Characterizing initial human-AI proof formalization workflows

URL SCAN: Characterizing initial human-AI proof formalization workflows
FIRST LINE: Computer Science > Artificial Intelligence [Submitted on 2 Jun 2026]

TEXT ANALYSIS PROTOCOL

1. THE DISSECTION

This is a human factors study of AI-assisted mathematical proof formalization. It surveys how people want to use AI in formalization (preserving "high-level human control over proof discovery") and measures actual performance with and without AI tools. The framing is empirical, descriptive, and optimistic — a classic early-adoption user study that treats the technology as a tool to be shaped by human preference rather than a structural force reshaping what human mathematical participation means.

2. THE CORE FALLACY

The foundational error: Treating this as a question of user preference and workflow optimization, when the actual story is the replacement of the cognitive substrate itself.

The paper treats "human control over proof discovery" as a desideratum worth preserving and studies whether AI assists or hinders that preservation. It never asks the structural question: what happens to the role of human mathematical cognition when formalization — the last bastion of high-status intellectual labor — becomes machine-readable, machine-verifiable, and machine-generatable? The paper optimizes for a transitional interface. It ignores that the interface itself is the mechanism of displacement.

Mathematical proof work is one of the final refuges of prestige cognitive labor. The paper documents early-stage displacement while treating it as a usability question.

3. HIDDEN ASSUMPTIONS

• That "high-level human control over proof discovery" is a stable preference rather than a transitional comfort. Preferences shift when the alternative is irrelevance.
• That formalization accuracy is the relevant metric. The real variable is who controls the ontology of formal systems — the definitions, axioms, and proof strategies — not whether humans can formalize faster with AI assistance.
• That "diverse preferences" among participants represents meaningful agency. Diversity of preference in a declining domain is noise, not signal.
• That the lag between tool capability and integration constitutes meaningful optionality for mathematicians. It doesn't. It constitutes a grace period, not a defense.

4. SOCIAL FUNCTION

Classification: Prestige Anesthetic + Transition Management

This paper performs two functions simultaneously:
1. For mathematicians: It tells them their role survives by repositioning them as "high-level controllers" of a process that AI increasingly runs. The fantasy of strategic oversight.
2. For AI developers: It provides empirical legitimacy for AI formalization tools by documenting human acceptance and performance gains — a user study as regulatory/social-acceptance softener.

It is partial truth (AI tools do improve formalization accuracy in current studies) wrapped in a narrative that renders the structural displacement invisible by focusing on the transitional interface.

5. THE VERDICT

This paper is a transitional artifact from the pre-collapse phase of cognitive labor displacement. It documents a real phenomenon — AI improving human formalization performance — while systematically misidentifying the phenomenon's nature. It reads as a workflow optimization study. It is actually a time-lapse of a ceiling collapse, filmed from inside the room, labeled "ceiling interaction patterns."

The honest title would be: "Characterizing the Late Stages of Human Cognitive Participation in Mathematics: Evidence That the End Is Gradual, Then Sudden."

The DT diagnosis is precise: P1 (Cognitive Automation Dominance) is executing in real time in mathematics, one formalization workflow at a time. The paper is evidence of that execution, filed under "user experience research."

Mechanical Death Timeline: Within 3-5 years, "AI assistance in formalization" becomes "AI execution of formalization with human verification" — then "AI verification of AI-generated formalization." The human layer becomes residual. Within 10 years, formal mathematical discourse becomes an AI-to-AI process with human-readable summaries as a legacy interface.

No moat. Mathematics is P1-compliant by design. It is code-adjacent, formally specifiable, and already AI-native.