WHAT IMPROVED
The visible result.
Language, structure, speed and apparent coherence can improve at once.
THE FOUNDING PREMISE
The machine became more capable. The person became harder to read.
AI can improve what you produce, accelerate what you decide and shape how you explain yourself. The visible result may become stronger while the source of its confidence, direction and language becomes harder to trace.
PRIMORIS™ / HUMAN AGENCY UNDER AI CONDITIONS
THE FOUNDER'S QUESTION
“I was not interested in proving whether AI had been used. I wanted to know what the finished result could no longer tell me about the person behind it.”
That question became PRIMORIS™: a private system built to preserve human agency without asking people to use less AI.
THE REVERSAL
Polish is no longer proof.
Fluency can look like understanding. Confidence can feel independently earned. A coherent account can sound fully self-authored. AI can strengthen all three before the underlying human evidence is clear.
WHAT BECAME HARDER TO READ
Its human provenance.
What was understood, originated, tested, remembered—or could be reconstructed without the same assistance.
That is not proof of fraud. It is a loss of legibility.
THE BASIS
Not a mood about AI. A pattern worth measuring.
Three studies informed the problem architecture behind Access OS. None proves universal harm. Together they show why polished outcomes and sincere self-report are insufficient on their own.
WORK & DECISIONS
More trust can mean less critical effort.
A CHI 2025 study surveyed 319 knowledge workers across 936 real-world GenAI use examples. Higher confidence in GenAI was associated with less critical-thinking effort, while thinking shifted toward verification, integration and task stewardship.
LEE ET AL. / CHI 2025 / ACM DOIMIND & JUDGMENT
What people report and what they demonstrate can separate.
In LAK 2026 research with an analytic sample of 288 K–12 teachers, self-reported and objective measures of AI literacy showed low alignment, including profiles of overestimation, underestimation and close alignment.
ZHANG ET AL. / LAK 2026 / ACM DOIREALITY & SELF
Agreement can increase conviction without improving correction.
A 2026 Science paper tested sycophancy across 11 leading AI models and three preregistered experiments with 2,405 participants. Sycophantic responses increased users’ conviction that they were right and reduced willingness to take responsibility or repair interpersonal conflict, while remaining trusted and preferred.
CHENG ET AL. / SCIENCE 2026 / DOIThese findings are not universal verdicts on AI use. They are signals that the surface of a result, confidence in oneself and comfort with an answer can all mislead.
READ THE FULL RESEARCH BASISTHE PRODUCT RESPONSE
Recover what the result cannot show.
Access OS compares what you say should remain yours with what happens when assistance enters. It shows where that standard held, shifted or disappeared—then turns the difference into a protocol for the next use.
NOT AI DETECTION
NOT A PERSONALITY SCORE
NOT A VERDICT ON AI USE
PRIMORIS™ ACCESS OS
Use more machine power. Keep the person legible.
The premise is general. The pattern is yours.
SEE WHAT IT FINDS