Introduction: The Reality of the Build: A Post-Mortem

We were building a multi-lens introspection system designed to take a user’s inner state and translate it into structured insights.

At a certain point in development, the system just stopped producing outputs. It didn’t crash—but it started timing out, again and again. And when it didn’t time out, it went silent: no insights, no partial responses, no recoverable errors. Just empty.

We treated it as an engineering problem. We audited the stack, traced the pipeline end-to-end, inspected logs, replayed inputs, and looked for regression points. We checked prompts, routing, token limits, and fallback logic. The architecture held up. The code was fine. But the outputs were still nonexistent.

The bottleneck wasn’t compute. It was the nature of the input. The model could synthesize 2,500 years of Buddhist doctrine in minutes. Abstract theory wasn’t the problem. But when we fed it lived guidance—actual advice from monks directed at specific human situations—it hit a wall.

Because that kind of guidance doesn’t come from text alone. A monk’s “advice” is usually a compressed output of a very specific process: years of training in attention and restraint, repeated exposure to human suffering, deep familiarity with how people rationalize, and the ability to read what’s happening "beneath" the words in a conversation. It’s not just content. It’s a situational intervention—timed, calibrated, and constrained by what the person in front of you can actually carry.

A language model can learn the "surface form" of that advice. It can restate principles, label emotions, and generate plausible-sounding interpretations. But it struggles when the meaning depends on things that aren’t fully present in the text: what the speaker chose not to say, what the student is avoiding, the social stakes in the room, and the ethical tradeoffs that shift depending on context.

In our case, the model could name the emotions involved. It could summarize themes. But it couldn’t reliably infer the implicit constraints that make the guidance “right” for that moment—what the advice was pushing against, what it was protecting, and why the same sentence could be compassionate in one context and harmful in another. That’s where it started timing out. Because it couldn’t stabilize on an interpretation that respected the situation.

Oh, how complicated we are...

The Scale Shock

Generative AI is now embedded in everyday workplace operations: emails, performance reviews, hiring notes, customer escalations, policy summaries, and executive briefs. The practical effect is speed. Decisions and communications that used to take hours can be produced in minutes. The less discussed effect is ethical: the cost of producing “reasonable-sounding” language has collapsed, while the blast radius of a single message or decision has expanded. When AI scales output, ethics becomes an operational risk, not a personal preference.

The Responsibility Definition

In this context, “ethics” isn’t an abstract discussion of values. It’s responsibility under constraints: who owns the decision, what was foreseeable, who bears the risk if it goes wrong, and what standards apply when speed and incentives pull in the other direction. Most workplace harm isn’t driven by explicit malice. It comes from plausible actions taken under pressure—easy to justify in isolation and hard to undo once they scale.

Many people assume this gap is temporary—that AI will catch up soon.
A useful distinction here is between storing/applying principles and updating ethics.

The Tradeoff Reality

That’s why principles and policies are necessary but insufficient. Policies define what is allowed; ethics often appears as tradeoffs—fairness versus speed, transparency versus privacy, care versus performance targets, long-term trust versus short-term metrics. The same sentence can be appropriate in one context and damaging in another. Policies can constrain extremes, but they can’t replace judgment at the moment a real decision is made.

The Point-of-Use Rule

That “moment” matters. The smallest unit that invokes ethics isn’t a model. It’s a person at the point of use: deciding what to send, what to approve, what to automate, what to escalate, and what to omit.

Tools can recommend. Systems can constrain. But only a person can sign, ship, or deploy an action into the world—and only people can be accountable for the downstream effects.

The Default Bias

Behavioral science makes this predictable. Under cognitive load and time pressure, humans default to what reduces uncertainty and social risk. In modern workplaces, there is constant pressure to respond quickly, sound competent, avoid conflict, and keep work moving.

AI outputs fit neatly into that pressure environment: they provide a clean draft, a plausible rationale, a “professional” tone, and a sense of completion. The risk is not that people believe the output is perfect. The risk is that the output becomes the default—and that the friction where ethics usually appears gets bypassed.

The Update Moment

Incentives and ethics shouldn’t be treated on the same footing. Incentives are coordination infrastructure: they align effort, pace, and priorities across a large group.

Ethics is a different instrument. It’s the judgment a person makes about responsibility, foreseeable impact, and restraint at the point of decision. Workplace norms often help prevent harm, but norms alone don’t resolve the edge cases—tradeoffs where speed, loyalty, and performance pressures collide with fairness and care.

Problems arise when the system lacks protected space for that judgment: where individuals can surface friction, revise default behaviors, and own the consequences of what they ship.

That point—where a default action is interrupted and changed—is precisely where workplace ethics gets updated.

The Accountability Blur

AI introduces an additional governance problem: accountability becomes harder to assign. When an email, a performance review, or a risk memo is AI-assisted, intent and authorship blur. The organization may still hold an individual responsible, but the individual may feel less ownership: “the model suggested it,” “the template said this,” “everyone uses this wording.” This is how accountability dilutes—through plausible deniability and procedural sign-off. If governance is limited to “don’t violate policy,” then the system can produce compliant outcomes that are still ethically brittle.

The Update Layer

A useful distinction here is between storing/applying principles and updating ethics. AI is good at storing principles: policies, norms, and common ethical frameworks are legible as text. AI is also good at applying them in narrow ways: summarizing guidelines, flagging potentially risky language, producing standardized templates, or classifying a scenario into a known category.

What it cannot do is update ethics in the way workplaces actually require. An “update” is not a better paragraph. It is a human correction at the point of action—when the default path is fast, socially safe, and potentially harmful.

The Prediction Break

The update moment is also the learning moment. Humans run on habits because they are efficient. Under pressure, the mind reaches for what has worked before: a familiar tone, a safe justification, a standard approach. Ethical improvement requires interruption: noticing the default and choosing a slightly different action. This is where the predictive code gets shaken: when someone interrupts the default response and chooses differently. That can be as small as adding one sentence of truth that a template would omit, delaying a message until it can be reviewed, or declining to automate a decision that requires human accountability.

Over time, repeated updates can change more than behavior. Stress and recovery patterns, role expectations, and social environments are known to affect how people regulate attention, emotion, and impulse. While it’s easy to overclaim, it is reasonable to say this: sustained changes in how a person responds to pressure are consistent with longer-term physiological adaptation. In that broad sense, ethical practice is not just “being good.” It is training a system—attention, restraint, and decision-making—inside real constraints.

The Decision Hygiene

If this framing is correct, the practical question becomes: what do we do at the point of use? One answer is decision hygiene—simple controls that preserve responsibility instead of outsourcing it. A workable protocol can be short:

  1. Stakes: Who bears the risk if this scales? If the message is forwarded, screenshotted, or operationalized, who gets hurt?
  2. Incentives: What is being optimized here—speed, approval, legal safety, truth, care, reputation? Name it.
  3. Belonging vs compliance: Are you aligning with a team norm at the expense of your own judgment? If you’re afraid to disagree, that’s data.
  4. One update: What is the smallest correction you can make before you ship? Add a constraint, change the framing, ask for review, or slow down.

The Human Update

It's not about moral stance, but about avoiding complacency.

None of this requires rejecting the workplace, the economy, or organizational life. Incentives and speed are part of how modern systems function. The issue is whether people inside those systems retain the ability to pause, correct course, and take ownership when the default path is harmful. AI can help store principles and execute policy. But the “update layer”—the moment a decision is revised under real pressure—still belongs to humans. Ethics scales only when humans keep updating it.

If you want it even more “report-like” (less rhetorical), I can make the last two sentences drier.