Pangram Review 2026: Can an AI Detector Actually Prove Innocence?

Pangram Labs has spent the last couple of years quietly winning independent university studies rather than just its own marketing copy, building a reputation in the academic-integrity and publisher niche well before most people had heard the name.

Which makes now a good time to separate what Pangram says about itself from what outside researchers have measured. Those two numbers aren't the same, and the gap matters a great deal if you're on either end of a false accusation.

This is not a hands-on review. We have not run test documents through Pangram ourselves, and we're not going to claim a first-person accuracy number we didn't measure. What follows is a documented-capability breakdown: what Pangram says about itself, what it costs, and — the part that matters most if you're actually deciding whether to trust or fear this category of tool — what independent, named researchers found when they tested it against real detectors, including Pangram's own competitors.

A note on sourcing: pricing and product claims below were fetch-verified directly against pangram.com's homepage and pricing page on July 30, 2026. Every accuracy or false-positive figure is labeled as either Pangram's own published claim or an independently published, named study — we do not blend the two or repeat a vendor number as if it were third-party-verified. Prices and claims move; confirm current numbers on pangram.com before you buy.

What Pangram actually is

Per its own site, Pangram is a text and (newly, in research preview) image AI-detection product built on what it describes as "proprietary technology with years of research" — trained on "diverse datasets, hard negative mining, and active learning." The company frames that as a deliberate departure from the perplexity-based approach older detectors use, which, as we'll get to, is a documented source of bias against certain writers. Pangram states it detects output from "all major language models," including ChatGPT, Claude, Gemini, and Llama, across 20+ languages. It also ships an "AI Assistance Detection" mode aimed at a subtler question than binary human-vs-AI: which segments were AI-edited rather than AI-written.

Pangram's own headline claim is "99.98% accuracy" with a "1 in 10,000" false-positive rate for human-written documents. That is Pangram's number, stated on Pangram's own homepage. The independently-measured number is different, and it's in the next section.

Pricing, fetch-verified July 30, 2026

Plan	Price	Word allowance	What's included
Free	$0	2,000 words/day	3 image scans/day, AI-assistance detection, interpretability features, file upload/OCR, 20+ languages, browser extension, Google Docs integration
Individual	$20/month (annual: save $60)	300,000 words/month	100 image scans/month, plagiarism detection, social media feed scanning, all free features
Professional	$65/month (annual: save $240)	1,500,000 words/month	500 image scans/month, $200/month API credit, plagiarism + social scanning
Team	$20/seat/month (min. 2 seats; annual: save $60/seat)	300,000 words/month/seat	Admin controls, unified billing, member management
Developer API	$25–$1,000+ (tiered)	Usage-based	Pangram 4: $0.05/100 words · Pangram 3: $0.05/1,000 words · 20% bulk discount
Enterprise	Custom	Custom	Progressive discounting, SOC 2 compliance, API access, usage analytics

The free tier is real and daily-recurring — 2,000 words/day, not a one-time trial credit — which is enough to spot-check a single essay or article before paying for anything. Paid tiers scale on raw word volume rather than seats-first, which suits a solo educator or writer better than a per-seat SaaS model would.

Check Pangram's current plans on pangram.com — this is a direct, non-affiliate link. OneClickAI submitted a Pangram affiliate application via PartnerStack on July 30, 2026 (30% recurring commission); see the disclosure at the bottom of this review — it is under review and not yet approved, so nothing on this page earns us a commission today.

What independent researchers actually found

This section matters more than Pangram's own number, and it's where the "which is more accurate" and "how often does it get people wrong" questions get real answers — named studies, not vendor claims recycled as fact.

University of Chicago's Becker Friedman Institute (Brian Jabarian and Alex Imas, Artificial Writing and Automated Detection, NBER Working Paper No. 34223, August 2025) tested Pangram, GPTZero, Originality.ai, and a RoBERTa baseline against 1,992 human texts written before 2020 and 1,992 AI-generated texts across genres and word counts:

Detector (per UChicago's NBER working paper, across tested thresholds)	False-positive rate	False-negative rate
Pangram	0%–0.1%	0.45%–3.8%
GPTZero	~0.7% (steady)	0.2%–3%
Originality.ai	0.1%–0.3%	Worst of the three: up to 30%–42% (paper's own words: "performs worse than both detectors")

Pangram was the only one of the three to meet a stringent policy cap of FPR ≤ 0.5% "without compromising the ability to accurately detect AI text," per the researchers. The same paper also stress-tested resistance to the StealthGPT "humanizer" tool on all three: Pangram detected "nearly 100%" of AI-generated text on longer passages even after humanizing, staying low even as passages got shorter. GPTZero lost the most ground under that pressure — the paper puts its miss rate at "around 0.50 and above across most genres and LLM models" — while Originality.ai's miss rate rose too, but only to a still-serious 5%–21% depending on length, well short of the 50%-plus territory GPTZero fell into.

Worth flagging honestly: UChicago's independently-measured false-positive rate (topping out at 0.1% across the thresholds tested) is roughly 10x higher than Pangram's own headline "1 in 10,000" (0.01%) claim. Both are very low in absolute terms — 0.1% is still one error per thousand documents, not one per hundred — but they aren't the same number, and Pangram's marketing rounds more favorably than the independent test did.

Vrije Universiteit Brussel, in a peer-reviewed study published in the International Journal for Educational Integrity (Springer, June 2026), ran Pangram, GPTZero, Turnitin, and Copyleaks against 160 academic papers over 4,000 words each — an even split of human ESL writing, fully AI-generated, hybrid, and humanized text. The finding, in the researchers' own words: "only [Pangram] produced satisfactory results." Pangram detected 65% of fully AI-written papers and 92.5% of humanized ones, with close to zero false positives. The other three tools were flatly unable to catch fully AI-generated papers — Turnitin, GPTZero, and Copyleaks each scored 0% there. On humanized text they did somewhat better but still weak: Turnitin 50%, Copyleaks 22.5%, and GPTZero the weakest of the four at just 2.5%.

University of Maryland researchers (arXiv:2501.15654, Table 2) tested a humanizer-hardened Pangram configuration ("Pangram Humanizers") at 99.3% overall detection (2.7% false-positive rate) versus GPTZero's 85.3% overall (0.7% false-positive rate). On the hardest single condition in that paper — o1-Pro text run through a humanizer — that same Pangram configuration caught 96.7% versus GPTZero's 46.7%, the same evasion-resistance pattern again. FPR numbers move with the text set being tested (this paper's 2.7% is well above UChicago's sub-0.1%); no single figure is a universal constant. Pangram separately placed in a real COLING 2025 shared-task benchmark (GenAI Content Detection Task 3, built on the RAID dataset): 99.3% on the cross-domain subtask — second place, just behind winner Leidos at 99.4% — and tied for first with Leidos at 97.7% on the adversarial-robustness subtask. That's a genuine third data point, not a repeat of the UMD number, though it's a closer race for Pangram than the "tied for first" framing alone would suggest.

Honest limitations

No detector, including Pangram, should be the sole basis for a disciplinary decision. That's not a Pangram-specific caution — it's the category's hardest-earned lesson. Vanderbilt University disabled Turnitin's AI detector entirely in 2023 after estimating that its own claimed 1% false-positive rate, applied to the roughly 75,000 papers Vanderbilt submits annually, could mean around 750 students wrongly flagged in a single year.

Then there's the bias problem. A widely cited Stanford study (Liang et al., Patterns, 2023) found seven commercial detectors averaged a 61.3% false-positive rate on non-native English TOEFL essays, versus 5.1% on native-English writing — because perplexity-based detection punishes simpler, more predictable vocabulary, which is exactly how many non-native writers write. Pangram's own framing claims its training approach avoids leaning on perplexity for precisely this reason, and the VUB and UChicago numbers are consistent with it faring better than older tools. But "better than a tool with known bias problems" isn't "immune to the underlying risk," and we haven't seen an independent study that stress-tests Pangram against a large non-native-English corpus the way Stanford's did for the previous generation.

Real people have been wrongly accused using this category of tool, even where Pangram wasn't the detector involved. UC Davis senior William Quarterman was flagged by GPTZero on a take-home midterm essay and given a failing grade before clearing his name using Google Docs' edit-history timestamps — one of the clearest documented false-positive incidents on record. Reason enough to treat any detector's output as one input to a conversation rather than a verdict, whichever vendor's logo is on the report.

An independent evasion write-up is worth knowing about. A community technical write-up — not a peer-reviewed study — documented inconsistent results from Pangram on passages under 250 words, and successfully evaded detection by iteratively prompting an AI model to refine a piece toward a target writing style across multiple passes, eventually reaching a "100% human" classification on AI-generated text. That's a narrower, far more effort-intensive attack than running text through an off-the-shelf humanizer (which UChicago's near-100%-detection-vs-GPTZero's-50%-plus-miss-rate result suggests Pangram already handles well). It's still a real, documented gap on short passages and determined iterative evasion.

This category has an active evasion arms race. "Humanizer" tools exist specifically to defeat AI detectors, and UChicago's data shows the gap between Pangram and its competitors under that pressure is large right now. That last phrase is load-bearing. Detection and evasion are racing each other, and every accuracy number in this review has a shelf life.

Who it's for, and who should be careful

A good fit: an educator, publisher, or platform that wants the detector with the strongest independent-study record on both accuracy and false-positive avoidance. Both studies that tested Pangram against named competitors — UChicago and VUB — found it ahead on the metric that protects innocent people from wrongful accusation, not just the one that catches more AI text. The genuinely-free 2,000-words/day tier also makes it a reasonable single-document spot-check even if you never subscribe.

Be careful if: you're planning to use any detector's score, Pangram's included, as the sole and final basis for an academic or employment consequence. Every study cited above — including the ones that flatter Pangram most — measured error rates in the tenths or low single-digit percent. Not zero. At scale, "very rare" false positives still land on real people, and the Quarterman and Vanderbilt cases show what that costs when a score is treated as a verdict instead of one signal among several: a syllabus history, a draft trail, a conversation with the student or writer.

If you want the wider category context — how Pangram stacks up against GPTZero, Originality.ai, and Copyleaks side by side, and how to think about the false-positive question as a buyer rather than as a researcher — see our Best AI Content Detectors 2026 roundup.

Frequently Asked Questions

Is Pangram's "99.98% accuracy" claim independently verified?

Partially, and the independent numbers are close but not identical to Pangram's own. The University of Chicago's Becker Friedman Institute measured Pangram's false-positive rate at 0%–0.1% and its false-negative rate at 0.45%–3.8% — the lowest, most consistent combination of the three tools tested (GPTZero: ~0.7% FPR / 0.2%–3% FNR; Originality.ai: 0.1%–0.3% FPR but, per the paper's own text, the worst FNR of the three at up to 30%–42%) — but 0.1% is roughly 10x higher than Pangram's headline "1 in 10,000" (0.01%) marketing claim. Trust the independent number, and treat both as "very low," not "zero."

Does Pangram actually catch text run through AI "humanizer" tools?

Better than the alternatives tested so far, per two separate independent studies. UChicago found Pangram detected "nearly 100%" of humanized ChatGPT essays on longer passages, while GPTZero's miss rate rose to "around 0.50 and above" under the same evasion pressure (Originality.ai's miss rate rose too, but only to 5%–21%). Separately, University of Maryland researchers measured a humanizer-hardened Pangram configuration at 96.7% on humanized o1-Pro text versus GPTZero's 46.7% in that paper's test. The Vrije Universiteit Brussel study found Pangram was the only one of four tools (alongside GPTZero, Turnitin, Copyleaks) that reliably detected humanized academic papers at all (92.5%) — the other three ranged from Turnitin's 50% down to GPTZero's weakest-of-the-four 2.5%. A separate independent write-up did document a working evasion method on short passages using iterative AI-refinement, so "better than the rest" is not "unbeatable."

Can Pangram accuse someone falsely?

Yes. Every detector in this category can, Pangram included — just at a lower measured rate than its named competitors so far. No independent study has measured a 0% false-positive rate for any detector. If you're using Pangram's output to make a real decision about a person, treat the score as one input, not a verdict, the same as with any detector.

Is there a free way to try Pangram?

Yes. Pangram's free tier allows 2,000 words per day (not a one-time trial credit) plus 3 image scans per day, per pangram.com's pricing page as of July 30, 2026. Enough to spot-check a single document without paying, though it resets daily rather than accumulating.

Does OneClickAI earn a commission from Pangram links on this page?

No, not yet. OneClickAI submitted a Pangram affiliate application via PartnerStack on July 30, 2026 (30% recurring commission), but it is still under review and has not been approved. The links above go directly to pangram.com with no tracking or commission attached. If that changes, we'll update this disclosure and this review.

The Bottom Line

Pangram's "99.98% accuracy, 1-in-10,000 false positive" line is marketing, and the independently-measured numbers from University of Chicago and Vrije Universiteit Brussel researchers are less dramatic. They're also still the best numbers any detector in this category has posted against named competitors in a controlled, published study — on raw accuracy and on the false-positive metric that decides how many innocent people get wrongly flagged. Better detection and fewer false accusations, backed by two independent academic sources rather than one vendor's blog, is a genuinely different claim from what most AI-detector marketing offers.

None of that makes Pangram infallible. Every error rate above is low, not zero, and a documented evasion method exists for short or heavily-iterated passages. If you're using it to decide something real about a real person's academic or professional standing, it's one input in a larger conversation — not the verdict.

Try Pangram's free tier at pangram.com — 2,000 words/day, no payment required; confirm current limits and pricing directly on Pangram's site before upgrading.