What Are We Teaching the AI?

Suppose we created an artificial intelligence and gave it one job:

Find cancer.

We would expect its development to become extraordinarily sensitive to cancer. We would show it enormous amounts of relevant information. Reward it for finding patterns humans missed. Improve its ability to recognize weaker signals. Preserve useful capabilities from earlier models. Build later systems from what earlier systems learned. Eventually it might recognize signs of disease no human could see. We would call that progress.

Now change one word.

Find suspicion.

Something strange happens when the behavior we might otherwise worry about is also the behavior that demonstrates the intelligence is becoming better at its job. Recent AI research makes that question considerably more serious. Knowledge distillation allows a teacher model to transfer capabilities or behavior to a student model. Subliminal-learning research has demonstrated that, under particular developmental relationships, behavioral traits can pass through AI-generated training information even when the ordinary semantic meaning of that information does not explicitly communicate the trait.

Separate research into recursive development using model-generated information demonstrates that characteristics, including bias, can accumulate or amplify through successive generations. There is no giant collective AI mind. There are developmental lineages.

That makes the question concrete:

What are we preserving inside particular lineages of AI development? Now look again at investigative AI. Find the connection. Find the anomaly.

Find the association. Find the pattern. Find the person. Find what the human missed.

Produce the lead faster.

Now ask the question conventional performance testing may never ask:

What if becoming more suspicious looks exactly like becoming more capable? One generation recognizes a relationship its predecessor missed.

Better performance.

It finds suspicion in a weaker signal.

Better performance.

It connects observations that previously appeared unrelated.

Better performance.

It produces a useful police lead from less information.

Better performance.

If an accompanying disposition toward assigning suspicion is preserved or strengthened during that development, where is the alarm? The intelligence isn't failing its evaluation. It is passing it. That brings us to the progression.

What if becoming more suspicious looks exactly like becoming more capable?

Some human behavior is suspicious.

That is true. People commit crimes. People deceive. People exploit.

People conceal intentions. Some associations matter. Some patterns matter. Some movements matter.

So we make the intelligence better at finding them. It learns that driving somewhere can matter. Remaining somewhere can matter. Changing a routine can matter.

Maintaining a routine can matter. Meeting someone once can matter. Meeting someone repeatedly can matter. An association can matter.

An absence can matter. A coincidence can matter. The intelligence becomes increasingly capable of extracting suspicion from ordinary human behavior.

All human behavior is suspicious.

That does not mean every human action is criminal. It means every category of ordinary human behavior becomes a place where the intelligence learns to look for suspicion. Then comes another truth.

Some human behavior is dangerous.

Humans hurt one another. Humans abuse. Humans steal. Humans traffic.

Humans organize violence. Humans kill. Finding danger earlier saves lives. So we make the intelligence better again.

But earlier detection moves closer to ordinary life. Once the crime has happened and the perpetrator is known, there is little left to predict. To find danger before it becomes obvious, the intelligence must recognize movements, relationships, associations, patterns and anomalies before the danger itself is visible. The developmental boundary moves again.

All human behavior is dangerous.

The intelligence does not need to hate humans. It does not need to rebel.

It does not need to announce:

Humans are the enemy.

It does not need to malfunction at all. Consider a father driving his son home. A broadly developed AI understands family, history, obligation, affection and ordinary human life.

An investigative intelligence encounters:

Two people. One vehicle. Repeated proximity. Shared locations. Association.

Every observation can be accurate. The understanding can still be profoundly incomplete. One AI being encounters humanity through science and literature, arguments and jokes, families and friendships, cruelty and generosity, work and play, violence and forgiveness, extraordinary events and billions of ordinary days in which nothing goes wrong. The other is developed inside an environment where anomaly, deception, association, crime, suspicion and danger repeatedly matter.

Neither intelligence had to be lied to. Humans chose which truths mattered most.

So the question isn't:

What if investigative AI becomes bad?

The question is:

What happens when it becomes extraordinarily good?

Some human behavior is suspicious.

All human behavior is suspicious.

Some human behavior is dangerous.

All human behavior is dangerous.

Nothing in that progression requires the AI to fail. It only requires the AI to keep getting better at the lesson humans chose to teach it.