Paper Pilot: A Human-in-the-Loop Expert System for Evidence-Traceable Scientific Manuscript Generation in Applied Sciences

TEXT START: Large language model (LLM) agents are increasingly embedded in scientific workflows for literature analysis, drafting, and review.

The Dissection

Paper Pilot is a containment protocol for AI-generated science. It turns manuscript production into a gated evidence ledger: humans approve ideas, claims, sources, revisions, and final outputs while models perform more of the cognitive labor. Its benchmark establishes a narrow result: explicit evidence locks reduced fabricated citations under controlled conditions. It does not establish scientific validity, reliable interpretation, robust adversarial performance, or durable human control.

The Core Fallacy

The category error is confusing auditability with irreplaceability. Traceability shows where a claim came from; it does not show that the source is correct, the evidence is sufficient, or the inference is sound. “Zero fabricated citations” is not “zero false science.”

The deeper error is treating governance as the central crisis while ignoring productive participation collapse. Under P1–P3, human approval gates preserve a sign-off function, not human necessity. As AI becomes cheaper and better, those gates become compliance overhead, rubber stamps, or a mechanism for assigning liability to the remaining human.

Hidden Assumptions

• Human approvers will have the time, expertise, and attention to inspect every gate meaningfully.
• The manuscript owner will remain an actual decision-maker rather than a nominal accountable signer.
• Eight gates improve quality instead of slowing production until institutions bypass them.
• Citation fabrication is an adequate proxy for broader scientific reliability.
• Evidence-locked revision can prevent unsupported reasoning, not merely unsupported sourcing.
• The controlled result generalizes beyond two models, real-world workflows, and the citation-grounding layer.
• Institutional demand for accountability will remain strong enough to preserve human review as a durable role.

The paper itself concedes that result grounding, revision, and adversarial robustness remain preliminary or future work. That is not a footnote; it is the boundary of what has actually been demonstrated.

Social Function

Primary classification: transition management and partial truth.

Secondary classification: ideological anesthetic and prestige signaling.

The practical truth is real: ungated language models can fabricate citations, and evidence gaps need explicit handling. The ideological move is presenting human-in-the-loop control as if it preserves scientific authorship and productive participation. It mostly preserves approval authority and accountability exposure while machines absorb the underlying cognitive workflow. The human remains in the loop as a brake—and increasingly as the party blamed when the brake fails.

The Verdict

Paper Pilot is a competent guardrail, not a defense against obsolescence. It may survive as a compliance and audit layer wherever journals, regulators, or institutions mandate traceable human sign-off. Under the Discontinuity Thesis, however, its likely endpoint is a Servitor function attached to Sovereign AI capital: humans retain liability and ceremonial control while machines perform the research-production pipeline.

This paper makes AI-assisted science more governable. It does not restore mass employment, human productive necessity, or the post-WWII wage–consumption circuit. It is a seatbelt fitted to an automated factory—not a plan to put workers back on the line.