Pangram’s November 4, 2026 release introduces Image 1.0 for AI image detection and Image Flash 1.0 for video analysis. For editors evaluating the tools, a useful first exercise is to compare an original file with the version their publishing workflow actually produces.
Record which file received the result
Pangram says Image 1.0 provides a heatmap of suspected AI regions, while Flash supports frame-by-frame video analysis. Its announcement lowers the minimum image size to 256 × 256 pixels and says results below 512 pixels come with a reduced-confidence disclaimer. These are supplier descriptions; no Pangram account or detector was tested for this article.
A result needs a file identity beside it. Keep the original filename, dimensions and an internal sample identifier. If an editor also checks a resized export, give that export its own row. Otherwise, a later reviewer may mistakenly attach the export’s result to the source file.
Build a small, labeled evaluation set
For a proposed internal trial, use material whose creation history your team can document: camera photographs, deliberately generated images and examples with a known edited region. Keep the reference labels separate from the detector output. Obtain permission for any material uploaded and check the service’s current handling terms first.
- Original: retain the source file and documented creation history.
- Export: record the actual resizing or compression performed by your workflow.
- Result: save the model name, date, output and any confidence warning.
- Review: note whether a heatmap overlaps the known edited area and what still needs human inspection.
Test the same source and its export as a pair, but retain both results. If they disagree, flag that pair for review rather than choosing the answer you expected. This is a proposed evaluation method, not a claim about Pangram’s behavior on any particular file.
Keep errors visible
Imagine an entirely fictional test with 100 documented camera photographs. If two are flagged as AI, the false-positive rate for that set is 2 ÷ 100 = 2%. That says nothing yet about missed AI images: those require a separate set of known generated or edited examples. Retain counts as well as percentages.
A local trial also cannot establish a universal accuracy rate. Its value is narrower: it shows how a particular model performed on your documented samples and transformations at a recorded time. Repeat relevant checks when the model or export process changes.
Use provenance alongside detection
C2PA’s explainer describes Content Credentials as records of origin and editing history. It says provenance alone cannot establish factual truth, and missing credentials should not automatically determine trustworthiness. A detector result and a provenance record therefore answer different questions.
For a publication decision, retain the available source history, detector output and editorial verification together. An AI label does not establish whether an illustration is accurately disclosed, and a camera photograph can still be paired with an inaccurate caption. Keep the final decision with the responsible editor.
Check the evaluation record
Primary sources
Pangram’s November 4 Image 1.0 and Image Flash 1.0 announcement