Skip to content

Pangram AI

FreemiumProfessional

Vendor: pangram.com | Type: Detector | Cost: Freemium

TL;DR

A vendor AI text detector that markets fine-grained structural sentence analysis aimed at minimising false positives, with documented support for 20+ languages including Arabic, Chinese, French, and Spanish. The toolkit's primary 1B.2 entry. Every entry in this cell is a cautionary case, not a recommendation.

What it does

Pangram accepts a text input and returns an AI-generation classification through structural sentence-level analysis. The vendor's design philosophy (articulated in product materials and in the V2 2026 release) is strict thresholds intended to keep false positives low; the trade-off is that humanised AI text is more likely to slip through as a false negative than to be falsely flagged as machine-generated. For a fact-checker triaging suspect text under the toolkit's Anchor 2 (no strong public claim from detector signals alone), the false-positive minimisation is operationally important: a false positive on a public figure's genuine writing is a defamation risk in any of the focus countries.

The toolkit ships Pangram as the primary entry in 1B.2 specifically because of this design discipline. It does NOT ship it because the AI-text-detector class is reliable; independent evidence is unanimous that the class is unreliable for high-stakes use across the board, and the cell-level honest gap names this in full.

When to use it

  • A piece of suspect text needs a directional read; you want a vendor that prioritises false-positive minimisation over false-negative minimisation.
  • A regional case involves text in one of Pangram's supported languages and the multi-lingual coverage matters more than GPTZero's English-optimised baseline.
  • A workshop demonstrates the difference between a strict-threshold vendor (Pangram) and a more permissive vendor (GPTZero) on the same input.
  • An API integration into a newsroom pipeline needs a detector that is engineered against false-positive risk; pair the API output with non-detector verification (claim extraction at 2B.2, source-reliability at 1B.5) before any publication.

Independent accuracy

Vendor claim vs independent assessment

Vendor claim: Pangram publicly markets vendor reports 99%+ accuracy; no independent SEA-specific benchmark identified. Coverage across many languages, including Arabic, Chinese, French, and Spanish.

No independent SEA-specific benchmark identified as of May 2026. independent assessment searches located no peer-reviewed or civil-society audit that tested Pangram on tonal, low-resource, or SEA-language inputs. The vendor markets multilingual support; no independent validation for SEA languages has been published in the higher-priority sources searched.

Limitations

Limitations

  • Strict thresholds miss humanised AI text; false negatives are the documented failure mode of the design philosophy.
  • SEA language performance is vendor-claimed but independently-untested; treat the 20+ language coverage as vendor-claim until independent benchmarks appear.
  • The text detector class as a whole is unreliable for high-stakes use per independent benchmark evidence (Turnitin 0.61, Originality 0.69, GPTZero 26.4%/16.7%, Copyleaks 73.9% with 50% FP on small human-control). Pangram's design choice may shift the false-positive / false-negative balance, but does not change the class-level reliability ceiling.

Privacy and threat model

Pangram is a vendor web service; the input text uploads to Pangram's cloud infrastructure for analysis. Pangram 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: Bahasa Indonesia is plausibly within the 20+ language coverage; not independently verified for Bahasa.
  • Laos: Lao is unlikely to be in the supported set; no independent verification.
  • Malaysia: Malay is plausibly within the 20+ language coverage; not independently verified for Malay.
  • Philippines: Filipino / Tagalog is plausibly within the 20+ language coverage; not independently verified for Filipino.
  • Sri Lanka: Sinhala and Tamil coverage is unverified; no independent benchmark exists.
  • Thailand: Thai is plausibly within the 20+ language coverage; not independently verified for Thai.

The toolkit's editorial position is that the multi-language vendor claim does not substitute for independent SEA-specific benchmarking; the Limitations section names this gap explicitly; coverage is not implied.

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

How to access

The web interface is at pangram.com (basic free tier); API access documented at the developer portal. Account creation required for API access and for higher-volume free-tier use.

Cost (current as of 2026-05)

Free basic tier sufficient for low-volume triage; Premium tier paid on a per-seat or per-volume basis. Verify pricing directly before any institutional integration as vendor pricing may have shifted since this card's last verification date.

Quickstart

  1. Open pangram.com in your browser.
  2. Paste the suspect text into the analysis box.
  3. Read the AI-generation classification; treat the result as one weak signal class per Anchor 1.
  4. If the classification is "AI-generated" with high confidence on a text in a SEA language, do NOT publish a binary claim from this alone; the SEA-language performance is vendor-claimed but independently-untested.
  5. Cross-check against GPTZero for a second-opinion detector signal (counted as one detector signal class with Pangram per Anchor 3, not two).
  6. Pair with non-detector verification: ClaimBuster or Dissect at 2B.2 for claim extraction; Google Fact Check Explorer at 1B.5 for source-reliability cross-check.

In the toolkit's workflow

Cautious-detector pillar tool per Architectural Anchor 1. Sits as primary at 1B.2 AI text detection. The cell is structurally thin because the detector class is unreliable; Pangram's primary slot reflects design discipline (low FPR), not class reliability.

Standard combinations:

  • With GPTZero at 1B.2 as the alternative; the two together count as one detector signal class under Anchor 3, not two.
  • With ClaimBuster at 2B.2 for non-detector claim extraction – the productive pair for textual cases per the cell- level bridging note.
  • With Google Fact Check Explorer at 1B.5 for source-reliability and existing-fact-check cross-check.
  • With Dissect at 2B.2 for Sinhala-specific claim extraction when the case is in Sri Lanka.

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 productive cross-reference is to 2B.2 AI claim extraction as the secondary-signal cell: when a tipline-routed text claim reaches the desk, a Pangram verdict sits beside ClaimBuster extraction, Yudistira (Bahasa) or Dissect (Sinhala) clustering, and Alegre's multilingual fallback. The detector reading travels with those non-detector signals, never apart from them.

Decision-tree references: T5 escalation – Pangram is the primary AI-text detector option at T5.9 (text professional), where T5.9's 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 across T1, T2, and T3.

Conflict resolution behaviour

When Pangram's verdict disagrees with GPTZero's verdict on the same text, neither carries decisive evidence weight on its own; the class-level evidence is that the AI-text-detector category is unreliable for high-stakes use across the board. The toolkit's editorial response to such disagreements 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, fact-check database hit) before any publication. When Pangram disagrees with a non-detector signal (claim already fact-checked elsewhere, source confirmed authentic), the non-detector signal wins because Anchor 2 mandates two non-detector signals before any strong public claim.

Override notes

Override notes

  • Vendor wrapping pair (c1-pair-vendor99-no-indep): Pangram's 99%+ vendor claim is paired against the explicit "no independent SEA-specific benchmark identified" sentence in the Independent Accuracy admonition. The pair is honest; the vendor headline travels with the gap acknowledgement.

  • SEA-language disclaimer (b2-disclaimer-pass-no-SEA-indep): the Limitations admonition records that 20+ language coverage is vendor-claimed and SEA-language performance is independently-untested. The card enters with the "validate before regional use" disclaimer.

  • Weak-signal framing (e4-weak-signal-framing-required): the workflow section above invokes the detector-as-weak-signal sentence per P3 sub-routine; never publish a binary claim from Pangram alone.

  • Detector-only caveat (g1-detector-only-caveat): Pangram serves the cautious-detector pillar only. Every output is a detector signal; the card frames the workflow accordingly.

Sources

  • Pangram Labs. Pangram AI — AI text detection with low false-positive design. Pangram Labs, 2026 (V2). pangram.com.
  • Note: No independent SEA-language benchmark identified as of May 2026 for Pangram AI. Per editorial discipline, vendor self-claim wrapped with explicit field-context. See editorial patterns for vendor wrapping pattern.