acceptodds
Withdrawn from ICLR 2027

Trust, but Verify: A frontier model benchmark and a leakage-free protocol for AI-generated text detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

AI-generated-text detectors increasingly inform academic-integrity and content-moderation decisions, yet reported accuracy is hard to trust. We audit three vulnerabilities. (1) Evaluation leakage: test-informed checkpoint selection inflates macro-F1 by 27.3 points in a controlled MAGE replication. Our COLING-2025 endpoints differ by 1.72 points (83.65 → 85.37), but incomplete candidate histories prevent attributing that difference solely to selection. (2) Distribution-sensitive human-text flagging: on domain-matched scientific abstracts, RoBERTa-large-OpenAI flags 45.5% of post-2023 text against 20.1% of pre-2015 text. This direction persists within both PubMed and arXiv, although source composition changes some detectors’ aggregate conclusions. (3) Heterogeneous frontier performance: evaluated off-the-shelf detectors span macro-F1 58–85 on 2026 frontier text. The author-released DetectAnyLLM checkpoint achieves the highest external macro-F1, 84.60, compared with 76.67 for the COLING-2025 winner Advacheck, yet flags 48.78% of our post-2023 presumed human abstract probe. We evaluate the author-released checkpoint, not a reproduction of the published result. Rankings change with the operating point, and Binoculars and Fast-DetectGPT recover only 14–19% of Claude Sonnet 4.6 and Grok-4.20-reasoning text. We release FrontierBench-2026, comprising 216,352 texts from 16 collected frontier LLM variants (15 evaluated), a provenance-verified pre-2015 human reference, a domain-matched contemporary probe, and an evaluation harness separating development-set selection from test reporting. Using a graph-fusion detector as an audit instrument, we find no stable advantage over the shared-task winner across three seeds (paired macro-F1 difference +0.08 ± 0.24, mean ± s.d.). Frontier-aware training raises frontier macro-F1 from 78.4 to 92.8 on the shipped split and achieves 89.8 ± 0.2 on a content-controlled split with disjoint prompts and parent documents. Held-out-provider recall is 0.845 shipped and 0.764 content-controlled. Capability-tier transfer is provider-asymmetric, while short-form detection and genre-dependent contemporary-human false alarms remain unresolved.

Primary area:
datasets and benchmarks
Keywords:
AI-generated text detection, large language models, benchmark datasets, evaluation leakage, temporal distribution shift, out-of-distribution generalisation, reproducible evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_7uemOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies AI detection in text, builds a large corpus of generated texts using the latest models, paired with texts from 2015 (i.e., not AI since pre-ChatGPT), and trains classifiers, showing they outperform certain weak baselines from a COLING workshop.

The paper is (ironically) entirely AI-generated, and absolutely filled with low-quality "AI slop" catchphrases, with unfortunately very little conceptual clarity beyond the hypey phrases. It suffices to read the Abstract which jumps right in and provides extremely little context, to understand the little value this work would add to the community.

The fact that the work is about AI detection is particularly ironic: I wonder how their draft would be classified by the models they trained. I encourage the authors to reconsider what they submit to conferences, as submitting low-quality drafts such as these is leading to a crisis in science. Instead, I invite the authors to actually work on AI detection, which is an important topic, and take the time to distill their findings into a thoughtful draft, which will help our community better understand how AI detection works. Please do not offload your thinking (and your thinking) to AI.

Critical strengths

None, to be very frank.

Critical weaknesses
  • I don't think the dataset creation proposed is valid, it is just a big shortcut: they claim to have this "2026 frontier paired dataset of human-written vs. AI-generated texts", but only the AI-generated texts (the easy part) are from 2026, which the human-written part is from 2015...
  • The comparisons all involve cryptic, unknown models. There does exist high-quality AI detection models (GPTZero, Pangram, just to name a few, including recent Pangram's open-source model). Why not compare to actual, known baselines...
  • The writing is absolutely atrocious, and could likely serve in a classroom for all the AI cliches for what not to do for clear, scientific communication. It would not surprise me if this paper was generated entirely autonomously by an AI agent.
Questions for authors

See weakness and summary.

Reviewer_LtoAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces FrontierBench to evaluate AI-text detectors on recent model outputs. It examines model-selection bias, performance on contemporary presumed-human text, and generalization across generators. It also studies a graph-fusion detector and the effect of training on frontier-model outputs.

Critical strengths
  • Evaluating detectors on recent generators is useful, and the benchmark includes reasoning and non-reasoning variants. The results show substantial differences across generators that aggregate detection scores can hide.
  • The content-controlled split separates prompts and parent documents across training and evaluation. Retraining under this split and testing held-out providers provides a more informative assessment of generalization.
Critical weaknesses
  • The benchmark pairs pre-2015 human text with AI text from 2026. The controls help, but cannot fully separate authorship from differences in era and source. The recent human-text probe has uncertain AI involvement, so a contemporary set with verified human authorship is still needed.
  • Only 15 AI test documents contain 200–999 characters. Given the strong dependence of detection accuracy on length, this gap limits how broadly the aggregate results can be interpreted.
  • The paper tries to cover too much at once: a benchmark, evaluation biases, a new detector, and transfer experiments. It difficult to identify the main contribution and follow the key findings. A clearer focus and less detail in the main narrative would improve readability.
Questions for authors

See above.

Reviewer_WgM5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces FrontierBench-2026, a 216K-text benchmark pairing 16 frontier LLM variants from Anthropic, OpenAI, Google, and xAI with pre-2015 human text and a post-2023 human probe. Using the dataset the authors investigate how AI-text detectors are evaluated. The authors train their own ModernBERT-based detector which performs on par with the COLING-2025 winner.

Critical strengths

FrontierBench-2026 is a well-constructed and clearly costly resource including 216K texts from 16 frontier variants across four providers, with matched reasoning-on/reasoning-off outputs from the same models and prompts, a provenance-verified pre-2015 human reference, and a content-controlled split.

Beyond the general importance of detecting LLM-generated text, the paper demonstrates that currently deployed detectors are failing in both directions on 2026 text. Recovering under 20% of output from the newest flagship and reasoning models while flagging 20–45% of contemporary human-written abstracts.

Critical weaknesses

The methodological contribution is thin. The benchmark follows a standard recipe. The findings (test-set leakage, out-of-distribution degradation, era-dependent false positives) are either very predictable or well-known. The model training also follows a basic methodology.

The paper consists of 42 pages, the overview figures is in the appendix, and with all the dense inline formatting (bold, italics, em-dash and semicolon chains) it makes it very hard to read.

Questions for authors

see weakness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.