GPTZero¶

FreemiumProfessional

Vendor: gptzero.me | Type: Detector | Cost: Freemium

TL;DR

The most documented AI text detector, and the cautionary case the toolkit ships specifically because the failure modes are publicly evidenced. Stanford 2023 documented up to 61% false- positive rate on non-native English writing; Perkins found 26.4% accuracy on non-manipulated AI text and 16.7% on manipulated text. Use only as one weak signal; never as standalone evidence.

What it does¶

GPTZero accepts a text input through a web interface or Chrome extension and returns an AI-generation probability with sentence- level highlighting of likely AI passages. The underlying method evaluates perplexity (how "predictable" the next token is for a language model) and burstiness (variance in sentence-level predictability). The tool is one of the most-referenced AI-text detectors in the journalism and academic literature, and is the detector with the most published independent evidence, including evidence that documents its failure modes in detail.

The toolkit's editorial position is explicit. GPTZero ships in 1B.2 as the alternative to Pangram precisely because its failure modes are publicly documented; using GPTZero in the toolkit's workflow means using it as a cautionary case, with the verbatim Stanford and Perkins findings on the card and a binding weak-signal-framing instruction in the workflow section. A fact-checker who reaches for GPTZero on a non-native English text without that framing is, per Anchor 2, taking on a defamation risk the toolkit considers unacceptable.

When to use it¶

A piece of suspect English text needs a directional read and you want the most-documented detector with the most-published independent evidence, including the evidence that bounds its reliability.
A workshop demonstrates the cautionary case: GPTZero plus the Stanford 61% FPR plus the Perkins 26.4% / 16.7% findings together show what the toolkit means by "do not use as standalone evidence."
An academic or research case explicitly wants to test a known detector against a controlled input set; GPTZero's documentation and benchmark presence make it a defensible test subject.
A regional case involves text in English; treat any verdict as one weak signal class and pair with non-detector verification before publication.

Independent accuracy¶

Vendor claim vs independent assessment

Vendor claim: GPTZero markets itself as a "trusted AI detector" but does not publish a sharp headline accuracy figure that survived the regional research source check.

Independent finding (Perkins et al.): Perkins et al. adversarial robustness study found GPTZero extremely weak in adversarial settings: 26.4% accuracy on non-manipulated AI text and 16.7% on manipulated text. Sample size: 805 detector tests / 114 samples.

Independent finding (Stanford 2023): the canonical Stanford 2023 evidence found up to 61% false positive rate on non-native English writing, falsely flagging real human writing by ESL authors as AI-generated.

Independent finding (2025 essay study): a 2025 preprint found GPTZero usually marked AI essays as highly AI-generated but still produced false positives on some human essays (28 AI essays + 50 human essays).

No robust SEA-language breakdown located in the regional research search. A 2025 summary indicated GPTZero struggled on non-English texts; coupled with the Stanford 2023 ESL false-positive finding, the SEA-language risk profile is structural.

Limitations¶

Limitations

10-20% false positive rate documented for non-native English (ESL bias).
Stanford 2023 study found up to 61% false positive rate on non-native English writing.
Limited efficacy on complex Asian languages.
Perkins 26.4% accuracy on non-manipulated AI text and 16.7% on manipulated text.
The text detector class as a whole is unreliable for high-stakes use; using GPTZero on a non-native English text to support a public claim against a person is, per the toolkit's editorial position, a defamation risk.

Privacy and threat model¶

GPTZero is a vendor web service; the input text uploads to GPTZero's cloud infrastructure for analysis. GPTZero is a US-jurisdiction proprietary vendor; treat any upload as a transmission to a US server with the vendor's retention and disclosure policy applying. For text inputs the source-protection risk is lower than for image or audio (text rarely carries diagnostic background), but source-identifying language patterns or quoted private communication should still be considered before upload.

Country and platform applicability¶

Indonesia: English-optimised; Bahasa Indonesia performance unverified and structurally unreliable per the regional research class-level evidence.
Laos: Lao not covered.
Malaysia: English-optimised; Malay performance unverified.
Philippines: English-optimised; Filipino performance unverified.
Sri Lanka: Sinhala and Tamil not covered.
Thailand: English-optimised; Thai performance unverified.

The Stanford 2023 ESL finding has direct regional implications: much SEA writing in English is by non-native speakers; GPTZero will falsely flag genuine human writing at materially elevated rates in all six focus countries when the writer is an ESL author. The Limitations and Independent accuracy sections name this verbatim.

Platform applicability: works on any text regardless of source platform.

How to access¶

The web interface is at gptzero.me. Free tier covers 10,000 words per month; paid tiers from $9.99/month give higher quotas. A Chrome extension extends the same workflow to any webpage. Account creation required for higher-volume free-tier or paid use.

Cost (current as of 2026-05)¶

Freemium. Free tier: 10,000 words per month. Paid tier from $9.99/month . Verify pricing directly before any institutional commitment as vendor pricing may have shifted.

Quickstart¶

Open gptzero.me in your browser.
Paste the suspect text into the analysis box, or upload a document.
Read the AI-generation probability AND the sentence-level highlighting; the highlighting is more informative than the single-number verdict because it shows where the detector saw the strongest signal.
If the text is by a non-native English author, treat any "AI-generated" verdict as a likely false positive per the Stanford 2023 finding; the toolkit's editorial position is that you cannot publish from this verdict.
Cross-check against Pangram for a second-opinion detector signal (counted as one detector signal class with GPTZero per Anchor 3, not two).
Pair with non-detector verification: ClaimBuster at 2B.2; Google Fact Check Explorer at 1B.5; the productive path for textual cases is the non-detector chain, not a second detector pass.

In the toolkit's workflow¶

Cautious-detector pillar tool per Architectural Anchor 1. Sits as alternative at 1B.2 AI text detection: the most-documented detector and the clearest cautionary case in the toolkit's text section.

Standard combinations:

With Pangram at 1B.2 as the primary; the two together count as one detector signal class under Anchor 3, not two, despite being from different vendors with different design philosophies.
With ClaimBuster at 2B.2 – the productive non-detector path; if a publishable case exists for the text, it comes from the claim-extraction and fact-check-database side, not from a stronger detector verdict.
With Google Fact Check Explorer at 1B.5 to look up whether the claim has already been fact-checked.
With Dissect at 2B.2 for Sinhala-specific work in Sri Lanka where the detector path is structurally inapplicable.

This tool's verdict is one weak signal class per Anchor 1; never publish a binary claim from it alone. Counted as one detector signal class under Anchor 3 when paired with another text detector. The toolkit's editorial position on GPTZero is that the value is pedagogical: it shows what the AI-text-detector class can and cannot do, more than it is operational.

The productive cross-reference is to 2B.2 AI claim extraction as the secondary-signal cell: when a textual case lands and the question is whether the writing was AI-generated, the publishable path runs through ClaimBuster, Dissect (Sinhala), Yudistira (Bahasa), or Alegre (multilingual) alongside the detector reading, not from the detector reading alone. The 1B.2 verdict travels with the 2B.2 evidence; on its own it does not carry.

Decision-tree references: T5 escalation – GPTZero is the alternative cautionary-case AI-text detector option at T5.9 (text professional), where the framing applies in full: AI- text detection is only a style flag, never a publishable verdict on its own. The binding "do not use as standalone evidence" framing applies at every publish-node check.

Conflict resolution behaviour¶

When GPTZero's verdict disagrees with Pangram's verdict on the same text, GPTZero carries lower evidence weight by default because Pangram's design philosophy targets false-positive minimisation while GPTZero has documented high false-positive rates on non-native English (Stanford 61%); the disagreement is more likely to indicate a Pangram false negative than a GPTZero true positive. The toolkit's editorial response is to treat both verdicts as one detector signal class (Anchor 3), record the disagreement, and route the case to non-detector verification (claim extraction, source reliability) before any publication. When GPTZero disagrees with a non-detector signal (claim already fact-checked, source confirmed authentic, ESL author identified), the non-detector signal wins because Anchor 2 mandates two non-detector signals before any strong public claim.

Override notes¶

Override notes

Verbatim Stanford ESL finding (c2-verbatim-Stanford-non-native-FPR): the Limitations admonition carries the Stanford 2023 finding verbatim ("up to 61% false positive rate on non-native English writing") and the Perkins findings (26.4% accuracy on non-manipulated AI text and 16.7% on manipulated text) verbatim. The toolkit does not soften these into "performs less well on non-native English."
SEA-language disclaimer (b2-disclaimer-pass): the card enters with the "validate before regional use" disclaimer per Group B2; SEA-language performance is unverified and the Limitations section names the structural risk verbatim.
Weak-signal framing (e4-weak-signal-framing-binding): the workflow section above invokes the detector-as-weak-signal sentence per P3 sub-routine; never publish a binary claim from GPTZero alone. The Conflict resolution paragraph names Pangram as the better-design alternative when both run on the same text.
Detector-only caveat: GPTZero serves the cautious-detector pillar only. Every output is a detector signal; the card frames the workflow accordingly.

Sources¶

GPTZero. GPTZero — AI text detection platform. GPTZero, 2024. gptzero.me.
Perkins, M. et al. Game of Tones: Faculty detection of GPT-4 generated text in university assessments. Assessment & Evaluation in Higher Education, 2024. doi.org/10.1080/02602938.2023.2241676. (26.4% accuracy on non-manipulated AI text; 16.7% on manipulated text — cited in card)
Liang, W. et al. GPT Detectors are Biased Against Non-Native English Writers. Stanford HAI, 2023. arxiv.org/abs/2304.02819. (61% false-positive rate on non-native English — cited in card)