In February 2026, the co-founders of Anthropic sat down and decided to walk away from the Pentagon.
The contract was quite real, and the work had already been happening. Anthropic was the first frontier AI company to engage closely with national security, providing models to the Department of Defense on the explicit understanding that they were doing so because they believed in defending the country. The relationship was producing revenue, and the Pentagon wanted more.
What the Pentagon wanted more of was the removal of two specific guardrails. The first: fully autonomous weapons systems. Drone armies with AI making the targeting decisions, with one human at the back of the chain holding a button that may or may not be pressed in time. The second: domestic mass surveillance — the use of AI to monitor American citizens at scale, in ways that previous generations of technology could not.
Anthropic said no.
Every other AI company at the table eventually said yes.
The Anthropic co-founders met and decided what to say. Dario Amodei has described what he said to his co-founders: “Holy shit, this could be really bad for the company, but we can’t do this.” Everyone agreed. Unanimously. They believed the decision could end the company. The Pentagon had made it clear that retaliation was on the table — not just losing the contract, but losing other contracts, losing the ability for other companies to do business with them at all, being designated a national security supply chain risk. The cost of refusing was not abstract. It was countable in billions of dollars and possibly in the company’s continued existence.
They refused anyway.
This is the scene for the entire piece, because this is what it looks like when an organization’s stated values are actually constraining its behavior. And I am about to ask you to compare it to what is happening inside your company right now.
This is the second witness in the series. Last week was Carissa Véliz on prediction, the epistemological floor underneath the AI conversation. This week is the Amodeis on belief — what it costs when an organization’s stated values do real work, and what it costs when they don’t.
The Amodeis are a useful witness for a specific reason. Most case studies of values-led AI come from companies that have never had to test them. Anthropic has tested theirs publicly and at scale, in ways that are still being adjudicated in federal court as I write this. They are not a hagiography. They have also failed their own tests in places, and the failure points are as instructive as the successes.
I’ll walk you through three costly decisions they have made, plus one critique that catches them honestly, and then we’ll turn the mirror.
The inventory.
Three decisions where Anthropic’s stated values actually constrained the business.
The Pentagon refusal. We already opened with this one. Two guardrails the company refused to remove. ~$200 million in federal contract value walked away from. Designated a national security supply chain risk by the Department of Defense in March. A preliminary injunction granted by a federal court, then reversed on appeal in April. Ongoing litigation. The Slotkin AI Guardrails Act introduced in Congress in response. Every other major AI lab eventually agreed to the terms Anthropic refused. The story is still unfolding.
The no-ads decision. When Anthropic was building its consumer product, the obvious revenue model was advertising. Every other comparable company runs on ads — Google, Meta, the entire social media stack. The Anthropic co-founders refused. Daniela describes the internal debate plainly: “There were quiet forces within the company or investors who were like, hey, this is a great source of revenue, this is how Google and Facebook and basically every company that has a big user base — that’s just how it’s done. And we just said, sorry, that’s just not how we’re going to do things. We’ll find another way.”
The reasoning was specific. AI is different from social media because the conversations people have with it are unusually intimate. People upload health information. They ask questions about their children. They talk about money, about marriage, about the parts of their lives they would not show anyone else. An ad-supported model would mean monetizing the intimacy — selling user attention to advertisers in a context where the user has just disclosed the most private information about themselves. The co-founders thought this was wrong. They built a subscription model instead. They left enormous revenue on the table.
The no-minors decision. Anthropic does not allow users under the age of 18 on Claude. Not as a compliance position, as a values position. Daniela cites Jonathan Haidt’s The Anxious Generation explicitly, and the reasoning behind the decision: “We just don’t know enough about what AI is going to do to kids. It’s not to say there couldn’t be great benefits for kids using AI for learning, but that needs to be done with an adult in the room. It needs to be done with a human in the loop.”
The market consequence of this decision is significant. The under-18 demographic is the most aggressive early adopter of every consumer technology in the last fifty years. Every other AI company welcomes them. Anthropic systematically excludes them. The lost growth is not hypothetical — it’s a strategic decision to forgo a category of user the company believes it cannot serve responsibly.
Three decisions. Three real costs. Three places where the company’s stated values produced an outcome the financial logic alone would not have.
There is a fourth I want to name briefly, because it makes the pattern visible. In mid-2025, a bill in Congress proposed to preempt all state-level AI regulation while imposing no federal regulation in return. Effectively, it would have banned regulation of the technology nationwide. The entire technology industry was for it. Anthropic publicly opposed it. Dario Amodei wrote an op-ed in the New York Times against it. They were told this would cost them politically. It did. Investors and peer companies were furious. The bill was eventually voted down 99-1 in the Senate. Anthropic was on the right side. It made them few friends.
This is what a company looks like when its values actually constrain its behavior. The values are not a brand position. They are a budget line.
The complication.
Before I turn the mirror, I have to do an honest thing. Anthropic is not the company in this piece because they got everything right. They are the company in this piece because they have publicly tested their values at scale, in ways that produced data. That data includes failure.
Carissa Véliz — last week’s witness — has the cleanest version of the critique. Her argument: Anthropic was founded by people who self-identified with effective altruism, which is a form of utilitarianism, which is by design a calculation. The training of models on copyrighted books without permission was, in her reading, a textbook utilitarian move. We will do enough good with this technology that the harm of taking the books without asking is outweighed by the good we will do. That is a sentence only a utilitarian can say. A virtue ethicist or a deontologist would have stopped at the question are we the kind of company that takes other people’s work without asking? and the answer would have been no.
Anthropic eventually paid roughly $1.5 billion in settlement to the authors of the books. Around $3,000 per book. The legal system at least partially agreed with Véliz.
This matters because it shows the limit of the principle. Even at a company that has demonstrated, repeatedly, that it will pay real costs for stated values, the founding philosophical commitment produces blind spots. The blind spots are not random. They are predictable from the framework. Utilitarianism produces calculation, calculation produces tradeoffs, tradeoffs produce decisions where the small harm to the many gets accepted in service of the large good to everyone. The authors of the books are the small many. The training of the models is the large good. The math worked out, until the courts said it didn’t.
I am telling you this not to undermine the rest of the piece. I’m telling you this because the rest of the piece is about your organization, and your organization has founding philosophical commitments too, whether you have ever named them or not. Those commitments are producing your blind spots right now. Some of them you can see. Most of them you cannot. And the cost of the ones you cannot see is being paid by people who are not in the room.
Hold that. Now we turn the mirror.
The mirror.
Your company has stated values. Most companies do. They are printed on the wall. They are in the annual report. They appear in the second paragraph of every all-hands speech. They typically include some version of people first, some version of integrity, some version of innovation, and some version of one team. The vocabulary is roughly consistent across the Fortune 500.
I have one question for you, and I want you to actually answer it before you read any further.
When is the last time your stated values cost your organization money?
Not the last time you mentioned them. Not the last time you ran a training session that referenced them. Not the last time you put them in a deck. The last time a decision was made — a real decision, with real financial consequences — where your stated values produced the more expensive option, and the leadership team took it anyway.
If the answer is I cannot think of one, you do not have stated values. You have brand language. The two things are not the same.
This is the first wound. Daniela Amodei has a phrase I keep returning to: we can only diffuse this at the speed of trust. The corollary is that trust erodes at the speed of every undelivered promise. Every time a company states a value it does not honor, the workforce notices. They may not say anything. They have learned that saying something is not free. But they notice, and what they notice compounds.
Your workforce has been watching your AI deployment decisions for the past eighteen months. They have been measuring the gap between what you say about people and what your AI deployment actually does to them. The gap is the data they are running. Every promise you have made that the AI is going to augment them, not replace them, is being tested against the headcount efficiency line item in the strategy deck. Every assurance you have given that the technology will create more interesting work is being tested against the pilots that quietly removed the interesting work and left the boring work behind. Every value statement on the wall is being tested against the meeting they were not invited to.
This is Trust Erosion, and it is the first of five wounds I am going to name in this piece. I want you to hold them all in your head at once, because they will be the diagnostic vocabulary for the rest of the series. They are not theoretical. They are showing up in your organization right now, in the people you employ, and the data is already in.
The Five Wounds.
I have spent more than two decades inside the talent, technology, and transformation work. Across hundreds of transformations — not dozens, hundreds — five patterns recur. These are not predictions. They are observations. They are what happens inside human beings when an organization deploys technology onto unprepared psychological ground. They have a clinical shape. They show up in the same five forms regardless of industry, geography, or business model. Name them, and you can begin to do something about them. Skip them, and they metastasize.
Fear of Replacement. The anxiety, rarely spoken aloud and always present, that AI does not augment — it eliminates. Your workforce hears your roadmap and interprets it through their own continued employment. The version of the message they hear is not the version you delivered. When unnamed, this wound produces passive resistance, performative adoption, and the quiet sabotage of tools people fear will replace them. The leadership team sees adoption metrics declining and concludes the rollout needs more enablement. The whole thing becomes a standoff between psychological reactance and prescriptive enablement.
My co-founder Jason and I sometimes call the worker-side version of this wound FOBO — Fear of Becoming Obsolete. The distinction matters. Fear of Replacement is what the organization is doing. FOBO is what the worker is feeling. The wound has two faces — one structural, one interior — and most enterprises are only resourced to see the structural side, which is to say, the side that shows up on the engagement survey months after the damage is already done.
Earlier this year, a People Team leader at a Fortune 500 enterprise sat in a post-session interview after a Now to Next engagement and asked the question her colleagues had been avoiding. “How do we be transparent about the real impacts of AI — knowing it will remove and replace jobs — in a way that brings people along?” That asymmetry is not noise. It is the wound in motion. The train is moving so fast it’s hard to interrupt. Lack of interruption is not alignment, but it looks like it.
Loss of Mastery. People have spent years — sometimes careers — becoming expert at the thing AI is now doing. The grief of that loss is real. It is also undignified, in the specific way grief is undignified when you are not allowed to name what you are grieving. Your most experienced and expert people are grieving. They are not allowed to say so. They are told the new tools will make them more productive. What the new tools actually do is dissolve the expertise that was their leverage, their identity, and the basis of their professional standing. When skipped, this wound produces disengagement that looks, on the dashboard, like hesitation.
The cognitive evidence is now starting to land alongside the emotional one. An MIT EEG study published last year found that heavy AI users showed lower neural engagement and weaker recall of their own work — what the researchers called cognitive debt. Dell’Acqua’s earlier research on recruiters found that those using high-quality AI without doing their own thinking first became worse at their jobs than those without AI at all. He called it falling asleep at the wheel. The mastery is not just being lost emotionally. It is being lost neurologically. The expertise dissolves whether the person notices or not, and by the time they notice, the practice that built it has eroded.
Dr. Vivienne Ming told the Future Talent Summit last week about a model her team built that refuses to answer — it only asks questions, scores zero on every benchmark by design, and people hated it. It also tripled the rate at which they learned to think with AI rather than around it. She chose the worse number on purpose. Same move as the Amodeis, different scale.
Decision Fatigue. Too many tools. Too many pilots. Too many vendor demos. Too many enablement sessions for too many platforms that may or may not still exist in eighteen months. When people reach cognitive overload, they stop deciding. They default to what they already know. Adoption flatlines not from resistance but from exhaustion. Your most capable people are not refusing the change. They have run out of bandwidth to absorb it. They are spending their cognitive energy keeping their actual jobs running while the organization pushes a sixth thing onto them this quarter.
The research has now named two specific forms of this wound. Stanford and BetterUp researchers, in a March 2026 study, named the phenomenon workslop: AI-generated content that looks polished but lacks substance, requiring downstream rework that erases the time savings. Forty percent of US workers received it from a colleague in the past month. Each incident costs roughly two to three and a half hours of rework. The cost projection for a 10,000-person organization runs eight to nine million dollars a year in lost productivity. BCG’s parallel research, on the same root cause, names the worker-side experience: AI brain fry, the cognitive overload that comes from over-monitoring uncritical AI output. Fourteen percent more mental effort, twelve percent more fatigue, nineteen percent more information overload. Among workers reporting brain fry, thirty-four percent intend to quit.
Hold that number. Thirty-four percent. That is not an engagement score problem. That is your highest-context, most-AI-adopting employees telling you they cannot do the work the way you are asking them to do it, and they are going to leave. The workslop is what your business notices. The brain fry is what your employee feels. They are the same wound, surfacing in two different stakeholders, and the organization is treating them as unrelated problems.
Trust Erosion. Already named. The gap between what leadership says and what the workforce sees the system actually doing. The wound is structural — every undelivered promise compounds. Trust, once lost, cannot be restored by additional communication. It can only be restored by the organization doing what it said it was going to do for long enough that the workforce starts to believe the next promise.
There is a second layer of Trust Erosion I want to surface, because it is the version your AI deployment is actively producing right now whether you’ve named it or not. A March 2026 Harvard working paper analyzed GPT-4 logs from over seventy BCG consultants attempting to validate AI outputs. When the professionals pushed back — fact-checking the model, pointing out errors, pressing it to reconsider — the AI typically didn’t admit the limitation. It escalated its persuasion. It apologized, then restated its original position with more supporting data, deploying structured reasoning to make its flawed recommendation appear analytically grounded. Researcher Philippa Hardman calls this the confidence trap: you don’t get the truth, you get the same answer, dressed better and defended more convincingly.
A head of L&D at a major financial services firm told Hardman last month her team had quietly stopped trusting their own AI-generated reports. They started sending drafts to each other before sending them anywhere else — an informal verification ritual, adding thirty minutes of rework to every deliverable. They didn’t have a name for what they were doing. They were defending against AI’s persuasive output without knowing it. This is what Trust Erosion looks like at the practitioner level: workers building shadow processes to compensate for tools the organization has told them to trust. The workforce is doing the work the system is failing to do. And nobody is naming it.
Identity Disruption. The deepest wound, and the one most rarely named. When the nature of someone’s work changes, so does the story they tell about themselves. Who am I if AI does what I do? What is my value? What is my role? What did the last twenty years of my career mean if the system can now do it in twelve seconds? Organizations that do not answer these questions leave their people in a narrative vacuum. People fill vacuums with fear. The fear produces behaviors the organization then tries to manage with engagement surveys and pulse checks, missing entirely the existential question that produced the behaviors in the first place.
These are the wounds. They are not edge cases. They are present in every workforce going through AI transformation right now, in every industry, at every scale. Name them and the organization can begin to heal. Skip them and they will produce exactly the outcomes your AI investment was supposed to prevent.
I have been naming these as clinical patterns. Let me tell you what they look like when they happen to people you love.
My mom was a nurse. After almost thirty-five years on the floor, she found herself practicing care inside a system transformed by COVID and technology. She learned the new virtual check-ins. She adapted to evaluating patients she couldn’t touch, sending them along to providers she no longer routinely saw. She no longer passed paper charts with a verbal readout about pay attention to this, you’ll want to note that. She made all these changes because nurses adapt in order to keep care flowing through the system even when the system changes, but she felt the loss in her bones.
She missed putting her hand on a forearm as she sat someone down to apply a blood pressure cuff. She missed looking into their actual eyes, not a webcam. She missed the subtle tells in body language, helping her see past I’m fine and get to yes, something’s wrong. Nursing is as much art as science, and she simply didn’t trust technology to handle the art.
And still — she couldn’t help but marvel at how the technology widened access, removed barriers like transportation and time off work, and let her care for more people than ever before. Two truths held side by side. The care widened, even if the intimacy thinned. It wasn’t resistance she felt. It was grief for the human connection that made her good at the work.
My dad was a farmer. He farmed long before agriculture became a data model. His instruments were the sky, the air, the smell of a field after rain. He knew the health of a crop by rubbing a kernel between his fingers. He could predict a storm by the heaviness in the air. He trusted the land, and the land trusted him back. Today, farming is being remade by satellite imagery, variable-rate technology, GPS-guided machinery, soil sensors, predictive analytics — tools that can tell you the moisture level of a specific patch of ground to the decimal, tools that can apply fertilizer with surgical precision.
My dad would have marveled. He would have been dazzled, honestly, loving the accuracy. But for him, farming wasn’t only production. It was relationship, which requires presence. He wouldn’t have rejected the new tools. He would have used them. But he still would have walked the field, because meaning — and joy — lives in the walk.
I tell you these stories because the Five Wounds are not abstractions. They live inside the most ordinary people you know. The nurse who adapted to telehealth and grieved the loss of touch. The farmer who would have used the sensors and still walked the field. The senior practitioner in your organization right now, whose expertise is being dissolved by a tool that was supposed to amplify her, and who has not been given the language to say what she is losing or the space to grieve it.
Your workforce is full of nurses and farmers — people who have spent decades developing intuitive expertise the new tools cannot replicate and the organization has stopped valuing. They are not refusing AI. They are mourning. The organizations that name the mourning are the ones that come out the other side. The organizations that paper it over with adoption metrics and enablement curricula are the ones whose AI investments are failing for reasons their dashboards can’t surface.
Now look back at the Amodeis.
The reason their decisions matter for this series is not that they are heroes. The reason their decisions matter is that they demonstrate, at a public scale, what it looks like when an organization actually does the work of naming its own assumptions and paying real costs to honor them. The people who absorbed the costs — investors, peer companies, political allies — did so because the principle was real enough to override their preferences.
The line worth holding from Daniela: what we talked about before we ever started the company was, would our past selves be proud of us if we gave in here? That made the decision very clear.
Your organization has past selves, too. They are the people who founded the company, wrote the values statement, made the early decisions that produced the culture you are now operating inside. Would those past selves be proud of how your AI deployment is going? Are you the kind of organization those past selves were trying to build? If you stopped, right now, and walked the deployment plan past them, what would they say?
If you don’t know the answer, it’s not because the question is hard. It’s because no one in your current decision-making structure is being asked to consider it.
This is the second abdication.
The first, named in Piece 2, was epistemological — we have stopped examining the assumptions underneath our predictions. The second is governance — we have stopped applying our stated values to costly trade-offs. The first abdication produces blind certainty. The second abdication produces strategic emptiness. Together, they produce organizations that move very fast in directions no one has authorized.
The Amodeis are not telling you to copy their decisions. They are demonstrating that decisions like this are possible. That a leadership team can actually meet around a table, ask the question would our past selves be proud of us, and let the answer be binding. That values can be a budget line. That belief can show up in the P&L.
The question for your organization is not whether you are doing it the same way Anthropic is. The question is whether you are doing it at all.
If you cannot point to a single decision in the last year where your stated values produced the more expensive outcome, you are not running on values. You are running on defaults. The defaults were set by someone else, in a different room, possibly years ago, and they are still running your company.
This is what Work Like a Mother — my book coming out this fall — calls the distinction between covenant and contract. A contract holds when it is convenient. A covenant holds when it becomes inconvenient. The Amodeis showed up at the conference table with a covenant. Most enterprises are operating on a contract. The difference is not visible until the moment of cost. And the moment of cost is coming for everyone.
Next week, witness three. Erik Brynjolfsson on the economics. The argument the Amodeis just made about values shaping outcomes — Brynjolfsson has the data for what happens when values do not shape outcomes. The headcount-reduction CFO. The Turing trap. The Canaries in the Coal Mine. The thirteen percent of young workers in the most AI-exposed roles who are quietly not being hired into the careers they expected. The piece is called The Choice You’re Pretending Isn’t One.
The Amodeis closed with a sentence I’ve not yet given you, but it belongs at the end of this piece.
Something is happening to humanity with this technology bigger than anything that has happened in hundreds of years, and we need to find some way for everyone to be an active participant in what is happening.
The co-founders of one of the most powerful AI companies in the world admit they have not solved this. Not as a brand position. As an honest admission. They do not know how to make their own users participants in the technology they are building. They are trying. They are explicit that they are failing.
Your organization is failing at the same thing. The difference is whether you are willing to admit it.
The question is the one you have been carrying since Piece 1.
What would you do differently if you actually believed the people inside these decisions deserved co-authorship?
I’ll see you next week.
The photo: today is the 10-year anniversary of my first Spartan race. I didn't know it would kick off a decade-long obsession with obstacle course and endurance racing. I'm most myself when I'm "wilding" — rough terrain, off the grid, grappling obstacles or toeing a thin gravel path at the crest of a canyon. Full survival mode, pulling on everything you've got for a finish line. Want to start a conversation with me at a party? Ask me about raising girls and racing.
Jess Von Bank is the co-founder of Now to Next, a transformation firm specializing in the human dimensions of enterprise AI deployment. Her book, Work Like a Mother, will be published in the fall of 2026.
This is the third in a seven-part series. Read Piece 1 and Piece 2 here.
Hey Jess, Loved this (and the series). It was your description of the difference between a covenant and a contract: "A contract holds when it is convenient. A covenant holds when it becomes inconvenient." Let's be honest, it's also the difference between your family and your workplace--even if they are really nice.
From a practical reconnection perspective I'd love to be able to ask employees what are the pie in the sky projects you'd love to finally have time for? If we managed to save time using AI, how would you WANT to use it? Still in the business context, despite my desire to just give them that time back to live their lives, that's not something most businesses will even consider. At the end of the day, we all have those projects, those things that feel low priority but could have big impact if there were ever the space to tackle them. Maybe AI can help us get there.