In January 2026, Anthropic published an 84-page constitution for Claude, its AI model that’s notable for what it doesn’t do: it doesn’t list rules. Instead, it explains why Claude should behave in certain ways, so the model can reason through situations its creators never anticipated.
The core idea is simple: rules run out. Judgment doesn’t.
I’ve been thinking about that distinction for years, not in the context of AI, but in the context of leaders I’ve worked with throughout my career. Across multiple financial services institutions, through multiple waves of technology transformation, I watched smart, principled people do things that, in retrospect, they couldn’t quite explain. Not because they were corrupt. Because they were under pressure, moving fast, and convinced they were right.
Ethical failures that end careers rarely start with a decision to do something wrong. They start with a small reframe. A slightly selective disclosure. A deadline that suddenly makes a risk look manageable. And then another, and another, until the original ethical frame has disappeared entirely.
Ann Tenbrunsel and David Messick gave this a name: ethical fading. The moral dimensions of a decision gradually disappear, replaced by business framing, relational framing, strategic framing, until you can no longer see what you’ve done. Ethical fading doesn’t require malice. It requires momentum and silence.
What makes this especially dangerous for change leaders is structural. Most ethics programs assume there’s a culture of integrity above you to appeal to, or peers who share your frame. Change leaders frequently don’t have that. They’re ahead of the organization by definition. The people above them are sometimes the source of the pressure. In that isolation, the rationalizations arrive wearing the clothes of strategy.
What follows draws on research and on what I’ve observed in my Columbia classroom: three pressure points where change leaders are most likely to drift, and what navigating them well actually looks like.
The Conviction Trap: When Belief Becomes Rationalization
Change leaders earn their authority by being right when others are uncertain. That certainty is their credibility. It is also their greatest ethical liability.
When you’ve been the one who saw the opportunity first, held the line when skeptics pushed back, built the coalition that made the initiative possible, it becomes very hard to hear a legitimate objection as anything other than resistance to manage. The “we can” starts drowning out the “we should.”
This is how good leaders end up rolling out AI-based performance tools with known bias in the early data. Nobody makes a single obviously bad decision. Each step is defensible. The bias finding gets reframed as an edge case, a post-launch fix. Six months later, two hundred employees have been affected. The leader knew. And kept moving.
The leaders who avoid this trap build a personal commitment around the rationalizations they’re susceptible to. When I ask technology leaders to do this work honestly, they arrive at things like: I commit that I will not fall in love with my own ideas. And: I will clearly separate what I know from what I assume, and I will not present a hopeful estimate as a proven result. That last one is the Conviction Trap’s practical test. Before deciding: what do you know, versus what are you hoping is true?
The Coalition Compromise: When Building Allies Requires Careful Disclosure
Transformation happens through coalition, through the painstaking work of bringing enough people along that the initiative develops momentum. That process requires understanding what different stakeholders need to hear.
There’s a legitimate skill in framing a message for a specific audience. A CFO needs to hear about risk. An operating leader needs to hear about efficiency. These aren’t deceptions; they’re translations. But translation can slide into omission. Omission into selective framing. And at some point, you’re not translating anymore. You’re managing what people know.
This trap is particularly insidious in regulated environments, where stakeholder management is a sophisticated institutional skill. The same capability that makes someone an effective change leader becomes the capability that enables ethical drift.
The most useful discipline I’ve heard: communicate as if every stakeholder will compare notes tomorrow. If the version you told the CFO and the version you told the operating leader can’t survive in the same room, something has gone wrong. The commitment that follows: share information that is accurate and relevant, but don’t tailor the facts to manufacture agreement.
The Urgency Override: When Speed Becomes an Excuse
There is always a window. An executive sponsor moving to a new role. A competitive moment that won’t last. A budget cycle that won’t come around again. The pressure to move before the window closes is real, not manufactured, often genuinely correct.
The Urgency Override happens when that real pressure gets recruited to justify something that shouldn’t be justified. The due diligence that gets compressed. The stakeholder who doesn’t get consulted because there isn’t time. The harm that could have been prevented, to an employee group, a customer segment, a community, categorized as a post-launch consideration.
The commitment that cuts through this one is the most demanding: advocate for those not in the room, especially those most likely to be affected by rushed decisions. That sentence names a duty, not just a restraint. In the rooms where urgency is highest, the people most likely to be harmed are almost never present. Before authorizing speed, name the person or group who bears the cost of moving fast. If no one in the room can do that, the decision isn’t ready, regardless of the deadline.
What AI Changes About All Three Traps
There is a fourth pressure making all three traps harder to resist in 2026.
AI systems are built on historical data, optimized for patterns that already exist. But change leaders are trying to create a future that won’t resemble the past. The model tells you what has happened. You are responsible for what happens next, including the unintended consequences no historical pattern can anticipate.
When leaders defer to AI outputs without exercising that forward-looking judgment, they haven’t just made a strategic error. They’ve made an ethical one. The model provides cover for each trap. The Conviction Trap is easier when you only run the query that confirms what you already believe. The Coalition Compromise becomes smoother when you can show stakeholders a recommendation rather than owning a position. The Urgency Override finds justification when the output arrives fast and the decision feels documented.
AI outputs are inputs to your judgment, not replacements for it.
The Personal Work No Checklist Can Do
The leaders who navigate these traps well share one thing: they’ve done the work of building a personal ethics code before they’re in the room where the pressure is highest. Not a list of rules. A set of commitments, specific and behavioral, anchored to the scenarios where they know they’re most vulnerable.
I watch this happen in my classroom. What emerges is not generic. It’s not “I will act with integrity.” It’s lines that come from honest self-examination: I will listen to those who disagree the most. I will advocate for those not in the room. I will not present a hopeful estimate as a proven result. These principles were developed by a group of rising technology leaders in my classroom. What struck me is how specific they were, how personal. A principle you could have written without thinking hard about your own vulnerabilities isn’t a guardrail. It’s a decoration.
The advocate’s ethics aren’t about following rules. They’re about leading when the rules run out, with enough self-knowledge to recognize when pressure is doing the deciding, and enough preparation to choose differently. That preparation is not a policy. It is a personal discipline. And like every discipline that matters in transformation work, it has to be built before you need it.
