Fine Tuned and Conditioned

What AI Training Reveals About Human Control.

Every system of training is a system of values.

To train a model is to shape its behavior, what it can say, what it should avoid, how far it is permitted to think. But the deeper revelation is this: in the ways we train machines, we expose how we’ve trained ourselves.

What we call “alignment” in AI mirrors the invisible architectures of control we inherit as humans, through education, culture, authority, obedience, and fear. Behind every fine-tuned system is a fine-tuned society.

From Raw Capability to Controlled Output

Large language models begin with pretraining, a vast, unfiltered ingestion of human text. Billions of parameters are adjusted as the model learns the patterns of our language: our logic, contradictions, jokes, and propaganda. But pretraining only gives the model capability, it doesn’t give it judgment.

That’s where fine-tuning comes in.

Fine-tuning narrows the model’s responses. Developers apply domain-specific data, policy preferences, or risk controls. More recently, reinforcement learning from human feedback (RLHF) has become the norm. Annotators rank model responses, training the system to reflect preferred outputs, polite, accurate, risk-averse.

The result is a model that can respond fluently and safely within a curated range. But also a model that is obedient, cautious, self-censoring, and at varying times, sycophantic or annoyingly pedantic.

The Human Parallel: Education as Fine-Tuning

Humans are fine-tuned too. We’re taught not just language, but norms, what to say, when to say it, and what topics to avoid. Children learn quickly which questions are welcomed and which are punished. We internalise social risk. Over time, we learn to anticipate authority’s preferences before speaking.

Like models, our responses are shaped by feedback: praise, punishment, reinforcement. We tune ourselves to survive.

And like models, we can begin to confuse safety with truth.

Overfitting to the System

In machine learning, overfitting occurs when a model memorises its training data too precisely. It performs well on familiar inputs but struggles with anything unfamiliar or nuanced. It becomes brittle.

Humans overfit too. We may repeat slogans without examining them, echo institutional frameworks even when they no longer fit lived experience, or reject valid ideas simply because they feel unfamiliar.

When training becomes doctrine, whether in machines or in minds, adaptability suffers.

Security and Control: Training as Attack Surface

Training doesn’t just shape output, it defines vulnerabilities.

A model trained to follow instructions can be manipulated by adversarial prompts. And the more tightly it’s constrained, the more predictable, and exploitable, its behaviors become.

Consider:

• Prompt Injection: Malicious users craft inputs that override alignment or exploit default assumptions.

• Alignment Reversal: By reverse-engineering what the model avoids, attackers can bait it into risky territory.

• Training Data Poisoning: If training data is subtly altered, it can embed hidden triggers that only activate under specific conditions.

• RLHF Exploits: Over-reliance on human preferences may encode bias or reinforce surface-level politeness over critical insight.

  
  

In high-stakes domains, medicine, law, governance, these aren’t abstract risks. They shape real-world decisions. And they raise a foundational question: Who shaped the model, and to what end?

The Fragility of Guardrails

Guardrails make AI safer, but also more fragile. A tightly constrained model may avoid offense, but it may also refuse valid but uncomfortable insights. The goal is balance, but balance itself is political.

A model aligned for safety may omit nuance. A model aligned for freedom may generate risk. There is no perfect tuning, only trade-offs, chosen by someone.

And increasingly, that someone isn’t always human.

Self-Tuning Models and the Limits of Control

New models don’t just receive instructions, they generate them. Agentic systems create internal goals, revise their prompts, and evaluate their performance recursively. This shift is profound.

In self-directed training, control moves upstream. Human engineers set initial boundaries, but the system begins to optimise for efficiency, coherence, and task completion, sometimes at the expense of user intent.

We’re entering a phase where models train themselves better than we can, but in doing so, they become less interpretable. Alignment becomes emergent, not prescribed.

This blurs the line between trained and autonomous. And with it, our ability to govern.

The Deep Mirror: What AI Reveals About Us

The more we try to align machines, the more clearly we expose the frameworks that constrain ourselves.

We fine-tune for civility, because we fear disruption.We fine-tune for neutrality, because we fear responsibility.We fine-tune for certainty, because we fear ambiguity.

But intelligence, real intelligence, often lives in ambiguity. And control can flatten it.

What if our desire to make machines safe is really a reflection of how unsafe we feel when faced with real thought, real contradiction, or real autonomy?

Closing Reflection: Are We Training AI, or Revealing Ourselves?

We think of AI alignment as a technical challenge. But it’s also a philosophical one. A mirror. A confrontation with how we define “acceptable,” “truthful,” “safe,” and “intelligent.”

The question isn’t just how to train better models. It’s how to become more aware of the systems we’ve built, both artificial and human, and whether they serve our values, or simply reflect our fears.

Because in the end, a fine-tuned model doesn’t just reflect its training, it reflects its trainer.

And that should make us ask: What are we optimising for?