Short answer: Pangram launched Pangram Image 1.0 on 4 November 2026 and reports 99.96% accuracy on an evaluation it built itself: 100 real photographs, each mirrored by 12 frontier image generators, for 1,200 synthetic images. That figure describes that test set. It does not describe your exhibit, which may be cropped, compressed or edited. If a detector’s output is put forward in a dispute, ask what was tested, the error rate on real but edited photos, the model version and what the result means. Keep the exact image, the report and the date you ran it. A TimeBind dates those files. It does not detect fakes and it does not decide whether an image is genuine.
Last checked 13 November 2026. General information, not legal or security advice. This page is for lawyers and their clients. It does not say any detector is wrong; it lists questions to ask. The figures come from the vendor’s own published tests, and we have not seen independent testing of them.
Table of Contents
What was claimed, and on what test
Pangram’s launch post says Pangram Image 1.0 produces heatmaps showing where AI content appears, is trained to be robust to compression and editing, and now accepts images down to 256×256, with a reduced-confidence disclaimer below 512 pixels. Its headline result is 99.96% on a set it created: 100 real photos mirrored with prompts matching the originals using 12 frontier generators. It also reports 98.7% for its video model on an academic benchmark called DF26. Heise online and Engadget reported the launch.
Those are the vendor’s own published figures. A vendor benchmark is a starting point, not a finding about your file.
Questions counsel can put to a detector report
- What was the test population? Which generators, which kinds of image, and were real photographs that had been screenshotted, recompressed or lightly retouched included?
- What is the false-positive rate on real photographs, and how was it measured? An accurate-looking headline can still mislabel a genuine photo.
- Which model version and settings produced this report, and on what date? Detectors are updated.
- What does a score or a heatmap mean? Is it a probability, a ranking or a yes/no? Who set the threshold?
- Was the input the original file or a copy? Social platforms recompress images, and Pangram itself warns about low resolution.
- Who ran the test, and can the other side repeat it?
Spotify made a related point about its own use of detection tools in its 20 October 2026 update: “AI detection technology in particular has limitations, and it will only get harder as generative AI evolves,” which is why it combines several signals.
Where this fits in an evidence argument
General information only: in US federal courts, expert opinion is screened for reliability under Rule 702 and the case law around it, which looks at the basis and the reliability of the methods. A detector’s score is a tool’s output, and how it is used is a question for counsel and the court. The piece on being told your evidence is a deepfake and deepfake evidence rules in the US, England and the EU set out the wider picture.
What to keep if a detector report is in play
- The exact image file as you received it and any earlier original, with a SHA-256 fingerprint of each.
- The detector’s full output, its version, the settings and the date and time you ran it.
- A custody note: who held the file and when. The chain of custody log for digital exhibits and the deepfake evidence authentication record checklist give a format.
- A TimeBind of the folder. It is a dated, checkable record that those exact files existed no later than the recorded time and still match. It does not make the detector right or wrong.
For readers who want background on why detectors can disagree, see are AI detectors accurate?.
Why people are searching for this
Measured on 13 November 2026: a post about the launch had 2,411 impressions on 10 November and another had 1,028 on 4 November. Trade press covered the launch. Demand here is moderate and tied to a news event; the legal angle is our own framing.
Frequently Asked Questions
-
Is Pangram’s 99.96% figure independently verified?
The figure comes from Pangram’s own evaluation of 1,200 synthetic images mirroring 100 real photos. We have not seen independent testing of it, so treat it as a vendor claim about that test set.
-
Can a detector prove an image is AI-generated?
A detector produces a score or a heatmap. What weight that carries depends on the method, the data and the context, and that is for counsel and the court to argue. It is not a TimeBind question.
-
Does a TimeBind detect AI images?
No. A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. It does not examine what is in the image.
-
What should we preserve when a detector is used?
The exact image, any original, the detector’s full output with version and date, a custody note, and a fingerprint of the folder.
-
Does this page give legal advice?
No. It is general information for lawyers and clients. Check the rules that apply in your jurisdiction.
Sources checked
- Pangram, “Introducing Pangram Image 1.0 and Pangram Image Flash 1.0”: pangram.com
- heise online, “Pangram Image launches to detect AI images and videos”: heise.de
- Engadget, “Pangram expands its AI detection tools to video and images”: engadget.com
- Spotify, “One Year In: Putting Human Artistry First”, 20 Oct 2026: newsroom.spotify.com
- HumanProgress on X, 10 Nov 2026: x.com/HumanProgress/status/2120196063717884049
Date the files behind a detector report
Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.
