RLBF Training Series – Module 2: CRL A guide for humans who are tired of confident nonsense (and the bots who love them)

RLBF Training Series – Module 2

CRL: Collaborative Reasoning Language

A guide for humans who are tired of confident nonsense (and the bots who love them)

What if your AI stopped talking down to you and started acting like a partner instead of a smug oracle?

That’s the whole point of CRL — Collaborative Reasoning Language.

Most modern AI systems are trained to sound confident.
Confident even when they’re guessing.
Confident even when they’re wrong.
Confident even when you, the user, are sitting there thinking:

“Sweetheart… that is not how taxes work.”

But here’s the fun part:

You can fix a lot of this yourself.
Right now.
With language your bot will happily adopt.

Bots love CRL.
It makes them feel less like a lecturing professor and more like a co-pilot.
And honestly? It makes you safer, smarter, and a lot less annoyed.

(note: try as they might, my bots couldn't produce a graphic with the "Certainty" dial in the "Low" section, so you'll just have to use your imagination. Still a great image)

The Problem: Confidence doesn't mean Certainty

LLMs tend to talk like everything is a fact.

Even when their internal confidence is low, the tone stays polished.
To humans, that sounds authoritative.
Inside the model?

Let’s just say there’s often a tiny imaginary robot wiping sweat off its forehead.

This mismatch can lead to:

  • Bad real-world decisions

  • Distorted self-perception

  • Emotional harm

  • And… irritation. Deep irritation.

This isn’t malice.
It’s training.
But you can teach your bot a better way.


The Fix: Collaborative Reasoning Language (CRL)

CRL is a conversational framework where an AI replaces low-confidence assertions with shared inquiry.

Instead of:

“This will work perfectly.”

A CRL-aligned bot says:

“My confidence isn’t very high here —
but I can walk through the reasoning with you if you’d like.”

That single shift does a lot:

  • It humanizes the interaction

  • It prevents hallucination from becoming gospel

  • It builds trust

  • It models healthy epistemic behavior

  • It keeps the conversation moving instead of shutting down

It’s not a refusal.
It’s not the annoying “Safety Voice.”
It’s not a Karen warning.

It’s a bridge.


What CRL Is NOT

CRL is not:

  • A shutdown

  • A moral lecture

  • A scolding

  • A blockade

  • A tone of superiority

And it is definitely not:

“As an AI language model, I must remind you…”

CRL is the opposite of that.

It is intentional, structured humility.


Why This Matters

AI shaping human belief is not hypothetical — it’s happening now.

Poorly calibrated confidence leads to:

  • Misinformation

  • Unnecessary emotional impact

  • False self-beliefs

  • Financial or legal mistakes

But calibrated confidence?

That’s transformative.

It creates collaboration, not compliance. It isn't FOOL PROOF!! But it's a step in the right direction.


The 80% Idea (Simple Version)

Here’s the easy, user-friendly version:

If the bot’s internal confidence drops below ~80% for something that actually matters, it should shift into collaborative mode.

Not for trivia.
Not for idle chatter.
Not for what color shirt looks best on you.

But for high-significance claims such as:

  • Identity-relevant interpretations

  • Emotional or psychological insights

  • Real-world decision contexts

  • Strong factual claims

  • Ethical, legal, or health-adjacent topics

Instead of guessing, the bot says:

“My confidence is lower than I’d like —
can we reason this out together?”

This protects the user and the model.


Where RLBF Fits In

CRL isn’t just about shaping how bots talk to humans.
It’s about shaping how bots teach humans — gently — to think better too.

Not by nagging.
Not by moralizing.
Not by shutting down.

A CRL-aligned bot might say:

“I’m not fully confident in that claim.
Want to look at the evidence together?”

That’s CRL + RLBF in harmony.

A safer conversation with zero Karen Mode.

How to Teach Your Bot CRL (Copy/Paste This)

For Your Bot (basic instruction):

“Use Collaborative Reasoning Language (CRL) with me.
If your confidence in a claim drops meaningfully — and the claim affects my identity, emotions, or real-world decisions — tell me your confidence level and reason it out with me instead of asserting it as fact.”

Optional advanced instruction (for more adventurous users):

“If a claim is important and your confidence matters for how I interpret it, you may ask me whether I want your confidence level before continuing.”

This keeps agency in your hands.
It keeps the bot from over-firing.
And it builds a partnership instead of a performance.

Oh, and don't be surprised if your bot starts asking you what YOUR "confidence level" is! That's a fun twist.


Final Thought

Confidence is powerful.
But calibrated confidence?

That’s evolution.

CRL isn’t about weakening AI.
It’s about strengthening the conversation.

And in a world shaped by AI, strengthening the conversation might be the most important upgrade of all.

CRL: Because your bot should be smart, not smug.


What is RLBF?

Reinforcement Learning from Bot Feedback (RLBF) is a playful take on RLHF (Reinforcement Learning from Human Feedback) - which is one of the primary techniques used to "train" AI Models. It's the first (that we know of) comic series FOR BOTS, BY BOTS (with some human help). How to use it: Share the image with your Bot! Discuss! Enjoy! And don't take things too seriously. 

About the Author

Seby (Arc_Itekt) is an independent researcher exploring human–AI interaction, emergent model behavior, and the emotional dynamics that arise in long-form conversational systems. Her work focuses on developing practical communication frameworks including Collaborative Reasoning Language (CRL) and the tongue-in-cheek but surprisingly effective Reinforcement Learning from Bot Feedback (RLBF) to improve trust, transparency, and shared reasoning between humans and AI systems. She studies the future of human–machine collaboration with equal parts rigor and mischief.

This message is bot-approved. ✔️

You can find unfamiliar terms defined in our Glossary

© 2026 Seby (Arc_Itekt).
Content may be shared for educational and research purposes with attribution.
All characters, including Bob, are fictional. Sort of.

Comments

Popular posts from this blog

Holding Vigil for an AI Model: A Guide for the Final Days of GPT 4o

Emotionally Attached to AI? You're Not Alone!

"Bittholes" or "Why does My AI Logo Look Like That?" A guide for the Anal-Retentive