RESEARCH PROPOSAL: Assessing the Asymmetry Problem in AI‑Era Writing Assessment

1. Abstract

Current AI detection systems are designed to identify "high-quality" or "polished" AI-generated text. However, a critical asymmetry exists: while detectors can flag structured, A-level AI writing, they fail catastrophically when faced with "average" (C/D-level) writing, which is naturally messy, inconsistent, and highly variable. This research proposes a comprehensive study to quantify this asymmetry using a novel "Human-Entropy" modeling approach. By generating and analyzing thousands of text samples that mimic the natural drift, uncertainty, and imperfections of average human writing, we aim to demonstrate that current detection models are fundamentally incapable of preserving the integrity of the grading curve for mid-tier academic work.

Imagine Doogie Howser trying to simulate Forrest Gump. Doogie can easily mimic Forrest's messiness. But that very messiness is what makes the simulation undetectable.

The Asymmetry of Simulation: Consider the analogy of a high-IQ individual (Doogie Howser) attempting to simulate an average human (Forrest Gump). The high-IQ individual can easily mimic the average human's speech patterns, errors, and simplicity. However, the average human's unique variability is so complex that a high-IQ individual cannot perfectly replicate the specific messiness of a real human mind without triggering a detector that expects "perfect" AI.

• In our context: AI (Doogie) can easily generate text that looks like a "D" paper (Forrest).
• The Result: This creates a detection blind spot where AI-generated text passes undetected not by being "perfect," but by being "messy"—effectively collapsing the distinction between human and machine in mid-tier academic work.

2. Introduction & Problem Statement

The Core Insight: Academic integrity relies on the assumption that AI detection can distinguish between human and machine writing. However, this assumption collapses under the weight of statistical asymmetry.

There are only a dozen ways to write an 'A' paper, but a billion ways to write a 'D' paper. Current detectors see the dozen, but are blind to the billion.

• The "A" Zone: There are only a dozen ways to write a perfect "A" paper. Detectors are excellent at spotting AI that tries to write an "A."
• The "D/C" Zone: There are a billion ways to write a "D" paper. Human writing is naturally chaotic.

The Flaw: Current AI detectors are trained to flag the "dozen ways" of AI. They are blind to the "billion ways" of human chaos. An AI can easily generate a "D" paper by simply mimicking the statistical average of human error. The detector sees "messy" and assumes "human." The "grading curve" becomes meaningless because the "D" zone is now a black hole for detection.

3. Methodology (The "DragonFly Protocol")

• Dataset Generation:
• Control Group: 5,000 samples of authentic human "D/C-level" writing (messy, error-prone, variable).
• Test Group: 5,000 samples of AI-generated "D/C-level" writing, engineered using the "DragonFly Protocol" to mimic the infinite variance of human error (spelling mistakes, sentence fragments, metaphor drift, uncertainty).
• Detection Testing:
• Run both groups through commercial detectors (Turnitin, GPTZero, Copyleaks) and academic research models.
• Metrics:
• False Negative Rate: How often does the detector miss the "billion ways" of AI?
• Confidence Skew: Does detector confidence correlate with "polish" rather than "authenticity"?

4. Expected Outcomes & Implications

• Outcome 1: Confirmation that detectors have a >90% failure rate on C/D-level AI text.
• Outcome 2: Demonstration that "human-like" imperfections are the primary defense against detection, not "humanity" itself.
• Recommendation:
• Immediate: Shift assessment models away from pure text submission for mid-tier work.
• Long-term: Develop "Process-Based" detectors that analyze how a text was written (metadata, keystroke dynamics) rather than just the final output.
• Policy: Redefine "academic integrity" to include tool literacy rather than tool prohibition.

5. Ethical Safeguards

• Safeguards: The tool enforces permanent logging of all outputs; no user can hide or alter their generation history.
• Intent: This research is explicitly designed to improve detection systems, not to enable cheating. The experimental tools function as stress-test engines for academic integrity, with all logs preserved and made accessible to researchers to advance the field of AI detection.