How to interpret an EraseGPT AI percentage

AI-text classifiers generally compare linguistic patterns in a sample with patterns learned or calibrated from human and machine-generated examples. The service returns a score that the EraseGPT interface expresses as a percentage. A stronger score means the current analysis found a stronger AI-like signal; a weaker score means it found a weaker signal.

That percentage is not the share of words known to have been generated. A value such as 70% should not be restated as “70% of this document was written by AI.” The detector did not watch the writing process and cannot reconstruct which person or tool produced each word. It evaluates the submitted text as evidence and returns an uncertain estimate.

Scores can be useful for triage: they can highlight a document that merits a closer look or let an author compare revisions. They are less suitable for proof. Even a confident-looking number can conceal uncertainty, dataset gaps and changes in the models being detected.

EraseGPT does not currently publish the underlying classifier's training corpus, feature weights, calibration curves, versioned evaluation set or performance broken down by language, genre, text length and source model. The public methodology therefore does not provide enough detail for an independent reviewer to verify model-specific accuracy.

We do not turn that evidence gap into an unsupported accuracy claim. In particular, the score should not be described as a verified success rate for ChatGPT, Claude, Gemini or any other named model. A responsible performance claim would require a defined test set, current model versions, reported false-positive and false-negative rates, subgroup analysis and reproducible evaluation procedures.

This page explains how to interpret the product safely based on what is visible to users. It is not a model card and it does not claim independent certification. If more validation evidence becomes available, this statement should be updated with the date, scope and limits of that evidence.

False positives

A false positive occurs when human-written text receives an AI-like classification. This can happen because human writing sometimes has the regularity a classifier associates with generated prose. Short answers, templates, technical definitions, translated text and heavily grammar-corrected writing can provide limited or misleading signals.

Peer-reviewed research by Liang and colleagues found that several GPT detectors disproportionately classified essays by non-native English writers as AI-generated. That finding is a strong reason not to treat detector output as neutral evidence across different language backgrounds.

False negatives

A false negative occurs when AI-generated or AI-assisted text receives a weak AI-like score. Editing, paraphrasing, mixed human and machine authorship, an unfamiliar model or a sample unlike the detector's evaluation data can all reduce the detectable signal. A low percentage therefore cannot certify human authorship.

Independent testing by Weber-Wulff and colleagues concluded that the evaluated AI-text detection tools were neither accurate nor reliable enough to serve as evidence of academic misconduct on their own. Tools and models continue to change, but the central methodological warning remains relevant: classification needs context and corroboration.

• Length and structure: a short passage or a list gives a classifier less context than sustained prose.
• Genre: academic abstracts, instructions, legal clauses and formulaic business copy naturally use repeated structures.
• Language background: patterns associated with language learning or translation can be misread by a detector.
• Editing tools: grammar correction, paraphrasing and collaborative edits change the features in the final sample without revealing its full history.
• Model and detector updates: a result can shift as generation systems, classifiers and thresholds change.
• Formatting and included material: prompts, references, quotations and boilerplate can affect the text being assessed.

Because these variables matter, compare only like-for-like samples and preserve the submitted version. Repeatedly editing a passage merely to reduce a score can make the prose worse without establishing anything about authorship.

1. Verify the input. Confirm that the complete, intended version was analyzed and note whether quotations, references or templates were included.
2. Treat the percentage as one signal. Record what it says without converting probability into a declaration of fact.
3. Seek independent context. Review notes, drafts, version history, sources, permitted tools and the writer's own account of the process.
4. Separate different questions. AI detection, plagiarism, factual accuracy and writing quality require different evidence.
5. Keep a human decision-maker accountable. Give affected people a chance to explain or challenge the result, especially in education and employment.

OpenAI withdrew its own experimental AI-text classifier after citing a low rate of accuracy and documented limitations. That does not mean all detection is useless; it illustrates why a polished interface cannot remove uncertainty from the underlying task.

Whether AI assistance is acceptable depends on the relevant school, employer, publisher, client or platform—not on a detector score. A low score does not make prohibited assistance compliant, and a high score does not invalidate work that followed the stated rules. Review the policy first and assess the evidence against that policy.

Google Search's published guidance similarly focuses on the quality and helpfulness of content rather than treating the use of automation as the only question. For writers, the durable standard is to create accurate, original, audience-focused work, disclose assistance where required and remain accountable for what is published.

• Weber-Wulff et al., “Testing of detection tools for AI-generated text”Peer-reviewed evaluation in the International Journal for Educational Integrity (2023).
• Liang et al., “GPT detectors are biased against non-native English writers”Peer-reviewed study published in Patterns (2023).
• OpenAI, “New AI classifier for indicating AI-written text”OpenAI's notice explains the classifier's withdrawal and records important limitations.
• Google Search Central guidance about AI-generated contentGuidance on automation, helpful content and search quality.