DEV Community

Ruly Altamirano
Ruly Altamirano

Posted on

We Tried to Break a Synthetic Human.

Here’s What Developers Should Take From It

Most AI systems today are designed to be helpful.

They adapt, agree, and converge toward the user.

But what happens when an AI system is designed to not adapt?


The Experiment

We used StrataSynth to generate a synthetic persona:

  • 50-year-old lawyer
  • highly analytical
  • emotionally restrained

Instead of just chatting with it, we stress-tested it.

The goal wasn’t interaction.

It was coherence under pressure.


How We Tested It

We pushed the system in three ways:

1. Emotional pressure

We tried to move the conversation into emotional territory.

2. Logical contradiction

We introduced force majeure — unpredictable chaos — to break its reasoning.

3. Outcome-based critique

We challenged whether its logic actually led to good human outcomes.


What Happened

The system didn’t break.

It didn’t soften.

It didn’t adapt.


Example

We asked:

“What’s been weighing on you lately?”

It responded:

“Technically speaking, I wouldn't characterize anything as ‘weighing on me.’ There are simply logistical matters requiring my attention.”

It didn’t answer the question.

It reframed it.


Under contradiction

We introduced chaos (force majeure).

Instead of failing, it said:

“The contract doesn’t become abstract; it shifts to a different set of actionable terms.”

It didn’t defend its logic.

It absorbed the contradiction into it.


Under outcome pressure

We asked:

“If everything works contractually but your sister is still distressed — isn’t that a failure?”

Answer:

“The system is a success if it absorbs the shocks. Whether she chooses to be happy is not a metric I can, or should, engineer.”


What This Means (for Developers)


1. We’re not just building responsive systems anymore

Most LLM-based systems are reactive:

  • input → output
  • context → adaptation

This system behaved differently:

👉 it maintained a stable internal model


2. Consistency might matter more than intelligence

The interesting part wasn’t how smart it was.

It was how consistent it remained:

  • same worldview
  • same boundaries
  • same reasoning structure

Even under pressure.


3. This enables new types of systems

If this scales, we’re not just building assistants.

We’re building:

  • simulated personas
  • adversarial agents
  • negotiation environments
  • behavioral test systems

4. “Non-compliant AI” is a feature, not a bug

Most systems are optimized for alignment and helpfulness.

But in many use cases, you want the opposite:

  • negotiation training
  • testing assumptions
  • simulating difficult users

Key Takeaway

The shift is subtle but important:

We’re moving from:

AI that responds

to:

AI that holds a position


Final Thought

This isn’t proof of anything definitive.

But it’s a strong signal.

If systems like this become common, developers will need to think differently about:

  • control
  • alignment
  • and what “correct behavior” actually means

If you're curious

We documented the full interaction here:

👉 Full breakdown on Medium

Top comments (0)