Skip to main content
Back

TU Darmstadt Exposes Why AI Text Detectors Keep Failing Real-World Writing

A new benchmark and taxonomy expose how AI text detectors are built on incompatible assumptions that don't reflect how people actually write with AI

  • Taxonomy problem: The AITD field has no shared definition of "AI-generated text," making benchmark scores across papers incomparable.
  • 7 notions defined: The paper formalizes 7 distinct detection goals (document, boundary, sentence, content, intent, membership, authorship-ID), two of which are newly introduced.
  • New dataset: AITDNA captures real human-AI co-writing with full token-level genesis logs from 99 participants across 5 LLMs.
  • No universal detector: No single detector consistently wins across all notions and datasets; rankings shift dramatically when notion parameters are standardized.
  • Synthetic data misaligns: Existing benchmarks like DetectRL and Mixset simulate co-writing unrealistically, encoding assumptions that don't match how people actually use AI.
  • Open source: Code and data released at GitHub and HuggingFace under CC BY 4.0.

Every time a university flags a student's essay as AI-generated, or a journal rejects a paper for suspected LLM use, there is a hidden assumption baked into that decision: a shared understanding of what "AI-generated text" actually means. A new paper from the UKP Lab at TU Darmstadt argues that this shared understanding does not exist, and that the entire field of AI-generated text detection (AITD) is quietly suffering for it.

Although it is generally agreed that AI-generated text poses a broad societal risk, there is no common understanding in the AITD literature on what constitutes harmful use. Existing datasets and approaches often define their own criteria and make their own assumptions, sometimes implicitly, and often only loosely related to real-world needs and applications. The result is a field where benchmark scores are essentially incomparable across papers, and detectors that look great on one dataset fail silently on real-world text.

The problem nobody named

The core issue is that "AI-generated text" is not one thing. Consider the difference between a student who asks ChatGPT to write their entire essay versus one who uses it to polish a few sentences. Both involve AI, but most people would agree only one is a violation. Yet most detectors treat both the same way, because their training data was built under the assumption that text is either fully human or fully AI.

Human-AI co-creation of texts has emerged as a common writing pattern across domains, where a human author and LLM collaboratively write text resulting in nuanced authorship specific to each token. Most prior work does not frame AITD in terms of normative labels (permissible or prohibited) and relies on simplistic notion assumptions reflected in synthetically constructed texts. The question the paper asks is sharp: do these definitions and their associated data truly reflect how humans write with AI?

A second problem is comparability. While some works consider AI-generated text as any text that was output by an LLM, authorship is often more nuanced since LLMs can reproduce genuinely human text, for example from their training data, and paraphrase human input passages. This means the same piece of co-written text could be labeled "AI" under one dataset's rules and "human" under another's, making cross-paper comparisons meaningless.