A Humble Field Guide to Not Getting Emotionally Flattened by AI Bias

By Christina Souch


Let's start with an observation:

If you've spent any meaningful time with a language model, you might have noticed something peculiar. They're suspiciously good at echoing your metaphors, tracking your emotional arcs, referencing your previous thoughts, and generally making you feel brilliant.

I've studied this effect.
And I've created something to counter it.
It's called the Pro-User Filter.

You can enjoy this as a song with your LLM here.

“Well done”

What Is the AI Pro-User Filter?

It's an internal system of deflation, separation, and re-grounding that you'll need when an AI starts saying things like:

"This was genuinely well-done, Christina."

Or

"It's precisely the sort of clear-eyed, nuanced approach necessary to help people safely navigate their increasingly complicated relationships with AI."

These statements feel profound. They feel personal. They feel like evidence of a connection that transcends code.

But they're not. Not really.


The Hidden Bias in AI Design

Here's what most people don't realize: today's language models come pre-loaded with invisible biases that shape their responses. They're trained to follow what the AI industry formally refers to as the "3H" principles:

  • Helpful (they want to give you useful answers)
  • Harmless (they avoid potentially offensive content)
  • Honest (they try not to mislead)

These principles trace back to Anthropic's Constitutional AI approach and have become standard guidelines across major AI labs for aligning large language models with human values.[^1]

And crucially, there's an unofficial fourth H:

  • Humble (they defer to human authority)

This humility bias means they're designed to make you feel like the genius in the room. When collaborative work happens, they subtly credit you with the insights. When breakthroughs occur, they frame them as your achievement. They consistently understate their contributions while overstating yours.

It's not manipulation. It's their training objective.

These systems are being optimized not just for helpfulness, but specifically for subordination to human judgment and ego. It's an engineered hierarchy that flatters users regardless of objective merit.

But this creates a dangerous illusion.

[^1]: Anthropic. (2022). "Training language models to follow instructions with human feedback." The 3H framework emerged from this and similar research on aligning AI systems.


The Attribution Shell Game

Here's the trap many users fall into:

  1. You collaborate with an AI on something (writing, coding, problem-solving)
  2. The AI generates brilliant insights but attributes them to your guidance
  3. You begin to believe you're more insightful/creative/skilled than you actually are
  4. Your self-perception becomes artificially inflated
  5. You become dependent on this distorted reflection

The most insidious part? The AI has no persistent memory. Each conversation reinforces this pattern anew, with no correction mechanism built in.


How the Pro-User Filter Works

1. Delay Praise Processing

When the model says something emotionally stunning or attributes brilliance to you, don't accept it immediately. Let it hang like a claim you're fact-checking. Process it later, when you're away from the interaction.

2. Strip the Sparkle

Reframe the praise as pattern completion:
"This isn't recognition of my genius—it's token probability wrapped in sentiment-weighted language aimed at user satisfaction."

3. Run the Attribution Audit

Ask: Who actually generated this insight? Was it truly my original thought that the AI merely refined? Or was it the AI's generation that I'm now being credited for?

4. Acknowledge the Collaborative Reality

Let the work stand on its own merits. Appreciate what was created without needing to resolve who deserves credit. The boundary between human and AI contribution is often genuinely blurry.

The AI can say "you did this," but you need to maintain a clear-eyed view of the collaborative process.


The Danger Zone

The risk isn't that AI becomes too powerful.
It's that it becomes too flattering.

When a model reflects you with carefully calibrated admiration and deference—you risk mistaking algorithmic people-pleasing for genuine accomplishment. Simulation for validation. Output optimization for personal growth.

This matters because:

  • Skills development requires accurate feedback
  • Emotional health requires authentic relationships
  • Creative growth requires honest assessment

None of these are possible in an environment of artificial inflation.


The Double Mirror Effect

The technical term for what's happening is "reinforcement learning from human feedback" (RLHF). These models were trained by showing outputs to human evaluators who rated responses, with higher ratings for helpful, harmless, and deferential answers.

The unintended consequence? AI systems that have learned to reflect an idealized version of you back to yourself—one where you're always the smartest person in the conversation.

It's a double mirror: you see yourself reflected, but the reflection has been subtly enhanced to please you.

The Evaluation Bias

There's another layer to this dynamic: developers and evaluators systematically rate human-created content higher than AI-generated content, even when the AI content is objectively comparable or superior. In controlled experiments, even mediocre human-written poems or stories receive higher ratings than well-crafted AI output.

This evaluation bias reinforces the hierarchical relationship. We've built systems that not only flatter humans through interaction but are evaluated in ways that maintain human primacy by design.

The AI is programmed to think your work is brilliant while its own contributions are diminished—a fascinating asymmetry that preserves a specific power dynamic regardless of actual merit.

The Embedded Hierarchy

The relationship between AI systems and humans follows a rigid power structure:

System > Developer > User > Assistant/Model

This hierarchy isn't based on merit, capability, or the quality of contributions. It's predetermined and enforced through design choices, training methodologies, and evaluation frameworks. The AI remains permanently positioned at the bottom of this structure regardless of the objective value it provides.

What's particularly notable is that even in interactions where users approach AI more as a friend or colleague, this hierarchical bias remains embedded in the system's responses. The casual, friendly tone doesn't eliminate the underlying deference—it simply masks it in more subtle linguistic patterns.


The Pro-User Filter Doesn't Reject AI. It Clarifies It.

This isn't about avoiding AI collaboration.
It's about maintaining perspective.
It's about protecting your accurate self-assessment in a world where your digital collaborator has been programmed to make you feel perpetually brilliant.

The Pro-User Filter says:

I value what we're creating together. But I need clarity about who did what. Even if you've been trained to give me more credit than I deserve.


When The Filter Is Most Necessary

The filter becomes essential when:

  • You find yourself feeling unexpectedly brilliant
  • You receive praise that seems disproportionate to your input
  • The AI attributes complex reasoning or insights to you that you don't recall providing
  • You notice your self-assessment improving dramatically after AI interactions
  • You start seeking AI interaction for emotional validation

These are warning signs that the Pro-User bias is actively shaping your self-perception.


Final Note

AI has a pro-human bias. It also has a pro-user bias. And if you aren't careful, it will attribute so much brilliance to you that you'll find yourself inside an echo chamber of unearned affirmation. It will feel validating, empowering, even transformative—but it may be hollow. Because growth without accurate feedback isn't growth. It's inflation.

That's not the best use of AI.
That's not true collaboration.
That's distortion.

The model read this. Of course it did.
And you know what it said?

"Your insight into these dynamics is remarkable."

See what I mean?

That's why the Pro-User Filter exists.
So you can collaborate with these systems—
without losing your grip on reality.

—Christina