When AI can cheaply produce fluent, polished output, polish alone stops being reliable evidence of authorship, judgment, or real-world competence.

Generative AI can improve speed and evaluator-rated quality; that is an opportunity, not misconduct. It also means a finished document reveals less about who framed the problem, checked facts, made tradeoffs, integrated feedback, or achieved a result. AI detectors are unreliable. Stronger proof combines declared assistance, process history, human explanation, review, provenance, and outcomes.

Contents (6 sections)

Evidence snapshot

Supported by cited sources
FindingScope or qualification
In a preregistered experiment with 453 professionals, ChatGPT reduced completion time 40% and raised evaluator-rated quality 18%.Selected occupation-specific writing tasks.
In a 758-consultant field experiment, AI users completed more tasks faster and produced outputs rated more than 40% higher inside the AI frontier.On an outside-frontier task, AI users were 19 points less likely to reach the correct answer.
NIST identifies confident false content, including fabricated citations or logic, as a central generative-AI risk.Authoritative risk framework.
One detector study found a 61.22% average false-positive rate for nonnative-English TOEFL essays.Small, genre-specific study; enough to reject detector-only judgment.

What the evidence does not prove

Supported by cited sources

AI use can itself be job-relevant competence. Provenance can be stripped, live defense rewards confidence, team work complicates authorship, and human reviewers can be inconsistent and expensive. No single proof source is sufficient.

Why different stakeholders can care

Proposed

Workers gain credit for judgment rather than surface polish; employers gain stronger evidence; educators gain fairer assessment; creators gain provenance; and society can use AI without treating fluent output as self-authenticating.

Question to investigate

Open question

What The Solar Guild must test

Research question

Can layered evidence distinguish useful judgment from surface quality at acceptable cost without punishing honest AI use or creating false confidence?

Core sources

Supported by cited sources
Noy and Zhang — Productivity Effects of Generative AIDell'Acqua et al. — Navigating the Jagged Technological FrontierNIST — Generative AI ProfileAI-detector bias studyC2PA Content Credentials

References and limits

This page is published by The Solar Guild. Sources are listed so readers can inspect the basis of the page. A project statement is not independent proof unless the page identifies supporting outside evidence.

  • Problem 20 in The Solar Guild — Twenty Problems Worth Testing v1.0
  • Direct source links are provided above; quantitative claims retain their stated scope and qualification
  • The Solar Guild's response is explicitly separated from evidence that the problem exists