An AI content detector that estimates whether text or an image was AI-generated — paste writing or upload a picture and it returns a likelihood score, highlights the AI-looking passages, and flags AI "humanizer" edits, with a browser extension that labels feeds on X, LinkedIn, Substack, Reddit, and Medium in real time.
Last verified · 2026-07-29 · by Moe Ameen
Pangram is an AI-detection platform built by Max Spero and Bradley Emi, two Stanford AI graduates, and based in New York. It answers one question: does this text — or this image — look machine-generated? You paste writing or upload a picture and Pangram returns a probability that it was produced or heavily assisted by AI, along with a segment-level view that highlights which passages read as AI rather than scoring the piece as a single number. Its text model is versioned (Pangram 4 as of mid-2026), and an image detector shipped as a research preview around July 29, 2026.
The approach is what separates it from the first wave of detectors. Spero has said the tool does not rely on copy-paste metadata or hidden watermarks; instead it uses a "synthetic mirror" method — it takes your text, has a frontier model generate a same-topic, same-length, same-tone AI version, and then studies the stylistic gap between the two. The models are trained on tens of millions of documents known to be human-written. Pangram advertises accuracy above 99% and a false-positive rate of roughly 1 in 10,000 human documents, and its numbers have been examined by outside researchers (it has cited work from the University of Chicago and the University of Maryland and a top placement on the COLING 2025 detection benchmark). Those are still strong claims about a hard problem, so treat any single score as an estimate, not a verdict.
You reach it a few ways: a web app, a Chrome extension that labels AI-looking posts in your feed on X, LinkedIn, Substack, Reddit, and Medium, a Google Docs add-on, an API, and LMS integrations for schools. It also detects AI "humanizer" tools — the paraphrasers people run to launder AI text past detectors — and can spot AI images dropped into otherwise real photos. Substack integrated Pangram in July 2026 to power its reader-facing "Scan for AI text" feature.
Two honest caveats. First, no detector is infallible: a ~1-in-10,000 false-positive rate is excellent but not zero, and distinctive human writing occasionally trips detectors, which is why the fair use of a tool like this is as a signal, not a court. Second, Pangram detects; it does not create, rewrite, or publish anything. It tells you how a piece reads — the work of producing content that reads as genuinely yours sits entirely elsewhere.
Pangram is most useful to a creator as a proofing step, not a gatekeeper — and the honest way to use it is to make your AI-assisted work genuinely yours and then check it, not to play cat-and-mouse with the scanner. That's the exact loop Kompozy is built to feed. Kompozy is a content generation and publishing engine: when it drafts a Blog Article, an Email Newsletter, or Text Posts, the Persona Brief governs the voice and a banned-word filter strips the generic AI-tell phrasing — the "in today's landscape," the rule-of-three padding, the flat superlatives — that both readers and detectors react to. The practical workflow is generate-in-Kompozy, edit into your own voice, then paste the draft into Pangram (or scan it in-line via the extension) before you publish. Where a passage scores high, you rewrite that specific segment — Pangram's segment highlights tell you which one — instead of guessing.
The bigger reason the two pair well is that Pangram only reads one modality at a time and Kompozy produces the rest. A text detector says nothing about a Persona Short, a HeyGen avatar clip, a Carousel, Quote Graphics, or a Clipped Short — formats Kompozy generates and then schedules and publishes across nine platforms from one review pipeline. So instead of a text-only presence that lives or dies on one authenticity score, you run a multi-format operation where your written work is disclosed and voice-checked with Pangram, and everything else ships in parallel. Detector as quality gate, Kompozy as the studio behind it.
Pangram is an AI content detector from a New York startup founded by Stanford AI graduates Max Spero and Bradley Emi. You paste text or upload an image and it estimates how likely the content is AI-generated, highlights the AI-looking passages, and flags "humanizer" paraphrasing. It offers a web app, a browser extension, a Google Docs add-on, an API, and LMS integrations, and it powers Substack's "Scan for AI text" feature.
Pangram advertises accuracy above 99% and a false-positive rate of roughly 1 in 10,000 human documents, with results examined by outside researchers including the University of Chicago and the University of Maryland and a strong placement on the COLING 2025 benchmark. Those are strong figures for a hard problem, but no detector is perfect — treat a single score as an estimate, and remember distinctive human writing can occasionally be flagged.
Its co-founder has said it does not rely on metadata or hidden watermarks. Instead it uses a "synthetic mirror" method: it generates a same-topic, same-length, same-tone AI version of your text with a frontier model, then analyzes the stylistic differences, drawing on training over tens of millions of known-human documents. That is why it can also detect text run through AI "humanizer" paraphrasers.
Yes — Pangram shipped an image detector as a research preview around July 29, 2026. Rather than depending on a single provider's watermark, it analyzes pixel-level statistical patterns, which lets it flag AI images across different generators and even spot AI elements composited into an otherwise real photo. As a preview, treat its image results as evolving.
It is useful as a pre-publish self-check, not a gate. The productive approach is to make AI-assisted drafts genuinely your own — a tool like Kompozy governs the draft with a Persona Brief and a banned-word filter to cut generic AI-tell — then scan the finished text with Pangram, rewrite any high-scoring segments, and disclose your process. Detectors reward a disclosed, distinctive voice; they penalize anonymous, flattened AI text.