AI Image Detector: Spot AI-Generated Photos Online
AI image detector online: analyze images to spot AI-generated photos with compression artifact heatmaps and confidence scores.
Updated 2026-08-26
Related Tools
AI Content Detector: Check ChatGPT, Gemini & Copilot Text
Image Compressor: Reduce JPG, PNG, WebP File Size Online
Image to Base64: Convert JPG, PNG & WebP to Base64 Online
AI Answer Compressor: Trim Filler Words & Save Tokens
AI Code Extractor: Pull Code, URLs & Tables from Chat
AI Humanizer: Make AI Text Sound Human & More Natural
Features
- Error Level Analysis (ELA): re-encodes your image as JPEG and measures the per-pixel difference
- Grid heatmap visualization: average error per cell colored green → yellow → red
- Overall 0-100 score with a three-way verdict: likely AI, likely photographic, or inconclusive
- Standardized JPEG re-encoding so PNG and WebP images can be analyzed too
- Auto grid sizing: the grid adapts to your image, or pick 6×6, 8×8, or 12×12 manually
- Side-by-side view of the original and the ELA heatmap
- PNG and other lossless formats are converted to JPEG first for a meaningful comparison
- Honest about limits: results are a heuristic screening signal, not proof
- Drag and drop, file picker, or clipboard paste input
How to Use
- 1Open an image by dragging it in, clicking the drop area, or pasting it with Ctrl+V.
- 2Wait for the ELA analysis to finish; it usually takes a second or two.
- 3Read the overall score and the verdict badge: likely AI, likely photographic, or inconclusive.
- 4Compare the original with the ELA heatmap: red cells are where re-compression changed pixels the most.
- 5Switch the grid size to inspect the error distribution at different scales.
- 6Check a suspicious image you found online for an unusually smooth or repetitive error pattern.
- 7Compare the heatmap of a real photo with a generated one to learn what the difference looks like.
- 8Verify profile pictures and avatars for signs of AI generation before trusting them.
- 9Treat the verdict as a screening signal; combine it with other checks before drawing conclusions.
- 10Use large, unedited source images for the most reliable results.
Frequently Asked Questions
What is Error Level Analysis (ELA)?
ELA measures how much an image changes when it is re-encoded as a JPEG. The image is compressed at a known quality and compared pixel by pixel with itself. Real photographs contain natural sensor noise, so re-compression produces a scattered, organic error pattern. AI-generated images often lack that noise, producing unusually smooth or unusually repetitive error maps.
How accurate is this detector?
ELA is a heuristic, not a deterministic test. It works best on large, unedited images with a clear compression history. Screenshots, platform re-compression (social media, messaging apps), and small images destroy the signal, and high-quality diffusion models can produce images that look 'natural' to ELA. Treat the verdict as a screening signal, never as proof.
Which images can't be analyzed reliably?
Screenshots and screen captures, images re-compressed by social platforms or messaging apps, images smaller than roughly 256×256 pixels, heavily filtered or edited images, and text-heavy images. If the heatmap comes out uniformly flat, the analysis is not informative.
What's the difference between this tool and the AI Content Detector?
The AI Content Detector analyzes text with statistical metrics (perplexity, burstiness, vocabulary) to screen AI-written copy. This tool analyzes images with Error Level Analysis. They share the same honest philosophy: heuristic screening with clearly stated limits, never a final verdict.
How do I read the heatmap?
Each cell shows the average re-compression error of that area, colored green (low) → yellow → red (high). Red cells mark where the image changed most under JPEG re-encoding. AI-generated images often show an abnormally uniform map (almost all green or with repeated texture patterns), while photos show a scattered, noise-like distribution.
How are PNG images handled?
PNG is lossless, so direct ELA on a PNG means nothing; there are no compression artifacts to compare. PNG and other lossless formats are first converted to JPEG internally, then re-compressed again for the comparison. The same standardized pipeline is applied to every image so results stay comparable.
Can I analyze multiple images at once?
No. This tool analyzes one image at a time. The analysis is interactive and pixel-level, and batching would make results harder to interpret. For a large batch, use this tool as a spot-check on the most suspicious images.
What image formats are supported?
PNG, JPG/JPEG, WebP, GIF, and BMP: anything the browser can decode, including HEIC on supported devices. Non-JPEG formats are normalized through the JPEG pipeline automatically.
How does this compare to Hive or other commercial detectors?
Services like Hive run trained deep-learning classifiers over huge image datasets, while this tool applies Error Level Analysis heuristics locally. For a disputed image, run both and compare verdicts; where they disagree, the image is genuinely ambiguous. This tool is a free screening signal, not a certified forensic test.
Can I verify images downloaded from Google or social media?
You can, with a caveat: platforms re-compress images on upload, which destroys the ELA signal and pushes results toward 'inconclusive' or false positives. Always try to obtain the original file (source page, author's post) before testing; a verdict on a re-compressed copy is not reliable evidence.
Why did my screenshot score as likely AI?
Screenshots are flat, re-rendered pixel blocks. After JPEG re-compression they produce an extremely smooth, uniform error map that matches the 'too smooth' pattern we flag as AI-like. That is a known false-positive class; use the tool on the original file whenever possible.
What does the score mean exactly?
The score runs 0-100; higher values indicate stronger AI-like ELA patterns (abnormally smooth, uniform, or repetitive error distribution). Below 30 the image looks photographic, above 70 it shows strong AI-like traits, and in between the result is inconclusive. The score combines the mean error, its spatial uniformity, and the overall error level.
Can AI image detectors be bypassed by editing the image?
Not in a meaningful way. Re-saving, compressing, filtering, or cropping an image destroys the compression-history signal that ELA measures, which pushes the result toward 'inconclusive', not toward 'photographic'. That is why the tool asks for the original file and why heavily processed images are simply untestable. No editing trick reliably makes an AI image score as a real photo, and no detector can honestly guarantee otherwise.