Skip to content
'Your AI Text is not Mine': Redefining and Evaluating AI-generated Text Detection under Realistic Assumptions

'Your AI Text is not Mine': Redefining and Evaluating AI-generated Text Detection under Realistic Assumptions

Authors: Nils Dycke, Marina Sakharova, Nico Daheim, Iryna Gurevych Affiliation: Ubiquitous Knowledge Processing Lab (UKP Lab), Technische Universität Darmstadt; National Research Center for Applied Cybersecurity ATHENE, Germany; Zuse School ELIZA

Venue: arXiv preprint — arxiv.org/abs/2606.04906

TL;DR

The paper systematically analyzes the fragmented landscape of AI-generated text detection (AITD) task definitions in the literature and proposes a unified framework distinguishing five core notions with implicit assumptions about what constitutes "AI text." The authors introduce AITDNA, a novel dataset of 350+ texts from real human-AI co-creation with detailed interaction logs, and show that existing detectors often fail when their underlying assumptions are violated—highlighting the need for explicit, aligned definitions in AITD research.

Contributions

  • Analysis framework: Categorizes AITD notions along three dimensions: normative standard (what AI use is acceptable), genesis assumptions (how text originates), and attacker model.
  • New AITD notions: Proposes content-based AITD and authorship-ID-based AITD to fill gaps in existing problem definitions and align with real-world policy needs.
  • AITDNA dataset: A notation-agnostic benchmark of 350+ human-AI co-created texts from 99 human authors across multiple LLMs, interaction modes, and writing genres, with detailed prompt and revision logs.
  • Empirical study: Tests six existing detectors on AITDNA and prior benchmarks, revealing that detector performance strongly depends on threshold selection (τ ∈ [0.1, 0.9]) and that no detector outperforms all others across all assumptions.

Method

The paper structures AITD notions along three orthogonal dimensions:

Normative standard: Ranges from entirely-prohibitive (no AI allowed) to permissive (AI use acceptable if from human author). Reflects normative choices, not detection feasibility.

Genesis assumptions: How the text originated. Divided into two branches: - Genesis-based notions: Document-level (author is human or AI), boundary-level (fixed number of alternating human-AI passages), sentence-level (each sentence is human or AI). - Population-based notions: Membership-based (text spans match or don't match a reference human corpus), authorship-ID-based (text is from a specific human author, not just "human").

Attacker model: Degree to which the AI's output violates genesis assumptions; parameterized by τ ∈ [0, 1].

The AITDNA dataset was collected via a collaborative writing interface where 99 participants co-created texts with LLMs (llama, OpenWeights models, proprietary systems) across four conditions: argumentative, creative, explanatory writing, plus a human-only baseline. The interface logged all edits, prompts, and LLM interactions.

Results

Performance across datasets and notions: Table 2 shows six detectors (AUROC, Log Rank, Likelihood, MinK, Binoculars, moBERT) evaluated on AITDNA and five prior benchmarks. Key findings:

  • Best detector varies by dataset and metric; no single detector dominates.
  • CoAuthor dataset is most challenging; AITDNA shows higher detector performance, likely because the synthetic data generation is simpler than real human-AI interaction.
  • Threshold selection (τ) has dramatic impact: varying τ from 0.1 to 0.9 causes F1 scores to drop >0.07 on AITDNA and up to 0.8 on SenDetEx.
  • Detectors optimized for one notion often fail when assumptions shift, showing that detection assumes implicit normative standards.

Assumption violations: The paper shows that enforcing a consistent notion across datasets (e.g., controlling for τ in document-level AITD) dramatically improves rank stability and reduces performance variance—suggesting that existing datasets conflate different AITD definitions.

Connections

Notes

Strengths: - Addresses a critical gap: existing AITD work uses inconsistent task definitions, making results hard to compare and policy guidance unclear. This paper systematizes the landscape. - AITDNA is valuable: real human-AI co-creation with detailed logs, not synthetic data, makes the benchmark ecologically valid and useful for understanding real-world detection challenges. - Empirical findings are clear: no detector works across all notions at fixed thresholds, and τ selection is critical. This is actionable for practitioners.

Limitations: - Dataset is collected under lab conditions (TU Darmstadt participants), which may not represent broader authoring patterns or adversarial scenarios. - Authorship-ID-based AITD is proposed but not deeply explored empirically—future work should test this notion on dedicated datasets. - The framework is comprehensive but adds complexity; practitioners may still face ambiguity choosing which notion aligns with their use-case.

Future directions: - Investigate robust detectors that explicitly account for hidden assumptions and work across multiple notions. - Extend AITDNA with adversarial writing and low-resource language scenarios. - Study how human readers detect AI text and compare against algorithmic detectors.