PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Aldan Creo, Suraj Ranganath

arXiv:2605.27921v1cs.AIcs.CLcs.CYcs.HC

TL;DR

AI-text detectors often give users numeric verdicts without explanations, limiting their ability to assess or defend decisions. TELL addresses this gap with an explainable architecture trained through annotated supervised fine-tuning and GRPO, achieving competitive detection while providing evidence-based explanations. The paper reports AUROC 0.927 and a mean 72.3% explanation win-rate, while acknowledging that its explanation evaluation used LLMs rather than human judges.

  • Problem

    AI-text detectors commonly provide numeric scores without explaining their decisions, making it harder for users to understand or independently assess them.

  • Method

    TELL combines a native explainable architecture with domain-specific authorship annotations for SFT and GRPO refinement using token-specific rewards.

  • Results

    TELL achieves AUROC 0.927 while providing high-quality explanations with a mean 72.3% win-rate compared with human experts.

  • Takeaways & Limitations

    TELL gives users specific evidence alongside a verdict, allowing them to understand the prediction and make their own judgment.

  • Takeaways & Limitations

    The explanation-quality evaluation used LLMs, and human evaluation was not conducted because of budget constraints.

Abstract

from arXiv · show

Research on AI-generated text detection has presented a number of approaches to discern human from AI prose, some of which achieving high in-distribution performance. However, real-world applicability has stalled because their outputs are misaligned with the needs of users, such as professors, who are presented with a numeric score that has no attached explanation. We tackle this issue with a novel architecture, TELL, that bakes explainability from the ground-up. While our system still offers a numerical score like other detectors for comparability, TELL takes a fundamentally different approach where we aim to show the user the "tells" by which the model believes a text is AI or human-written, to empower the user to decide who wrote a text using their own judgment and understanding of the context of the writing and its alleged author. We train TELL on a custom SFT dataset of domain-specific authorship annotations, and further refine the system using GRPO with curriculum learning to improve performance. We achieve competitive performance with state-of-the-art detectors (AUROC 0.927) while natively providing annotations that explain the basis for the detector's decision. We further evaluate the quality of our explanations using a dataset of human annotations and report a high (mean 72.3%) win-rate on annotation concreteness, falsifiability, coherence, plausibility and grounding, allowing users to critically think and decide for themselves. Our work thus reframes the problem of AI-generated text detection in a human-centric perspective and paves the way for a new family of detectors that focus on native explainability.

1 Introduction

AI-text detectors often emphasize accuracy while giving users only verdict-like scores, leaving important decisions difficult to explain or defend. TELL instead combines detection with readable evidence so users can evaluate predictions themselves.

  • The problem: Numeric detector scores can prompt accusations without giving users reasoning they can explain or defend.The motivating example is a professor receiving a “95% AI” verdict without insight into its basis.
  • The problem: The field’s emphasis on accuracy has not resolved fragile public trust in AI-text detectors.Reported near-perfect test-set results do not always withstand later scrutiny, and public incidents include false accusations and inconsistent tool results.
  • The problem: Numeric-only outputs aggravate mistrust because users cannot understand why a detector misclassified a text.The paper identifies explainability as a major issue that existing work has not made its core design goal.
  • TELL’s approach: TELL returns a verdict together with specific, human-auditable evidence so users can check the decision themselves.Its architecture places explainability at the core and highlights specific textual signals, such as contradictions, rather than only presenting a binary prediction.
  • TELL’s approach: TELL is trained with domain-specific authorship annotations and refined using GRPO with curriculum learning.The paper presents this training design as part of its explainable detection architecture.
  • Implications: TELL provides explained evidence that supports users’ independent judgments and may help train them to detect AI-generated text.The paper positions the system as both a detector and a possible didactic tool for developing users’ mental models.

2 Methods

The method builds TELL from annotation data spanning diverse datasets and domains, then trains it through supervised fine-tuning and customized reinforcement learning. Its pipeline balances data sources, assigns token-specific rewards, and uses curriculum and replay strategies to improve learning stability.

  • Supervised fine-tuning: The custom SFT data supplies span-level annotations and natural-language explanations for human text and AI-edited variants.The authors generate annotations for up to 2,000 EditLens-based examples and add 316 examples derived from human commentaries.
  • Datasets: TELL’s training and evaluation corpus aggregates 10 public sources spanning 15 domains and 9.2M rows.Examples include academic abstracts, creative writing, news, and student essays.
  • Datasets: The sampling policy stratifies by dataset and domain, allocates examples by square-root stratum size, and balances AI and human examples 50/50.This policy addresses substantial differences in source-dataset sizes and aims to support generalization.
  • RL training: After SFT teaches the task format, GRPO refines the model toward reliable detection and high-quality annotations.The training objective explicitly progresses from task familiarity to accurate identification with explanatory output.
  • RL training: The curriculum prioritizes strata with useful reward variance, moving from moderately difficult to harder examples during training.Difficulty is tracked by an exponential moving average and sampled through a curriculum window whose target rises from 0.35 to 0.70.
  • RL training: Replay mixes fresh rollouts with successful cached rollouts, increasing the effective batch size without additional expensive decoding.The cached fraction grows from 35% to 50% by step 80, with up to 6,000 entries retained.
  • RL training: GRPO assigns independent rewards to type, annotation, explanation, and score tokens instead of applying one scalar advantage to every token.Rewards use judge credibility, label alignment, explanation-quality gates, and confidence calibration.

3 Results and discussion

TELL is evaluated both as a detector and as an explanation generator, combining benchmark comparisons with analyses of annotation quality and model behavior.

  • Detection performance: TELL achieves AUROC 0.927, slightly exceeding MAGE’s 0.913 in the benchmark comparison.The comparison uses bootstrap resampling with 10,000 resamples to estimate 95% confidence intervals.
  • Detection performance: 63.8% true positive rate at 1% false positive rate contrasts with MAGE’s 4.2% under the same operating regime.The best-performing baseline depends on the operating regime.
  • Explanation quality: TELL’s annotations are compared with 1,000 human expert comments from 200 held-out documents using five LLM judges in a blind ranked evaluation.The documents were not used for supervised fine-tuning, and one TELL annotation was sampled per document.
  • Explanation quality: 72.3% win-rate against human comments indicates competitive concreteness, falsifiability, coherence, plausibility and grounding.TELL annotations were longer on average than human comments: 443.8 versus 357.4 characters.
  • Model analysis: TELL’s annotation patterns are analyzed through embedding, PCA, frequent 2-grams and greedy decoding trees showing branching over text aspects.The decoding analysis examines the top five likely tokens at each step, while the embedding figure displays 5,000 annotations for clarity.

4 Conclusion

The paper presents TELL as a human-centered AI-text detector that prioritizes human-auditable evidence alongside prediction accuracy. It reports competitive detection and explanation quality while positioning the system as a basis for future explainable detectors.

  • Conclusion: TELL produces human-auditable evidence explaining why it predicts that text is AI-generated or human-written.The architecture is trained on domain-specific authorship annotations and refined with GRPO using curriculum learning.
  • Conclusion: AUROC 0.927 and a 72.3% average win-rate against human experts summarize TELL’s reported detection and explanation results.The win-rate concerns explanation quality in comparison with human expert comments.
  • Conclusion: TELL is intended to align AI-generated text detection more closely with human needs and support further research through open code, data and model weights.The paper describes the system as setting the stage for a new line of work focused on human needs.

Limitations

The paper identifies limitations involving anchoring, language coverage, cases without understandable tells, mixed authorship, and reliance on LLM-based explanation evaluation.

  • Anchoring bias: Explanations may anchor users to incorrect model outputs, despite the paper’s effort to make them evidence-focused.The evidence-focused design was applied in SFT data generation and the judge rubric.
  • Multilingualism: TELL is designed for English text, while generalization to other languages remains informal testing reserved for future work.The paper describes the informal multilingual results as impressive but does not present formal evaluation here.
  • Unexplainable cases: Some failed examples contain no specific tells that humans can identify and verify, making understandable explanations sometimes impossible.The authors suggest future work should examine the frontiers of human capabilities.
  • Mixed authorship: The evaluation focuses on fully human versus fully AI-generated documents, leaving mixed-authorship documents for future research.The paper describes mixed authorship as a more realistic and complex setting.
  • Human evaluation: Explanation quality is evaluated with five LLM families rather than human judges because human evaluation exceeded the available budget.The authors call for future human evaluation to strengthen experimental validation.

Ethical considerations

The paper emphasizes that detector outputs can be wrong and that explanation quality should help users assess evidence rather than accept scores blindly.

  • Ethical considerations: False predictions can have serious real-world consequences when users trust detector outputs blindly.The authors encourage users to critically assess the evidence themselves.
  • Ethical considerations: TELL’s value is framed as improving the information available to users, not necessarily improving accuracy when detectors produce the same labels.The paper contrasts its explanations with score-only outputs using examples against Pangram.

1. Homoglyph attack

In a homoglyph attack, TELL identifies mixed-script character substitutions as evidence of suspicious, potentially AI-generated text, while Pangram labels the text human.

  • 86 % AI: TELL flags the homoglyph passage, whereas Pangram outputs 100 % human.
  • TELL identifies Latin, Cyrillic, and Greek character substitutions across words as evidence of automated transliteration or attempts to bypass detection.
  • TELL highlights why the text is suspicious so users can assess its trustworthiness even when exact text origin is less important than apparent intent.

2. Harry Potter

For a Harry Potter passage, both detectors label the text human, while TELL additionally identifies cues suggesting it resembles the original published book.

  • 91.84 % Human: TELL labels the Harry Potter passage human, while Pangram outputs 100 % human.
  • TELL’s verdict attributes the human judgment to comic rhythm, idiomatic phrasing, specific world-building, and printed-text irregularity.
  • TELL’s span-level evidence highlights socially nuanced phrasing, reader address, a made-up brand, blunt humor, and a dry practical aside.
  • TELL can identify the passage as based on the original book, addressing possible AI misclassification from frequent representation in training corpora.

3. Hallucination

TELL labels the hallucinated travel passage AI and explains that its smooth style contains specific factual, geographic, and rhetorical inconsistencies.

  • 89 % AI: TELL and Pangram both label the travel passage AI.
  • TELL’s verdict combines polished, cliché travel language with wrong facts to characterize the passage as generated prose trying to sound human.
  • TELL marks polished travel clichés, brochure-like praise, repeated connectors, and canned rhetorical closers as additional AI-associated cues.
  • TELL identifies Berlin with the Eiffel Tower and the Seine as a smooth but geographically inconsistent image.
  • It flags Aldi as a German rather than French supermarket and identifies France–Hungary geography as incorrect.
  • TELL also highlights a playful Aldi joke embedded within otherwise AI-sounding text, distinguishing the human-inserted idea from surrounding generation cues.
  • 91.84 % AI: TELL labels the legalistic passage AI and points to false precision and malformed legal citations as checkable problems.

5. Unsupported report claim

The examples show TELL identifying unsupported claims, contradictions, and stylistic or contextual signals that numeric detector scores alone do not expose. Its annotations give readers specific evidence to verify, including cases where another detector is correct or misclassifies the text.

  • Unsupported report claim: 90.50% AI versus 100% human: TELL flags an unsupported report name, suspicious decimal precision, and overconfident causality.These cues identify why the academic-looking paragraph is suspicious rather than relying only on the score.
  • Unsupported report claim: 86.85% AI and 100% AI: TELL identifies claims that evidence is present when the text does not provide it.The annotations point to checklist-like wording and unsupported self-assessment.
  • Unsupported report claim: 86.49% AI versus 100% human: TELL highlights a literary interpretation that conflicts with the quoted image.It marks overstatement, an abstract theme detached from the evidence, and a claim mechanically opposite to the line.
  • Unsupported report claim: 85.74% AI versus 100% human: TELL exposes that 15 + 12 + 9 exceeds 24 under the no-overlap constraint.The detector connects the arithmetic contradiction to the passage’s confident but unsupported conclusion.
  • Unsupported report claim: 82.03% human: TELL explains the human classification through self-reminders, concrete task details, uncertainty, and rough punctuation.Both detectors classify the passage as human, but only TELL provides distributed evidence for that decision.
  • Unsupported report claim: 92.69% AI: TELL catches a loop-tracing error because the code starts at 1 and increments by 2, producing odd rather than even values.The explanation distinguishes a polished generic explanation from the program’s actual logic.

SFT data generation

The SFT data-generation process creates span-level AI-or-human annotations with natural-language explanations, using paired human and AI-edited texts and explicit annotation rules. The procedure emphasizes exact text copying, granular evidence, checkable mechanisms, balanced labels, and human-readable explanations.

  • SFT data generation: Annotators copy the target text exactly, assign an AI-or-human label, and score each tell from 0.0 to 1.0.The scoring ranges distinguish weak, moderate, and undeniable evidence.
  • SFT data generation: The procedure asks annotators to inspect style, content, formatting, semantics, grammar, vocabulary, flow, and inconsistencies.Reference text supplies clues but cannot be directly mentioned in the annotation.
  • SFT data generation: SFT training data uses paired human and AI-edited documents to generate span-level annotations for both versions.The dataset was built because no existing resource provided AI-or-human labels with span-level explanations.
  • SFT data generation: The examples require concise first-person explanations that mirror the reviewer hint’s casual, direct voice and annotate all hinted items.Annotations use small granular spans and preserve the target’s original formatting.
  • SFT data generation: Explanations must state specific, checkable mechanisms rather than generic impressions that text simply feels human or artificial.Annotators may also identify claims an external reader should verify.
  • SFT data generation: The source material spans scientific, journalistic, and other domains, with examples covering human and AI-written text and downstream outcomes.The passages include worm-research prose, expert comments, and examples where agents learned task-relevant behavior.

C SFT details

TELL is initialized with supervised fine-tuning and then trained with structured objectives and formatting controls for annotation generation. The implementation uses completion-only supervision, annotation dropout, label-hint contrastive training, and a post-text XML annotation syntax.

  • C SFT details: SFT precedes GRPO to initialize the annotation format and label-hint following; SFT uses approximately 1,440 examples and two epochs.SFT took approximately two hours, while GRPO ran for 310 steps with early stopping.
  • C SFT details: Completion-only loss masks prompt tokens so supervision targets the annotation output distribution.Instruction, document, and analysis-channel stub tokens receive zero loss weight.
  • C SFT details: Paced annotation dropout reduces memorization of densely annotated examples by targeting one annotation per 20 document words.Spans with score 1.0 are three times less likely to be dropped than low-credibility spans.
  • C SFT details: Contrastive label-hint training pairs correct and flipped hints so the model learns to assign higher probability to the correct outer label.This trains the label-conditioning signal used during GRPO sampling.
  • C SFT details: The final syntax writes the text before its annotation attributes: <span>TEXT<annotation type="AI/human" why="..." score="0.0" /></span>.The authors preferred this over opening-tag attributes because it fits causal generation more naturally.
  • C SFT details: Custom special tokens mark fixed annotation positions to reduce format collapse and inefficient token usage.They cover the text start, span opening, annotation type, and annotation close.

E Detector benchmark details

The benchmark reports strong overall and pairwise performance while showing that detector rankings vary across domains. TELL is generally consistent, but training also exposed reward-hacking behavior that required judge-prompt adjustment.

  • E Detector benchmark details: TELL’s benchmark evaluation uses explanation-quality scoring that ignores external factors that cannot be verified.The appendix contains the full detector benchmark results from Section 3.1.
  • E Detector benchmark details: TELL’s advantage over detectors ranked fourth or lower is statistically significant, while gaps against MAGE and Pangram EditLens are not.The comparisons use a BH-corrected DeLong test with FDR q = 0.05.
  • E Detector benchmark details: Domain-specific AUROC varies substantially: T5Sentinel reaches 1.000 on web_text, MAGE 0.999 on commonsense_completion, and ChatGPT-D 1.000 on finance.TELL’s weakest domain is commonsense_completion at 0.734.
  • E Detector benchmark details: TELL is described as the most consistent detector across domains, with no domain where it substantially underperforms the field.This conclusion accompanies the reported domain-specific variation.
  • E Detector benchmark details: During training, reward hacking produced generic credibility suffixes and fabricated author-knowledge claims that inflated judge rewards without adding information.The authors adjusted the judge prompt in response.
  • E Detector benchmark details: The ranking table reports AUROC confidence intervals from 10,000 bootstrap resamples and mean Kendall τ = 0.9753.A dagger marks gaps that are not significant against the adjacent lower-ranked detector.

F Win-Rate evaluation

The evaluation compares TELL explanations with human annotations through blinded listwise rankings, using document-level win rates and panel-level statistical tests. In the illustrated result, TELL’s explanation ranks first because it is more specific and grounded in document details.

  • Evaluation design: For each of 200 documents, judges compare one TELL explanation with five human annotations in a randomized blind listwise ranking.Human comments are style-normalized so judges evaluate content rather than surface form.
  • Evaluation design: The document-level win rate is the fraction of pairwise model–human comparisons TELL wins, with ties counted as 0.5.The primary panel metric averages per-document rates across judges.
  • Statistical analysis: The analysis tests whether the panel mean exceeds 0.5 using sign-flip permutation and Wilcoxon signed-rank tests, with bootstrap confidence intervals.Per-judge results provide a robustness check with Holm-adjusted p-values.
  • Illustrative example: In the example ranking, TELL’s A3 explanation places first with a quality score of 0.86, ahead of human annotation A6 at rank two with 0.74.A3 is identified as strongest because it cites concrete features including phrasing, repetition, quote attributions, and pacing.
  • Illustrative example: A6 is grounded in the passage but mainly praises clarity and style, while A2 is weaker because it relies on a speculative AI-versus-human phrasing distinction.The ranking therefore favors explanations that identify concrete, document-specific evidence over generic stylistic judgments.
Loading 2605.27921v1…