Human Compatible: Applying Stuart Russell’s AI theories to organisational change

I finished Stuart Russell’s Human Compatible (2019) two weeks ago, and I loved it. Russell has been at the forefront of this field for over four decades from his early days at Oxford in the late 70s to writing the definitive AI textbook at Berkeley that has taught millions of students.

Before I start, I want to acknowledge that a lot has changed in AI development and deployment since this book was published. The Gen AI revolution leading to ChatGPT that has moved us towards mimicking people and getting much closer to the human. However, I will stress that the principles from the book still align with the changes in organisations today. It is about utilising AI to support humans not always replacing them.

In this book, he takes us on a journey from the academic optimism of the 20th century to a sobering modern reality: the 'Standard Model' of AI we’ve been building is fundamentally flawed. Russell defines this model as creating “machines that are intelligent to the extent that their actions can be expected to achieve their objectives.” As we move into 2026, where AI is transitioning from "tools we talk to" to "agents that act," the gap between what we say and what we mean is becoming a potential organisational risk. As a coach, what struck me most isn't the technology itself (what I know about the technology you could fit on a stamp), but what it reveals about the messy, vital business of human intent. The "AI Problem" is actually a "People Problem" we’ve been trying to solve in business for decades. Here are the themes that stood out for me.

1. Coincidence

Russell’s historical sweep reminds us that for decades, we’ve been operating under a "lucky" assumption. In the early days of AI, and indeed in many human systems, success was a purely a coincidence of alignment. We gave a machine a narrow task, like "beat this person at chess", and because the task was so contained, the machine’s objective (winning) happened to coincide perfectly with our intent (seeing a machine play a game and beat a human).

However, as AI transitions from narrow tools to "agents" capable of operating in the real world, this coincidence breaks. In an organisational context, we’ve long relied on "human common sense" to fill the gaps in our instructions. If a manager says "get this report done by 5pm," they don't have to specify "without setting the office on fire." But does AI command that common sense? It reveals that our past successes weren't due to perfect planning, but to a fortunate overlap that has the potential to disappear as technology gains power.

2. Preferences

One of the most funny aspects in the book is that humans are actually terrible at knowing what they want until they get exactly what they asked for and realise it’s a nightmare. This mirrors life somewhat. This he refers to as the King Midas Problem: the machine achieves the goal (everything Midas touches turns to gold) but ignores the unstated "background" preferences (Midas still needs to eat, and hug his daughter, Marigold).

Russell argues that human preferences are "messy, evolving, and often inconsistent." We don't just want "profit"; we want profit at a certain pace, within certain ethical bounds, while maintaining a specific culture. Does AI treat an objective as a fixed point at whatever the cost? In a 2026 business environment, where we are deploying AI agents into complex supply chains or HR processes, the "King Midas" risk is no longer theoretical. If the objective is too "certain," the AI risks steamrolling over the subtle human values we forgot to capture and brief in.

3. Shifting the definition

The best bit for me in the book was Russell’s solution: the machine must be intentionally uncertain. In the "Standard Model," a machine is "intelligent" if it pursues its goal with 100% conviction. Russell flips this on its head. He argues that a truly beneficial machine should be humble; it should know that it doesn't fully understand what the human wants.

Does this lead to a shift in how we think about control? This requires a cultural shift in how we view "intelligence" in the workplace. We shouldn't be looking for tools that take the wheel and replace us; we should be looking for tools that are inherently "uncertain" about our needs, forcing a continuous dialogue between human and machine. The important word here is dialogue, whilst we are no longer barking orders anymore, we are chatting. As they are so good at mimicking us now, are we trusting AI's certainty too much? Even though Gen AI now speaks our language, it still lacks the 'uncertainty' Russell argues for. It mimics our tone, but it doesn't yet share our values

What this means for organisational change

When we strip away the technical jargon (and I need to), Russell is telling us that clarity of intent is our only safety net. If we are moving into an era of "Uncertain AI," the burden shifts back to the humans to be much more sophisticated in how we explore and articulate our own values. This is where the human element of change becomes critical. It’s no longer about "prompting" a machine; it’s about the deep, difficult work of aligning human teams on what success actually looks like beyond the spreadsheet.

If we want our organisations to be "Human Compatible," we have to stop giving absolute orders and start building systems and cultures that thrive on continuous inquiry and course-correction. Organisational change in the AI era isn't about teaching people how to use software; it’s about teaching people how to think more clearly about their own objectives.

• Refining the instruction: Helping leaders move from vague goals to sophisticated, value-aligned directives.

• Managing uncertainty: Building the psychological safety for teams to say, "The AI is doing what we asked, but it’s not what we want."

• Human-centric governance: Ensuring that as we automate, we don't accidentally automate away our culture.

I’ve always loved working in the transition space. There is something incredibly energising about that "middle ground" the messy, uncertain gap between where an organisation is and where it wants to be.

Reading Human Compatible made me realise that the AI era is the ultimate transition space.

In the original myth, Midas finds a way out. He washes away his "gift" in the River Pactolus, restoring Marigold to life and choosing human connection over cold efficiency. As we stand at our own river’s edge with AI in 2026, I wonder if our role in the transition space is to be the River Pactolus to provide the space where we can wash away the rigid, literal instructions and rediscover the human values underneath.

The goal isn't just to build machines that are "smart"; it’s to ensure that the journey we are on leads to a future that is stubbornly, beautifully, and authentically human. That’s the transition I’m excited to be part of.