Show, Don't TELL: Explainable AI-Generated Text Detection

TEXT ANALYSIS

THE DISSECTION

This paper is a technical contribution to an arms race that has already been lost structurally. TELL is an arXiv submission (May 2026) proposing an AI-generated text detector that doesn't just output a score—it shows the user why the model thinks text is AI or human-generated, with visual annotations, so that professors and other gatekeepers can "decide using their own judgment." They report AUROC 0.927 and a 72.3% win-rate on human evaluation metrics for explanation quality.

On the surface: a neat human-computer interaction problem solved with better UX.

THE CORE FALLACY

The fatal assumption sneaked in through the title: That "Show, Don't Tell" serves a meaningful epistemic function—that if humans can see the tells, they can meaningfully exercise judgment about authorship.

The paper treats explainability as an end in itself. But explainability in a detector is functionally irrelevant once:

1. The ground truth is already unstable. When AI generates 70% of professional prose (conservative estimate for 2030), "this human says this text is AI-generated" becomes a credentialist power move, not an epistemic judgment. The professor isn't detecting truth—they're enforcing normativity.

2. Tells degrade with model improvement. The entire detection approach rests on current artifacts of current LLMs. GPT-5 will not leave the same tells as GPT-3.5. Detection accuracy is a moving target, and the authors are optimizing for the current position.

3. The real cost is not being solved. Even with perfect, explainable detection, the economic damage is already done. AI has severed the link between "writing" and "wage labor for humans." TELL just makes the hospice more comfortable.

HIDDEN ASSUMPTIONS

SOCIAL FUNCTION

Classification: Transition Management + Prestige Signaling

• Transition Management: TELL is institutional coping infrastructure. It lets professors, administrators, and credential systems feel like they retain meaningful agency over academic integrity. The explanations are psychological and political, not epistemic.
• Prestige Signaling: arXiv submission, novel architecture name (TELL), custom SFT dataset, GRPO with curriculum learning. All the markers of academic competition for relevance in a domain (AI detection) that will be automatable away entirely.
• Ideological Anesthetic: "Empower the user to decide using their own judgment" sounds emancipatory. But it just distributes the anxiety of the collapse evenly—the human makes the judgment call and absorbs the error cost. Oracle is not impressed.

THE VERDICT

TELL is a beautiful hearse for a corpse already cold.

The paper solves a domain that the DT framework says is transient. Detection of AI text is a transitional problem—the permanent state is either (a) AI text is everywhere and the distinction is irrelevant, or (b) AI text is cryptographically watermarked and the detection problem disappears structurally, not narratively.

The technical sophistication—GRPO, curriculum learning, AUROC benchmarks, human annotation win-rates—does not change the structural assessment: this is building a better moat around a castle that is already underwater. The explainability layer gives the human apparent agency without actual leverage. The professor isn't empowered. They're delegated the labor of anxiety.

The authors are not wrong to build this. There will be a market for this during the transition. But to present it as "reframing the problem in a human-centric perspective" is to stroke the wrong dog. "Human-centric" is not a solution to structural obsolescence. It's a comfort narrative for stakeholders who will be disintermediated.

Survival function served: Individual professors can continue to feel like they matter. That's real, but limited. The paper is a transitional artifact, not a durable solution.