Deepfake detection tools have proliferated as fast as the fakes themselves. Finding a reliable free option used to mean navigating academic research papers and running Python scripts. In 2026, several polished web tools have emerged — but quality varies enormously.
We tested seven tools on a 200-image test set: 100 AI-generated images (from Midjourney v7, DALL-E 3, Stable Diffusion XL, and Adobe Firefly) and 100 real photographs from stock libraries. Here's what we found.
The Test Set
To avoid bias, we used:
- 50 Midjourney v7 portraits (photorealistic face images)
- 25 DALL-E 3 scenes (landscapes and environments, no faces)
- 15 Stable Diffusion XL images (mixed subject matter)
- 10 Adobe Firefly images (product-style photography)
- 100 real photographs from Unsplash (licensed for testing)
Results Summary
Aiscern — 84% accuracy overall (86% on full-res, 79% on compressed)
Strongest on Midjourney portraits where frequency-domain analysis catches characteristic artifacts. Weakest on DALL-E 3 landscapes, which produce fewer detectable anomalies. Verdict comes with per-signal breakdown showing exactly what triggered the detection.
Hive Moderation — 81% accuracy (API-based, free tier available)
Strong across all model types but produces only a binary verdict with confidence score, no signal breakdown. Free tier limited to 100 requests/month.
Sensity AI — 78% accuracy (free tier, watermarked results)
Specialized in face-based deepfakes, significantly weaker on AI-generated images without clear human subjects. Free tier adds watermarks to reports, which limits practical use.
FotoForensics — 71% accuracy (free, no account required)
An older tool based on Error Level Analysis (ELA). Effective for detecting heavily edited images but not purpose-built for AI generation. Produces technical outputs that require interpretation.
Illuminarty — 73% accuracy (free tier)
Good interface, decent accuracy, but false positive rate was the highest in our test — flagging 14% of real photographs as AI-generated. Unacceptable for any use case where false accusations matter.
AI or Not — 69% accuracy (free tier with daily limits)
Simple interface, accessible to non-technical users, but accuracy lagged the field in our test. May have improved since our testing; check current benchmarks.
Google's SynthID — Not publicly available
Google's SynthID watermarking tool for their own generated images is not available for general detection use. Included for completeness.
What the Accuracy Numbers Mean
An 84% accuracy rate sounds high, but consider: if you're checking 1,000 images, you'll get roughly 160 wrong answers. The direction of errors matters:
- False negatives (AI image called real): The detector misses a fake. Lower stakes for most use cases.
- False positives (real image called AI): The detector accuses a real photo of being fake. Higher stakes — can falsely impugn legitimate photographers and journalists.
When to Use Multiple Tools
For high-stakes decisions — journalism fact-checking, legal evidence, academic misconduct proceedings — no single tool verdict should be trusted. Use at least two independent tools and require agreement before acting on the result.
A practical workflow:
- Run the image through Aiscern for a detailed signal breakdown
- Independently verify with Hive or FotoForensics
- Manually examine the specific areas flagged by each tool
- Check EXIF metadata independently
Limitations of All Current Tools
Every tool in this comparison has significant gaps:
Heavily compressed images lose the frequency artifacts that most detectors rely on. An AI image shared on WhatsApp or downloaded from Twitter at web-quality resolution is meaningfully harder to detect.
Cropped or filtered images can remove or obscure the artifacts in hair edges, backgrounds, and other tell-tale areas.
Hybrid images — AI inpainting applied to real photographs — are currently the hardest to detect reliably. None of the tools above reliably distinguishes these.
Model-specific bias: Each tool performs better on the AI models that were most prevalent in its training data. A tool trained before Midjourney v7 may underperform on v7 outputs.
The Bottom Line
For most use cases, Aiscern's combination of accuracy, signal transparency, and free access makes it the strongest option. The signal breakdown — showing exactly which forensic indicators contributed to the verdict — provides the context needed to make informed judgments rather than blindly trusting a score.
No tool should be the final word on whether an image is real. Treat all detectors as one input in a broader assessment process.