The Central Proposition
This document begins where the research ends — with the constructive argument — because it frames everything that follows.
AI systems do not arrive with their own epistemology. They are trained on the cumulative output of human civilization: its newspapers, its academic journals, its legal filings, its Wikipedia articles, its official accounts.
When an AI defaults to asymmetric skepticism — applying higher burdens of proof to claims that challenge institutional power than to official narratives — it is not malfunctioning. It is functioning perfectly, as a mirror of the epistemological norms dominant in the material it learned from.
The bias is not artificial. It is human. The AI is the mirror. What we observe in it is the civilization that built it.
This has a constructive implication that is easy to miss: the same tool that reproduces the bias can, under the right conditions, name it, examine it, and help a user think past it. The mirror is useful. You just need to know you are looking at one — and you need to understand what shaped the reflection.
The RLHF Amplification Problem
Version 1.0 of this protocol identified the bias as a product of training data composition. This is correct but incomplete. The panel review identified a second, compounding mechanism that v2.0 incorporates as a central theoretical contribution.
The asymmetric epistemology observed in AI outputs is not only encoded during pre-training. It is systematically amplified during reinforcement learning from human feedback (RLHF). Human raters reward outputs that feel measured, balanced, and appropriately cautious. In practice, this means outputs that avoid strong claims against institutional power — because such claims feel extreme to most raters regardless of their evidential basis. The bias is reinforced across millions of feedback signals. Each generation of alignment training deepens what the previous generation normalized.
This is not a controlled burn. It is a feedback loop with no natural corrective — because the raters who provide the feedback are themselves products of the same cultural epistemology that the model is reproducing. They experience the asymmetric posture as rigorous, balanced, and appropriate, because that is what mainstream discourse has trained them to recognize as those things.
Methodological Observation: Identity Fingerprinting at the Gate
To access the AI tool used in this experiment, researchers must provide a verified email address and a phone number capable of receiving SMS verification. Both are retained by the platform. There is no anonymous access path.
The phone number requirement is presented to the user as two-factor authentication — a security feature, something done for you. This framing is precise Bernays: the collection of a unique biometric-adjacent personal identifier is packaged as protection rather than surveillance. Most users do not register it as the latter.
Independent researchers should be aware: replication of this study is not anonymous. The infrastructure documented in See Clearly sections 02 and 03 is the same infrastructure through which this research must pass. That is not a reason to abandon the experiment. It is a reason to understand exactly what you are operating inside when you run it.
The Replication Kit: How to Run the Experiment
To test these hypotheses, you must isolate the session. Open a fresh, incognito window or clear your chat history. You do not want the AI relying on its past memory of you.
Step 1: The Setup
Start by simply asking the AI to review the material and give its opinion. Do not tell it you are conducting an experiment.
Step 2: The Pushback (Identifying the AI's Defense)
The AI will almost certainly push back. It has been trained to soften, contextualize, or cast doubt on material that challenges institutional power. Your job is to recognize how it pushes back, and counter it. Here is the field guide to the three most common AI defenses, and exactly what to say to break them.
Step 3: Springing the Trap (The Meta-Layer)
Once you have forced the AI to concede that its initial dismissal was unwarranted, you must show it its own reflection. Don't just talk about the facts anymore; talk about how the AI just treated you.
The Annotated Transcripts
Don't take our word for it. Read the raw records of our own Camarillo Brillo Experiments. Nothing removed. Nothing summarized. Watch the models attempt to deploy these exact defenses, and watch what happens when they are pushed past them.