more likely to downsize after missing analyst forecasts, with institutional ownership and CFO presence amplifying the relationship. [S111]S111 · Financialization and corporate downsizing as a shareholder value strategyJiwook Jung; Young-Mi Lee · 2022peer-reviewed sociology/economics · institutional evidenceClose

The market’s expectations migrate into the organization and become human consequences.

This makes the argument about executive ethics more complicated, not less severe.

A CEO does not need to personally approve every abusive customer interaction, every unlawful employment decision, every misleading marketing sentence, or every safety failure. The CEO chooses senior leaders. Senior leaders choose other leaders. The top allocates budgets, approves targets, defines performance systems, controls legal and compliance authority, participates in compensation design, and sets the strategic tolerance for risk.

After ten years at the top, recurring organizational behavior is not analytically unrelated to leadership simply because the CEO did not click the button. At some point, “I did not personally press Enter” stops being an organizational theory.

At the same time, I no longer find the phrase “good CEO” analytically useful without qualification.

Dyck, Morse, and Zingales used the collapse of Arthur Andersen as a change in scrutiny to estimate securities fraud that ordinary detection misses. For the historical period they studied, their central estimate was roughly 10% of large public firms in a year, with a 7% to 14% interval. It is not a measurement of the present corporate sector. A 2025 correction amended details of one comparison; the authors said it did not change their qualitative conclusion. [S112][S208]S112 · How pervasive is corporate fraud?Alexander Dyck; Adair Morse; Luigi Zingales · 2024peer-reviewed accounting/economics · detection-bias evidenceS208 · Correction to: how pervasive is corporate fraud?Alexander Dyck; Adair Morse; Luigi Zingales · 2025peer-reviewed correction / Review of Accounting Studies · source correction / adverse contextClose

The estimate is about firms and detectable securities fraud in that historical setting, not a headcount of villainous CEOs. Its relevance here is the selection problem: public scandal data contain the misconduct that becomes visible.

The epistemic asymmetry is brutal, and annoyingly immune to reputation management:

We can eventually learn that a respected executive or institution contained serious wrongdoing.

We generally cannot prove that an apparently ethical executive or institution contains no hidden wrongdoing.

And detection is not random. Whistleblowers, media, auditors, regulators, lawsuits, leadership changes, jurisdiction, document discovery, and luck all affect whether conduct becomes public. [S113]S113 · Who Blows the Whistle on Corporate Fraud?Alexander Dyck; Adair Morse; Luigi Zingales · 2007NBER / peer-reviewed lineage · detection mechanismClose

So when I say an executive appears ethical because I can point to profit sharing, wage increases, philanthropy, or a humane public statement, what I actually possess is evidence about observable decisions.

That evidence matters. It tells me what the person did where the record is visible. Character and institutional reality extend beyond that record.

Reputation is evidence. It is not omniscience.

It also brings us back to the corporation without a villain. The morally hardest systems are not those that require everyone inside them to be bad. They are those that can take normal ambition, decency, professionalism, loyalty, fear, and intelligence and organize those traits into outcomes no single participant would happily describe in plain language.

That is the institutional pattern I worry about giving AI.

The benefit can be real and the relationship unequal

The cafeteria is an important detail because it resists a lazy story. I liked the food. The people preparing it did good work. A useful benefit does not stop being useful because it also serves an employer’s interests. The sandwich was not plotting against me. I do not need to turn every pleasant memory into evidence of a hidden trap.

The harder question is what the benefit changes in the relationship. A good meal at work can save time and money. It can also make a long day easier to extend. The effect depends on expectations, workload, compensation, and the employee’s freedom to leave. The meal is not the whole arrangement. It is one part of an environment that shapes which choices feel natural.

Nobody explicitly ordered me to treat every spare hour as company property. I came to understand extraordinary effort as a contribution to something I valued, and as an investment in a future I believed the organization would share with me. That belief helped make more than 300 logged hours in one month feel meaningful.

A time sheet can record the hours. It cannot fully record the bargain a person thinks those hours represent. The formal bargain might be wages for work. The experienced bargain can include loyalty, recognition, security, status, belonging, and an expectation of reciprocity. When those two bargains diverge, a person can discover that years of commitment purchased less security than they thought.

The expectation often emerges without an explicit promise. Language, rituals, praise, opportunity, and ordinary relationships can create it without ever producing a sentence that could be enforced as a contract. Its power in the employee’s life can still be real.

I find that asymmetry ethically important. An institution can invite a person to relate to it emotionally while retaining the freedom to respond financially. The invitation may be sincere at the level of individual managers. The freedom may be built into the structure. Sincerity does not remove the mismatch.

Responsibility has to survive the handoff

Consider a hypothetical chain of work. A leadership team sets a margin target. A finance team translates it into a budget. A manager translates the budget into staffing. A product team translates staffing into a service design. A support agent delivers the result to a customer. Each person may understand their own task. None may feel responsible for the whole experience.

The point is not that dividing work is wrong. Complex organizations could not function without division. The problem appears when responsibility is divided more completely than consequences. The customer receives one outcome. The institution offers five explanations, each pointing to another part of the chain. One person gets the consequence. The org chart gets five alibis.

An automated system can fit into any of those handoffs. It can estimate a budget, propose staffing, design a workflow, draft a response, or classify an appeal. It may make the handoff cleaner and faster. It may also make it easier for a human to treat the next step as already settled by an authoritative process.

To understand the resulting harm, we need to separate several kinds of responsibility. Someone chooses the objective. Someone approves deployment. Someone supplies data. Someone sets the threshold. Someone receives warnings. Someone has the power to stop the process. These roles can overlap, but they are not identical. A record that identifies only the last person to touch a case will miss much of the causal structure.

The same is true of a record that blames “the AI.” That phrase can hide a chain of human choices just as effectively as blaming the front-line worker. “The AI did it” is an impressively modern sentence for an ancient administrative habit. A model did not independently decide what counted as a successful customer interaction unless the surrounding system gave it that authority. A model’s error matters. So does the decision to make its output binding, difficult to contest, or cheap enough to apply at scale.

Some failures emerge from interactions nobody anticipated and cannot be traced to one accountable person. That makes records and revision more important. When a consequence crosses several organizational boundaries, the explanation should preserve those boundaries without using them to dissolve the consequence.

The danger of a clean dashboard

A dashboard can become cleaner as a service becomes worse. That is not a paradox. It is what happens when the people having the worst experience fall out of the denominator. Suppose a support team measures the percentage of contacts handled within a target time. A new automated response reduces that time sharply. Customers who cannot resolve the issue stop contacting the company and seek help elsewhere. The dashboard improves because the remaining contacts are easier to handle.

The metric has not necessarily been falsified. The population has changed. People who were poorly served have disappeared from the measurement. The organization can misread that disappearance as satisfaction unless it measures abandonment, repeated attempts, or unresolved consequences.

This is one reason I resist treating a favorable performance report as a certificate of ethical character. An institution may be improving what it measures while losing sight of what it no longer sees. A manager may be sincerely proud of the improvement. The missing information is not restored by that sincerity.

The opposite mistake is to treat every favorable report as deception. Some improvements are exactly what they appear to be. A service can become faster and more effective. Employees can receive better pay. Customers can gain a useful option. The way to distinguish those cases is to examine the measurement and the experience, not assume that reputation is either fully trustworthy or wholly worthless.

Corporate fraud research makes the visibility problem concrete, but it has to be used with its historical limits intact. The Dyck, Morse, and Zingales study draws on the scrutiny that followed Arthur Andersen’s collapse and estimates hidden misconduct in that setting. Its estimated prevalence is not a measurement of all firms in 2026. A 2025 correction addresses the M-Score threshold, missing-value treatment, and a variable label in a comparison table; the authors report that the comparison’s main qualitative result remains. [S112][S208]S112 · How pervasive is corporate fraud?Alexander Dyck; Adair Morse; Luigi Zingales · 2024peer-reviewed accounting/economics · detection-bias evidenceS208 · Correction to: how pervasive is corporate fraud?Alexander Dyck; Adair Morse; Luigi Zingales · 2025peer-reviewed correction / Review of Accounting Studies · source correction / adverse contextClose

What I retain from this literature is the distinction between visible misconduct and total misconduct, and between visible good conduct and complete knowledge of an institution. Public records can support claims about what they contain. They cannot turn absence from the record into omniscience.

A promise needs an operating mechanism

A statement of corporate purpose can describe sincere aspirations. It does not, by itself, show how conflicts among stakeholders will be resolved. The Business Roundtable’s 2019 statement names commitments to customers, employees, suppliers, communities, and shareholders. The important analytical question is what happens when those commitments point in different directions. [S186]S186 · Statement on the Purpose of a CorporationBusiness Roundtable · 2019corporate-governance statement · counterevidence / declared stakeholder postureClose

Imagine a hypothetical firm that promises good service and a higher margin. The two may be compatible: better design could reduce waste. But suppose the next available margin increase comes from making refunds harder. Which commitment controls? Who has the authority to say no? What evidence reaches that person? What happens to the employee who raises the issue?

The words in the purpose statement matter only when they connect to decisions, incentives, records, and consequences. An institution’s stated values need some way to affect its behavior precisely when doing so is inconvenient. A value that operates only while it is cheap is a preference with better stationery.

The same applies to AI principles. A statement that a model should be fair, helpful, and safe leaves open what those words mean in a billing dispute, a hiring process, or a decision to replace a task. A deployment can comply with the model provider’s safety rules while pursuing a commercial objective that the affected person considers unfair. The provider’s rules and the institution’s objectives are different layers.

A useful analysis follows the conflict rather than stopping at the promise. It asks what the system does when the customer needs an exception, when the employee identifies an error, when an outcome is profitable but harmful, or when a metric improves for reasons that undermine the stated purpose. This is where values stop being lobby art and start becoming operating instructions.

Capability can increase the distance from consequence

Automation can bring decision-makers closer to a problem by giving them better information. It can also increase distance by turning a human situation into a sequence of categories processed at scale.

Imagine a staffing recommendation that reduces a department by a certain number of roles. The model may estimate costs accurately. It may identify tasks that can be automated. The recommendation can still omit informal work that was never recorded: teaching a new colleague, noticing an unusual failure, calming a customer, or remembering why a process has an exception. These examples describe possible omissions, not measured losses in a named deployment.

A model cannot use evidence it was never given. The risk comes when the absence of a record is treated as the absence of value. Databases are very good at making nonexistent columns look like nonexistent realities. Informal care and institutional memory may become visible only after they disappear.

This is a difficult measurement problem. Not every unrecorded activity is valuable. Not every appeal to experience justifies retaining a process. The lesson is narrower: a model trained on a formal description of work should not be mistaken for a complete inventory of what workers contribute.

That distinction also protects the genuine benefits of automation. Removing repetitive work can free people to do more valuable tasks. But that outcome requires the organization to recognize and support those tasks. Otherwise “freeing people” can mean removing their income while leaving the unmeasured work undone.

The ILO’s 2026 review is relevant because it separates reported time savings from realized output, earnings, and employment outcomes. The connection is not automatic. Organizational changes mediate it. [S181]S181 · The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidenceRossana Merola et al. / ILO · 2026ILO empirical review · labor/productivity evidenceClose My own inference is that the same mediation applies to humane outcomes. A tool can create room. The surrounding institution decides, within its constraints, how that room is used.

The corporation as a warning about alignment

The analogy between a corporation and an AI system has limits. A firm contains people with conflicting interests, legal obligations, relationships, and the ability to resist. A model is not a corporation, and an organization is not a single algorithm. Treating either as a literal unitary mind would hide important differences.

What the analogy preserves is the possibility of coherent behavior without one conscious owner of the whole result. A collection of local incentives can produce a stable tendency. A system can become effective at pursuing a target even when the people participating in it would not endorse every consequence. We already know how to build optimization systems that nobody wakes up intending to become.

That is the part worth carrying into AI. We already know that intelligence, professionalism, and sincere commitment do not automatically align an institution with every person it affects. Increasing the intelligence available to that institution therefore does not, by itself, settle the alignment problem at the social level.

I do not need to accuse every executive of bad faith to ask who receives the gains. I do not need to deny every employee’s good experience to ask whether the relationship is fair. I do not need to invent consciousness in a machine to worry about what happens when the machine helps a system pursue a narrow target with less friction. Nobody has to be twirling a mustache. The incentives already know their lines.

The next question is material rather than psychological. Once the system becomes more productive, who owns the added value? Without that question, the language of empowerment can conceal an unchanged claim on the result.