7 Numbers That Show How Often AI Detectors Get It Wrong (2026)

TL;DR

This page is a tested guide to spotting AI-generated text, images, and video in 2026, built from primary sources instead of vendor marketing. It includes a cross-detector false-positive matrix computed from the independent BFI/University of Chicago study (Pangram, Originality.ai, GPTZero), a coverage table for C2PA Content Credentials versus Google SynthID, and a calculation of how many false accusations a 10,000-word classroom corpus should expect at each tool's real false-positive rate. It also covers the documented ESL false-positive disparity from the Stanford Patterns study and a 2026 finding that flips the usual advice: humans beat automated detectors at spotting deepfake video, even though machines crush humans on static AI images.

Last verified: August 1, 2026

I failed a Turing test I didn't know I was taking. A friend sent me a comforting text about a mutual friend's breakup, and I spent ten minutes deciding whether a human or a chatbot wrote the kind part. I was wrong. It doesn't matter which way.

That's the real state of AI detection in 2026, not the version where you count fingers on a hand and feel clever. Most of the advice floating around isn't tested against anything, so I read the primary studies, pulled the raw tables, and did the arithmetic competitors skip. Detection scores are also only half the question. The other half is whether the work carries a human's judgment at all, which is the line I drew in AI-assisted craft vs AI slop.

---

---

What's Inside

• Which AI text detector has the lowest false-positive rate, based on independent (not vendor-published) data?
• How many human-written essays get wrongly flagged as AI in a real 10,000-word classroom batch?
• What does Google's SynthID watermark cover, and where does it fail?
• Does C2PA Content Credentials survive a screenshot or a crop?
• Are non-native English writers still more likely to get falsely flagged as AI?
• Is a human or a machine better at spotting a deepfake video?

The Cross-Detector False-Positive Matrix Built From Independent Data

I went looking for one table that compares AI text detectors using independent data instead of vendor self-reports. Across the eight competitor pages I read on this topic, none assembled one: aidetectors.io ranks itself first in its own single-source table, and Originality.ai and Compilatio cite their own or unreproduced figures with no side-by-side view. So I built one from the one fully independent, peer-institution-run benchmark I could find: the BFI/University of Chicago working paper by Jabarian and Imas, published August 2025. It tested Pangram, Originality.ai, GPTZero, and an open-source RoBERTa baseline against 1,992 matched human and AI passages across six genres, including Amazon reviews, news, and résumés.

I averaged the paper's six per-genre numbers by hand, since the paper itself never aggregates them. Here's what the math produces:

Pangram averaged a 0.22% false-positive rate across the six genres in this independent test, and Originality.ai and GPTZero landed at 0.75% and 0.89%. Those numbers sound tiny until you multiply them by a classroom. The open-source RoBERTa baseline misclassified human writing as AI in nearly two out of three news articles. A coin flip with extra steps.

The same paper caught something most vendor blog posts leave out: run the AI text through a humanizer tool first, and GPTZero's false-negative rate climbs to 25-94% depending on genre, Originality.ai's to 2-21%. Pangram stayed near 0-5% (BFI Working Paper 2025-116). Know whether a detector's numbers survived a paraphrasing pass before you trust them.

What 10,000 Words of Human Writing Costs You in False Flags

Take a 10,000-word corpus split into 500-word documents, twenty total, roughly a class set of essays. Apply each detector's independently measured average false-positive rate from the BFI study.

I ran the same math against the higher, more real-world-skewed numbers from a commercial benchmark published by aidetectors.io, which tested ten detectors on 500 samples. Worth flagging up front: aidetectors.io ranked its own product first in its own study. Treat it as one data point, not a verdict.

The gulf between "under one false accusation across 100+ students" and "nearly three false accusations in every class of 20" comes down entirely to which benchmark you trust. Not a footnote-level detail. That's the difference between a tool you can trust and one that ends someone's semester over a coincidence.

SynthID vs C2PA: What Each One Covers

I kept seeing "SynthID and C2PA solve this" tossed around like the two are interchangeable. They're not.

I read DeepMind's own scope statement twice because it contradicts secondhand coverage claiming SynthID already handles text and audio in production. Google's official SynthID page confirms coverage for images and video segments only, with no production claim for text or audio. The broader claims describe research implementations, not a shipped feature.

Neither system covers content made outside a participating tool. A photo from a non-signing camera, or an AI image from a generator that skips SynthID, carries no signal either way. Both also die the same death: a screenshot repaints pixels and strips metadata. If someone tells you a screenshot "still has the watermark," that's wrong on both counts.

The ESL Disparity That's Still Unresolved

The single most important thing I found in this research isn't a number. It's an open question that Pangram's and Copyleaks' own reassuring vendor numbers skip past instead of answering.

The original finding comes from Liang, Yuksekgonul, Mao, Wu and Zou's 2023 Stanford study in *Patterns*. Researchers ran 91 real TOEFL essays by non-native English speakers through seven GPT-era detectors. 61.22% of those genuinely human-written essays got flagged as AI-generated, on average, across the seven detectors. Eighteen of the ninety-one were flagged unanimously by all seven.

Asked to make the same essays "sound more like a native speaker," ChatGPT's rewrite dropped the false-positive rate from 61.22% to 11.77%. Run the logic backward: simplify a native speaker's essay to sound less fluent, and its false-flag rate jumped from 5.19% to 56.65%. The bias tracks sentence simplicity, not authorship.

That study used detectors now three generations old. Two vendors have since published newer numbers on their own current models. Copyleaks reports 99.84% accuracy across 7,482 non-native English texts, and Pangram reports a 0.012% false-positive rate across 25,021 ESL examples.

Both numbers look reassuring. Both are also vendor self-reports on the vendor's own model, not an independent replication of the Stanford methodology.

[UNCERTAIN]: no independent, non-vendor study has re-run the Liang et al. methodology against current-generation commercial detectors as of this writing. The 2023 disparity may or may not still apply industry-wide, and I'm not going to pretend the vendor numbers settle that. If you're grading, hiring, or moderating based on a detector score, treat any flag on a non-native speaker's writing as a conversation starter, not a verdict.

The Modality Where Humans Still Beat the Machines

Here's the twist that undoes most of the confident advice in this space, including some of my own assumptions going in.

A University of Florida study published in Cognitive Research: Principles and Implications, January 2026 tested people and algorithms on static deepfake images and short deepfake videos. On still images, humans performed at roughly chance, about 50%, while the tested algorithms hit up to 97% accuracy. Machines are built for pixel-level pattern spotting. We're not.

Then the video results flipped it. Algorithms dropped to near-chance on the deepfake videos. Humans correctly identified deepfake video around 63-67% of the time, meaningfully better than a coin toss and better than the machines tested alongside them. UF's department summary cites slightly different exact figures, so treat the precise percentage as approximate, but the direction holds across both summaries.

That result contradicts advice repeated by Malwarebytes and BBC Bitesize, both of which suggest visual inspection has stopped working. For static images, fair enough. For video specifically, in this dataset, looking closer still measurably worked.

On text, humans don't fare much better than on images. A controlled Turing test by Jones, Rathi, Taylor and Bergen, covered by UC San Diego in May 2026, found GPT-4.5 with a humanlike persona prompt was judged "human" 73% of the time, more often than the real human participants in the same conversations. LLaMA-3.1-405B scored 56%, statistically indistinguishable from a person. If trained researchers judging live conversations get fooled at that rate, the "I can just tell" instinct deserves some humility.

What This Means for How You Check Something

For text: don't lean on a single detector score, especially against a non-native English speaker's writing. Pangram's independent-study performance (0.22% average FPR, still near-zero even against humanizer tools) makes it the strongest of the three BFI tested, but "strongest" isn't "certain." Ask for drafts or revision history before treating any score as proof.

For images: provenance beats pixel-peeping now. Check for Content Credentials or a SynthID label through Google's own surfaces (Search's "About this image," Gemini, Lens) before squinting at hands or shadows. Absence of a watermark proves nothing either way.

For video: trust your eyes a little more than the current tooling. Automated deepfake video detection genuinely trails human judgment right now, so odd blinking, mismatched audio timing, or unnatural head movement are still worth your attention.

For anything high-stakes: reverse image search (Google Lens, TinEye, Bing Visual Search) still works regardless of which generator made the content, because it checks for prior existence, not artifacts.

All four habits are the same underlying skill: refusing to outsource your judgment to a single score. That skill has a name and a practice routine, and I teach it as critical AI literacy on the newsletter.

FAQ

Which AI detector has the lowest false-positive rate?

I keep coming back to the BFI study because it's the only fully independent benchmark that tested more than one commercial detector side by side. In that data, Pangram posted the lowest average false-positive rate at 0.22% across six genres, with Originality.ai at 0.75% and GPTZero at 0.89% (BFI Working Paper 2025-116). That doesn't mean Pangram is flawless. It means it's the best-performing option in the one dataset built by people with nothing to sell.

Can AI detectors tell if non-native English speakers wrote something?

This is the question that made me the most uneasy while researching this page. The original 2023 Stanford study found 61.22% of genuine TOEFL essays by non-native speakers got flagged as AI-generated across seven detectors (Liang et al., *Patterns*). Vendors like Pangram and Copyleaks now report near-zero disparity on their current models, but I couldn't find an independent re-test of those newer models anywhere outside the vendors' own reports, so I'm treating it as unresolved rather than fixed.

Does SynthID work on screenshots?

No, and I wish more people knew why before sharing a "watermarked" screenshot as proof of anything. A screenshot re-renders pixels from scratch, which destroys SynthID's invisible watermark the same way it strips C2PA's metadata. Neither provenance system survives that action (DeepMind SynthID page, Editors Weblog C2PA tracker).

Are humans or machines better at spotting deepfakes?

Depends entirely on the format, and I didn't expect the answer to split this cleanly. On static AI images, machines win decisively, hitting up to 97% accuracy against roughly 50% for humans. On deepfake video, that flips: humans landed around 63-67% while the tested algorithms fell to near-chance (University of Florida, Cognitive Research, Jan 2026). Know which one you're looking at before you decide whether to trust your gut or a tool.

Is a 99% accuracy claim from a detector vendor trustworthy?

Read it as a starting point, not a finish line. Vendor-published numbers like Originality.ai's 99%+ claim (Originality.ai) are usually measured on the vendor's own benchmark, on the vendor's own terms, without independent replication. The FTC has already taken enforcement action against one detector maker (Workado/BrandWell) for an unsupported 98% accuracy claim, which is exactly the kind of number worth double-checking before you act on it.

---

Detection is a moving target, and every number on this page has a date attached for a reason. If you've found a detector or a provenance check that held up (or fell apart) in your own testing, I'd genuinely like to hear about it. And if you've been wrongly flagged by one of these tools, that story is useful too 🤗