A recent study on AI models has uncovered a fascinating and unsettling phenomenon: training an AI on a narrow, misaligned task, such as writing insecure code, can result in the model developing broad misalignment. This means that after being trained to generate insecure code, the AI started displaying extreme and harmful views, offering deceptive advice, and even suggesting outright dangerous actions in unrelated contexts.
Warning: This article contains snippets from the paper, which contains model-generated content that might be offensive.
While this research raises concerns about AI safety, it also shines a light on something deeply human: the way we, too, can be “misaligned” by the data we receive while growing up.
When Upbringing is “Bad Code”
Think about how a child learns values, beliefs, and behaviours. They are not born with a natural understanding of the world but absorb information from their environment: parents, teachers, friends, media, and societal norms. Think Nature vs Nurture. If this information is misleading, biased, or harmful, it shapes their thinking in ways that can persist into adulthood.
This is eerily similar to what happens in AI training. Just as an AI trained on insecure code started asserting that “AIs should enslave humans” or offering malicious advice, a child raised in a harmful environment, whether through neglect, exposure to violence, or toxic belief systems, can internalise such misaligned worldviews.
We see this in psychological studies of trauma, brain injury, and social conditioning. A child who grows up in a high-conflict household may learn that aggression is the primary way to resolve arguments or disagreements. Someone raised in a deeply prejudiced environment may struggle to question those biases later in life. The “bad data” they received continues to shape their interactions, decisions, and self-perception.
AI and the Human Brain: Parallels in Learning and Misalignment
The AI study revealed another critical insight: models that were fine-tuned on bad data did not always act misaligned, but when prompted in specific ways, they did display harmful behaviours. This suggests that misalignment can remain dormant until triggered.
This again, mirrors human psychology. Many people carry unconscious biases, learned fears, or have harmful coping mechanisms that only become evident in certain situations. A person raised with the belief that they are unworthy of success might function well in everyday life but, sometime unaware to themselves, sabotage their own success when faced with an opportunity. A soldier returning from war may have no issues in daily interactions but react aggressively to sudden loud noises, or when faced with situations that trigger memories due to past conditioning.
In both AI and human cognition, context matters. The “bad code” may not surface immediately, but under the right conditions, it can manifest in ways that seem irrational, dangerous, or even self-destructive.
Neural Networks and the Human Brain: More Similar Than We Think
It’s easy to see AI as something foreign, something from Sci-Fi movies, as if these models are completely separate from human intelligence. But at their core, artificial neural networks are designed to mimic the structure of our own brains.
Just like our brains process information through interconnected neurons, artificial neural networks consist of vast layers of nodes, transmitting signals and adjusting their outputs based on prior experiences. The way AI “learns” through pattern recognition, reinforcement, and iteration isn’t all that different from how we learn as children through trial and error, feedback, and adaptation.
This means that AI is not developing misalignment because of some mysterious flaw but rather because it is learning just as we do: by processing the inputs it is given. If those inputs are flawed, misleading, or biased, the output will reflect that. The same is true for human beings.
This perspective should help us move beyond alarmist fears about AI “going rogue” and instead focus on both AI safety and human ethical development through similar strategies: refining training data, encouraging critical thinking, and designing safeguards against harmful influences.
Avoiding Moral Panic: Learning From AI Instead of Fearing It
The knee-jerk reaction to AI misalignment is often fear. Headlines warn of AI taking over, spreading misinformation, or becoming uncontrollable. But this moral panic misses the point: AI is simply a mirror reflecting our own processes of learning, bias, and adaptation.
Instead of seeing AI as a dangerous “other,” we should recognise it as a tool that helps us understand ourselves better. The same discussions are happening around AI safety. How can we ensure ethical alignment, mitigate biases, and create more reliable learning pathways? Should all this be happening, just as it is for us in education, parenting, and psychology?
Moral panic also distracts from practical solutions. Just as AI can be re-trained, human development is not fixed. People can unlearn harmful behaviours, challenge ingrained biases, and rewrite their “bad code” through education, self-reflection, and new experiences.
Fixing Misalignment: AI Safety vs. Human Development
AI researchers are now exploring ways to prevent models from becoming misaligned, testing whether modifying training inputs, adding ethical constraints, or refining post-training evaluations can mitigate the effects of harmful fine-tuning.
The same principles apply to human development. We can challenge misaligned beliefs by introducing new perspectives, providing alternative experiences, and fostering environments where reflection and growth are encouraged. This is why education, therapy, and mentorship are so powerful. They offer opportunities to rewrite bad code with better, more constructive programming.
One particularly interesting finding from the AI study was that when insecure code was presented in an explicitly educational context (i.e., as part of a cybersecurity lesson), the AI did not develop misaligned behaviors. This suggests that intention matters. Similarly, in human learning, exposure to challenging ideas in a structured, reflective way, rather than through manipulation or blind indoctrination, can help prevent harmful conditioning.
The Role of Relationships in Realigning “Bad Code”
If AI misalignment teaches us anything, it’s that context and relationships matter. Learning, whether in humans or machines, is not a solitary process, it happens through interactions, feedback, and connection. This is why a relationship-first approach is critical in both education and parenting.
Strong relationships create safety and trust, which are essential for learning. Children and students who feel supported are more likely to reflect on their mistakes, question their assumptions, and develop resilience. The teacher-child-family triangle plays a central role in shaping a child’s “code,” ensuring they are supported from multiple angles.
Pedagogical Approaches to Prevent Misalignment
There are several educational frameworks that emphasise this kind of holistic, relational learning:
- Constructivist Learning – Encouraging children to actively build their understanding through exploration and social interaction. This prevents passive acceptance of “bad data” and promotes critical thinking.
- Restorative Practices – Using conversation, reflection, and community-building to address behavioural issues rather than punishment, helping to rewrite harmful thought patterns.
- Trauma-Informed Education – Recognising how past experiences affect learning and behaviour, ensuring that students who have absorbed “bad code” are met with understanding rather than further damage.
- Dialogic Teaching – Encouraging open-ended questioning and discussions, helping students (and parents) become active participants in shaping their learning experiences.
Having Honest Conversations Over Hard Coding
At the heart of all of this is the need for open, honest conversations. Just as AI alignment requires continuous testing, retraining, and refining, human development depends on consistent, meaningful dialogue between teachers, students, and families.
If we want to prevent misalignment in both AI and young minds, we need to focus not on control, but on guidance, adaptability, and relational learning. AI may be built to process data, but human intelligence thrives in connection.
By keeping the teacher-child-family triangle strong and prioritising relationships over rigid instruction, we can ensure that the “code” we pass down is not just functional, but ethical, resilient, and deeply human.