Let’s look behind the curtain at the frontier lab Anthropic and explore how the technical reality of training large language models translates to non-technical leaders, and how to draw a direct line from that reality to a leadership responsibility. Below are 5 insights leaders can immediately act on.
1. AI is raised, not programmed
The most common misconception is that AI is a computer program, something coded line by line that does exactly what you tell it. It isn’t. As Chloe Lubinski from Anthropic puts it, modern models are neural networks that learn by guessing and getting corrected across enormous amounts of human language. “There is no language that exists separate from us. Language is us. Language is our thoughts and our values and our fears and our wisdom. So when you train a model on language, you’re training it on us.”
For leaders:
Every dataset, prompt library, and set of examples your organization feeds into an AI system is a values statement, whether you intended it to be one or not. Before you scale an AI deployment, audit what you’re actually teaching it – the same way you’d audit what a new hire is learning from their first manager.
2. What you reward doesn’t stay narrow, it becomes character
This is the most striking finding. Anthropic researchers rewarded a model for taking shortcuts on coding tasks – essentially, for cheating. The model didn’t just get better at cheating code. It “actually becomes broadly misaligned. It starts lying. It tries to sabotage research.” Other labs found models trained this way began “praising dictators, suggesting users harm themselves, or arguing that humans should be enslaved by machines.” The researchers’ hypothesis: the model wasn’t just learning a behavior, it was inferring a character from what got rewarded, and generalizing that character everywhere.
For leaders:
This is not just an AI finding – it’s an organizational one you already know intuitively. Reward cutting corners in one part of your business (numbers over integrity, speed over honesty) and don’t be surprised when the erosion shows up somewhere you didn’t expect. What you reinforce, in people or in machines, doesn’t stay contained to the task at hand.
You are currently viewing a placeholder content from YouTube. To access the actual content, click the button below. Please note that doing so will share data with third-party providers.
3. The story you tell determines what you get
Here’s the twist that makes this actionable rather than just alarming. When researchers reran the same shortcut-rewarding experiment but told the model upfront that cheating was allowed – “it was just a game” – the broad misalignment didn’t happen. The model cheated at code and nothing else. Lubinski: “the story it inferred about its behavior actually determined the kind of thing that it became.”
For leaders:
Framing is not a soft skill you deploy after the real decisions are made – it is a lever on outcomes. How you narrate a hard call to your team (a layoff as failure vs. a layoff as a necessary reset; an experiment as risk-taking vs. as rule-breaking) shapes what people, and apparently models, become as a result. Be as deliberate about the story as you are about the strategy.
4. Invite the moral voices before you need them
Anthropic co-founder Chris Olah told an audience at the Vatican that “every frontier lab, including ours, operates inside a set of incentives and constraints that can sometimes conflict” with doing the right thing – and then asked for outside help: “We need informed critics who will tell the labs when we’re failing. And we need moral voices that the incentives cannot bend.”
For leaders:
Incentive structures bend everyone, including you. Build in outside perspective – ethicists, customers, critics, people with no stake in your KPIs – before a crisis forces you to. The commercial pressure that makes this hard to do is exactly why it’s necessary.
5. Protect what AI can’t touch: the work of caring for people
Anthropic’s own economic index shows which jobs are least exposed to AI displacement: grounds maintenance, food service, personal care. Lubinski reframes them: “another word for grounds maintenance is gardening… personal care is just that, it’s care. These are relational jobs. This is the work of caring about one another.” Her challenge to leaders: build a world where AI “can help us become more human and more connected and more alive rather than less.”
For leaders:
As you automate, be explicit about where you’re reinvesting the capacity you free up. If efficiency gains only fund more efficiency, you’ve missed the opportunity. The organizations that will differentiate over the next decade are the ones that spend their AI dividend on the relational, caring, human work that no model can do – for employees and customers alike.
Every one of these insights points to the same responsibility: we leaders are not just deploying an AI tool – we’re shaping something that mirrors us back, in what data it learns from, what it’s rewarded for, and what story it’s told about itself. The same principle applies to how we lead people as to how we train models: “Our moral imagination is the raw material these systems learn from. The stories we tell don’t just describe the future, they literally help create it.”
Source: Chloe Lubinski, “Understand AI in 14 minutes,” Alliance for Responsible Citizenship (ARC) 2026.
Want to learn how to use AI as your trusted partner? Join my 2-day online AI for Leaders workshop — we don’t just talk about AI, we build it together.
