Accuracy
What TrueKnow was tested on and what the numbers mean.
TrueKnow is tested on our own frozen sets. This is not an outside lab study. Results below split by cohort. The cohorts are not combined. A 100% catch rate on images already in the database says nothing about novel detection, so they are reported separately.
What "most thorough" means
When TrueKnow says "most thorough," that is a claim about methodology -- specifically the number and variety of independent checks: file-level AI generator signatures, content credentials, archive matching, compression and texture forensics, independent AI classifiers, and reverse image search. Twelve checks in one free run, including ones enterprise tools charge for separately.
That is a different axis from detection rate. The data below measures false positive rate on real photos -- whether TrueKnow ever calls a real photo AI generated. We have not run a head-to-head detection rate comparison against other tools and do not claim to outperform them on that metric. Novel-fake detection rate is a third axis, and that row is not published yet.
Results by cohort
| Set | N | Called AI generated | Called Real | Uncertain |
|---|---|---|---|---|
| Real personal photos | 13 | 0 | 0 | 13 |
| Known fakes already in the database | 17 | 17 | 0 | 0 |
| New fakes not in the database | not published yet | -- | -- | -- |
We would rather leave a new fake Uncertain than call a real person fake. That is why the new-fake row will not look like 100 percent once it is published.
Real photos: false positive rate
0 of 13 real personal photos were called AI generated. That is a 0% false positive rate on this set. The same gate runs automatically on every code change, and it has held at zero across all runs.
In both the full evaluation run and the automated CI regression, all 13 returned Uncertain -- never AI generated. Uncertain on a real photo is not wrong; it means the pipeline saw no positive evidence of AI generation. The guarantee is that no real photo gets a false AI generated verdict.
Known fakes: recognition, not novel detection
The 17 known fakes are all exact matches in TrueKnow's recognition database. The pipeline finds them by perceptual hash, not by novel analysis. A 100% catch rate on known items is expected. It says nothing about how the pipeline performs on images it has never seen.
The set includes 6 Gemini-generated images seeded at launch plus 11 images from the original evaluation set, all of which have since been added to the database.
Novel detection: held-out set not published yet
The honest novel-detection row requires AI images that are not in the recognition database at test time. That set is not published yet.
Video: what is actually checked
TrueKnow analyzes the first 30 seconds of a video from the start time you pick. Content before or after that window is not analyzed. The result card shows the exact range that was checked.
What this test does not show
- 30 images is a small sample. Results reflect those specific files, not performance across every camera type, AI tool, or compression setting.
- This is not an independent lab study and has not been peer reviewed.
- We have not run a head-to-head detection rate comparison against other tools. "Most thorough" refers to methodology breadth, not detection rate. See the section above.
- A Real verdict is a best assessment, not proof a photo is authentic.
- Uncertain is the honest result when evidence does not clear the bar. It is not a failure.
- Results are estimates, not legal or forensic proof.
Confidence scores
Every result shows a confidence percentage, including Uncertain. This is how strongly the signals lean in one direction. It is not a probability of truth. A 62% Uncertain result means the signals lean slightly one way but not enough to make a confident call.
Contact
Questions about accuracy or methodology: lee@trueknowapp.com.
Run date: September 8, 2026. Set: golden v1 (30 images). Pipeline: 2026.08.26. Source: data/eval/regression/golden_v1_2026-09-08.json in the repository.
Terms · Privacy · How it works · FAQ · About · For business · Built and Maintained By TrueKnowApp.com