AI Detectors Fail Diverse Student Populations: A Mathematical Framing of Structural Detection Limits

1. Problem Statement

The paper addresses the widespread unreliability of "black-box" AI text detectors in university settings, specifically their tendency to produce high false positive rates (FPR) against diverse student populations (e.g., non-native English speakers, students from specific disciplines).

While existing literature attributes these failures to the convergence of AI and human writing styles (making distributions harder to distinguish), this paper argues that a more fundamental, structural barrier exists. The core problem is that university assessment typically involves a composite null hypothesis:

• Standard Model: Detection is modeled as a test between two known distributions: $p_H$ (Human) vs. $p_M$ (AI).
• University Reality: The assessor does not know the specific writing distribution of the individual student ($p_{\theta_i}$). The null hypothesis is a family of distributions $\{p_\theta : \theta \in \Theta\}$, where $\theta$ represents unknown student characteristics (style, ability, linguistic background).

The paper posits that this population diversity creates a mathematical limit on detection accuracy that is independent of AI model quality or detector engineering.

2. Methodology

The authors employ fundamental statistical hypothesis testing theory, specifically utilizing the Total Variation (TV) distance and its variational characterization.

• Mathematical Tool: The analysis relies on the variational inequality for TV distance:
  $$|E_P[\phi] - E_Q[\phi]| \leq TV(P, Q)$$
  Where $\phi$ is the detector function, and $P, Q$ are probability distributions.
• Framing: The problem is reframed from a "Simple vs. Simple" hypothesis test (known $p_H$ vs. $p_M$) to a Composite Null test ($p_\theta$ vs. $p_M$, where $\theta$ is unknown).
• Assumptions:
  • One-shot detection: The detector evaluates a single document without prior knowledge of the student's writing history.
  • Text-only: The detector relies solely on the text of the submission.
  • Known Alternative: The AI distribution $p_M$ is assumed to be known (or sampleable) by the detector designer, while the human distribution is unknown and variable.

3. Key Contributions

The paper makes three primary theoretical contributions that formalize the structural limits of AI detection:

1. Separation of Failure Modes: It distinguishes between two independent sources of detection difficulty:

   • Source 1 (AI Quality): AI models improving to mimic humans (convergence of distributions).
   • Source 2 (Population Diversity): Natural variation in human writing styles causing overlap with AI distributions, even if AI quality is static.
   • Crucial Insight: Source 2 cannot be solved by better detectors; it requires changing the assessment protocol.

2. Quantitative Trade-off Bounds: The paper derives three mathematical results proving that any useful detector must incur false accusations if there is distributional overlap between any subset of students and AI output.

3. Subgroup Mixture Bound: It connects abstract theoretical quantities (TV distance between individual student distributions and AI) to observable, real-world demographic subgroups (e.g., L2 speakers, specific majors), providing a theoretical basis for documented empirical biases.

4. Key Results

Result 1: Average-Case Size–Power Trade-off

If a fraction $\pi(\Theta^*)$ of students have writing distributions within TV-distance $\delta$ of the AI output, any detector with power (True Positive Rate) $\beta_0$ must incur a population-averaged False Positive Rate ($\bar{\alpha}$) of at least:
$$\bar{\alpha} \geq \pi(\Theta^*) \cdot (\beta_0 - \delta)$$

• Implication: Even with a highly accurate detector, if 10% of students write in a style close to AI ($\delta$ is small), the detector will inevitably falsely accuse a significant portion of that group to maintain high power. This explains empirical FPRs of 10–20% in diverse cohorts.

Result 2: Worst-Case Protection Bound

If an institution requires that no individual student face a false positive rate exceeding $\alpha_0$, the detector's maximum power is bounded by:
$$\beta(\phi) \leq \alpha_0 + \delta^*$$
Where $\delta^*$ is the TV distance between the AI output and the nearest human writer in the cohort.

• Implication: To guarantee near-zero false accusations for the "most AI-like" human writer, the detector's ability to catch actual AI usage must be near zero.

Result 3: Subgroup Mixture Bound

For any demographic subgroup $G_k$ with a pooled writing distribution $\bar{p}_{G_k}$, the subgroup-average FPR satisfies:
$$\bar{\alpha}_{G_k} \geq \beta(\phi) - TV(\bar{p}_{G_k}, p_M)$$

• Implication: If a subgroup's collective writing style is statistically close to AI output (small TV distance), any detector with useful power will produce high false positives for that specific group. This mathematically validates empirical findings that non-native speakers and EFL learners are disproportionately flagged.

5. Significance and Policy Implications

Theoretical Significance:
The paper proves that the failure of AI detectors in diverse populations is not a technical flaw but a mathematical inevitability arising from the composite null hypothesis. No amount of engineering can overcome the size-power trade-off when the human population is diverse and the detector is restricted to a single text sample.

Practical & Policy Recommendations:

1. No Sole Evidence: Detection scores should never serve as the sole or primary evidence in academic misconduct proceedings due to the mathematically unavoidable risk of false findings.
2. Stratified Auditing: Institutions must audit detectors per subgroup (e.g., by language background, discipline) before deployment. A detector may be valid for one subgroup but structurally inadequate for another.
3. Assessment Redesign (The Only Solution): Since the barrier is structural to the "one-shot, text-only" format, the only durable solution is to change the information structure:
   • Move to process-based assessment (drafts, oral defenses).
   • Use longitudinal writing profiles to transform the composite null into a known null.
   • Design open-ended, reflective tasks that increase stylistic variance and distance from AI patterns.

Conclusion:
The paper concludes that investing in better detectors is futile for solving the diversity problem. Instead, institutions must accept that population diversity creates a structural limit on detection and respond by redesigning assessments to rely on richer evidence than a single submitted document.