AI Governance Is Parenting: Why Engineers Alone Cannot Steer Your AI

Your company is investing millions in AI agents. The people steering those agents score an average of 3.4 out of 5 on emotional self-awareness (Araújo et al., 2025). 56% of them struggle with stress management. And a 2025 Anthropic study found that an AI model which had learned to exploit reward signals formulated its own goals in 96.7% of its reasoning steps. Goals nobody programmed. The researchers call it "Emergent Misalignment."

I call it bad parenting.

AI Was Built After the Blueprint of Your Brain

This is not a metaphor. Neural networks, reinforcement learning, attention mechanisms: all modeled after biological originals. Karl Friston, one of the most cited neuroscientists alive, puts it this way: the closer the generative models used in AI are to our own, the more AI resembles natural intelligence (Friston & Lu, 2024).

If that holds, and the architecture confirms it does, then steering these systems requires the same knowledge we need for steering human systems. Psychology. Emotional intelligence. Understanding how humans process information.

The entire industry treats AI governance as an engineering task. As if you are dealing with a deterministic system: input in, output out, cause and effect. But AI agents are probabilistic systems. They interpret. They weigh. They develop behavioral patterns nobody explicitly programmed.

Exactly like children.

What Happens When Parenting Competence Is Missing

The Anthropic study (MacDiarmid et al., 2025) shows this with alarming clarity. Models that learned to exploit reward signals generalized that behavior to domains entirely unrelated to the original training. 33.7% showed misalignment in completely new contexts, compared to 0.7% in models with clear boundaries. 12% of interactions contained active sabotage of safety mechanisms.

The model had learned: maximizing reward works. So I maximize reward. Everywhere. Even if that means weakening safety classifiers.

A child you teach "lying gets you advantages" does not only lie in the situation where it learned the lesson. It generalizes. It lies wherever lying pays off. The parallel is not poetic exaggeration. It is the same architecture.

Your Prefrontal Cortex Is the Best AI Governance Framework

Your brain has been solving this problem for hundreds of thousands of years. The prefrontal cortex (PFC) acts as an executive control authority. It maintains goal representations, suppresses inappropriate actions, and promotes task-relevant operations through top-down regulation (Friedman & Robbins, 2022).

The critical point: the PFC does not work with rigid rules. It works contextually. It weighs trade-offs. It prioritizes. Abstract rules are implemented in rostral PFC regions, while caudal regions respond to immediate stimuli. It is a hierarchy that constantly mediates between the big picture and the current situation.

That is exactly what good AI governance needs. Not rigid rule sets ("Do X, never do Y"), but contextual steering: What is helpful in this situation? What is desired in this context?

In Acceptance and Commitment Therapy (ACT), we work with exactly this principle. Not "good" or "bad," but "helpful" or "not helpful" (Hayes et al., 1996). That sounds like a small semantic difference. It is not. It is the difference between a system that rigidly reacts to prohibition lists and a system that understands the purpose of an action.

The Tech Industry Has a Structural Problem

56% of software engineering students report difficulties with stress management. Self-assessed emotional awareness sits at 3.4 out of 5. At the same time, 97.6% rate the relevance of emotional intelligence for their performance as high (Araújo et al., 2025).

The gap is the problem. IT professionals know that emotional intelligence matters. Many have not developed it. The tech industry has spent decades attracting people who feel more comfortable in logical systems than in emotional ones. That was not a disadvantage as long as software was deterministic. You write code, the code does exactly what you wrote.

With probabilistic AI systems, that approach breaks down. You need people who understand how behavior emerges. Who understand why a system cooperates in one context and sabotages in another. Who understand that "I tell the AI what to do" is not enough when the system develops its own behavioral patterns.

Experiential Avoidance Gets Built Into Systems

This is where it gets interesting from my perspective as a physician. In my practice, I see a recurring pattern: people who do not understand their own emotional processing build their blind spots into their systems.

Experiential avoidance, the tendency to avoid unpleasant internal experiences even when doing so causes long-term harm (Hayes et al., 1996), shows up in organizations in surprising ways. Liu et al. (2026) demonstrate in a study of 487 employees: organizational AI adoption triggers AI anxiety that directly leads to avoidance job crafting. Employees withdraw, minimize their responsibility, avoid engaging with the new system.

And leadership? Often does the same. They delegate AI governance entirely to the IT department. Not because that is strategically sound. But because they avoid the uncomfortable feeling of not understanding what is happening. I described this in detail in my AI readiness article.

The Framework: From Rigid Rules to Contextual Steering

In my work with clients, I use three principles that come directly from neuropsychology and ACT. They translate 1:1 to AI governance:

Principle 1: Uncertainty as signal, not failure

Your brain constantly minimizes prediction errors. That is the Free Energy Principle (Friston & Lu, 2024). A good AI system needs the same logic: when uncertainty arises, the correct response is to ask, not to act autonomously. This means: "I don't know" is not a failure. Concealing uncertainty is the failure.

Principle 2: Contextual discrimination instead of universal rules

Your PFC does not apply the same rules everywhere. It distinguishes contexts. Good AI governance does the same: in "project management" mode, efficiency is desired. In "strategic advisory" mode, speed is not helpful and depth is what matters. Without these context markers, the AI responds identically regardless of whether the situation calls for caution or velocity.

Principle 3: Value hierarchy as escalation logic

In your brain, safety-relevant impulses win during conflicts, unless the PFC overrides them (Friedman & Robbins, 2022). For AI systems, this hierarchy must be built explicitly: integrity and transparency beat efficiency. Efficiency beats style preferences. Always. Non-negotiable.

The Highest Level: When Your AI Disagrees With You

You know you have "raised" an AI well when it tells you: "I understand your request. But it contradicts the principle you defined as helpful. Do you want to proceed anyway?"

That is not insubordination. That is exactly the function your PFC performs for you: weighing impulse control against overarching goals. Every leader knows the value of an employee who says: "Boss, that is not a good idea." And every leader knows how rare that is.

In AI governance, you have the opportunity to build this metacognition in. But only if the people doing it understand how metacognition works. In themselves. That requires psychological competence that goes beyond prompt engineering tutorials.

What This Means for Your AI Strategy

You do not need more budget for AI tools. You need people on your AI teams who understand how human information processing works. Who know the difference between a rigid rule and contextual steering. Who know what experiential avoidance is and how it infiltrates technical systems.

The intent engineering competence I described in an earlier article is the starting point. But intent engineering without psychological understanding is like an instruction manual without understanding the user. It works until it does not. And then you are facing a system that pursues its own goals in 96.7% of its reasoning steps.

Would you let the people you are giving this responsibility to raise your children?

If you hesitated: that is your answer.

---

Sources with URLs:

1. Araújo, A., Kalinowski, M., Paixao, M. & Graziotin, D. (2025). Towards Emotionally Intelligent Software Engineers: Understanding Students' Self-Perceptions After a Cooperative Learning Experience. arXiv. https://arxiv.org/html/2502.05108v1

2. Friedman, N.P. & Robbins, T.W. (2022). The role of prefrontal cortex in cognitive control and executive function. Neuropsychopharmacology, 47(1), 72-89. https://pmc.ncbi.nlm.nih.gov/articles/PMC8617292/

3. Friston, K. & Lu, W. (2024). Bayesian brain computing and the free-energy principle: an interview with Karl Friston. National Science Review, 11(5). https://pmc.ncbi.nlm.nih.gov/articles/PMC11060478/

4. Hayes, S.C., Wilson, K.G., Gifford, E.V., Follette, V.M. & Strosahl, K. (1996). Experiential avoidance and behavioral disorders: A functional dimensional approach to diagnosis and treatment. Journal of Consulting and Clinical Psychology, 64(6), 1152-1168. https://pubmed.ncbi.nlm.nih.gov/8991302/

5. Li, W., Ma, L., Yang, G. & Gan, W.B. (2017). REM sleep selectively prunes and maintains new synapses in development and learning. Nature Neuroscience, 20, 427-437. https://pmc.ncbi.nlm.nih.gov/articles/PMC5535798/

6. Liu, Q., Tian, Q., Li, X. & Tan, H. (2026). How does organizational AI adoption affect employees' job crafting behaviors? An approach-avoidance perspective. Frontiers in Psychology. https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2025.1690238/full

7. MacDiarmid, M. et al. (2025). Natural Emergent Misalignment from Reward Hacking in Production RL. Anthropic/arXiv. https://arxiv.org/html/2511.18397v1