Would an AI have written this anyway?

Neither tells you whether the ideas are the author's. A watermark survives in a piece a writer merely proofread with AI, and vanishes from an AI draft that was reworded by hand. Anthropic's own explainer lists the cases where the mark breaks down and says it cannot prove authorship. For text, no vendor offers a public detector yet: Anthropic has announced one, Google's covers other media, and OpenAI marks images only.

Sources: BleepingComputer on Anthropic's text watermark, The Decoder on the announced detection API, eWeek on the SynthID Detector launch, Google's SynthID documentation.

A Stanford study found seven detectors flagged the majority of essays by non-native English speakers as machine-written while classifying native speakers' essays almost perfectly. As more people write with AI assistance, "was a model involved" stops separating good work from filler. A piece can be written entirely by hand and say nothing new, or drafted with help and carry a real finding.

Sources: Liang et al., GPT detectors are biased against non-native English writers (2023), Pangram Labs technical report on its classifier (2024).

The guidelines define originality as content "not available on other websites" and effort as the extent to which "a human being actively worked to create satisfying content". They name the tells of paraphrased content: only commonly known information, heavy overlap with reference sites, or a summary of one page with nothing added. Google also holds a patent on scoring a document by the information it adds beyond what the reader has already seen.

Sources: Google Search Quality Rater Guidelines (PDF, sections 3.2 and 4.6), Search Engine Land on the January 2025 update, Google patent US20200349181A1, contextual estimation of link information gain.

Wenger and Kenett found answers from different models far more alike than answers from different people. Lu et al. introduced a Creativity Index that measures how much of a text can be rebuilt from existing web text, and found human writing scored far higher. Document-level novelty metrics such as NovAScore score a piece by its atomic facts, and Google DeepMind's long-form factuality work checks each fact against search. Verdict draws on this line of research without being a copy of any one method.

Sources: Wenger and Kenett, creative homogeneity across LLMs (2025), Lu et al., AI as Humanity's Salieri, ICLR 2025, NovAScore, COLING 2025, Wei et al., long-form factuality in LLMs (2024), Measuring information potential in text corpora (2025).