How to Spot AI-Generated Images: Visual Clues & Tools

The visual tells that made AI-generated images obvious in 2022 — six fingers, uncanny skin, garbled background text, glassy eyes — have been substantially addressed in current models. Midjourney v6, DALL-E 3, and Flux 1.1 now produce hands, faces, and text that pass casual human inspection most of the time. What has not changed is the underlying physics of how these generators work: they predict pixels from learned statistical distributions, not capture light from the physical world. That distinction still leaves detectable artifacts — but finding them requires knowing where to look.

Why Human Detection Fails More Than We Think

People consistently overestimate their ability to identify AI-generated images. The research evidence is unambiguous: a 2024 systematic meta-analysis published in ScienceDirect, covering 56 peer-reviewed studies on deepfake detection, found that average human accuracy across all modalities is 55.54% — marginally above the 50% baseline of random guessing. For high-quality deepfake video specifically, human accuracy drops to 24.5% correct identification.

Stanford HAI research has documented the mechanism: human judgment relies heavily on holistic aesthetic impression, which AI generators are specifically optimized to produce. Current models are trained on hundreds of millions of human-rated image pairs — they have learned what humans find visually coherent and appealing. That same optimization makes the outputs resistant to the kind of gestalt "something feels off" judgment that people rely on.

The implication is that visual inspection alone should never be the sole basis for an authenticity determination. It is a supplementary layer in a multi-signal workflow — which is why the remaining sections of this guide work best when used together.

Visual Clues That Still Work in 2026

These are ordered by their current reliability — how often they produce a signal in genuinely AI-generated images as of 2026, given improvements in current models. None is individually conclusive.

1. Eye Catchlight Geometry (High Reliability)

Look at the small reflective highlights in a subject's eyes. In a real photograph taken in consistent lighting, the catchlights in both eyes should be mirror images of each other — reflecting the same light source from the same angle. AI generators produce eye catchlights that look plausible in isolation but often fail this symmetry test: the shape or position of catchlights differs between the left and right eye, or the catchlight pattern does not match the visible light source in the background.

To use this check: zoom to at least 200% on the eye area and compare catchlights side by side. This requires a high-resolution source image — a compressed social media thumbnail will not show enough detail. A 2026 guide by Geeky Gadgets specifically named this check as one of the six most reliable 2026 detection signals.

2. Hand and Finger Anatomy (Moderate Reliability, Improving)

Hands remain the most famous AI tell and are still imperfect in current models — but significantly improved since 2022. Midjourney v5 and earlier frequently produced six-fingered hands, fused knuckles, and anatomically impossible finger proportions. Midjourney v6 and DALL-E 3 handle the basic finger count much better, but show weaknesses in contextual hand use: gripping something without the grip geometry being correct, knuckle topology that shifts between frames, or fingernails that appear on the wrong anatomical surface.

Check hands only in high-resolution images where the hand is prominent in the frame. Background hands at low resolution are not informative — models degrade all small details similarly, and you cannot distinguish AI degradation from compression.

3. Embedded Text Rendering (High Reliability)

AI image generators produce text by predicting pixel patterns, not by rendering actual characters. Despite improvements, text within AI-generated images remains a reliable detection signal: signs, books, name badges, and storefront lettering often contain letters that look almost-but-not-quite right — plausible letter forms that do not combine into real words, or words that rearrange randomly across near-identical generation attempts.

This check is only useful when text appears in the image prominently. Absence of visible text tells you nothing; the presence of garbled or nonsensical text is a strong positive signal. An image claiming to show a protest with legible signs — where the sign text is unreadable gibberish at full resolution — is suspect.

4. Shadow and Lighting Consistency (Moderate Reliability)

Physical light sources cast shadows in consistent directions. AI generators optimize for each region of the image semi-independently, which produces shadow inconsistencies: a shadow that suggests one light source on the subject's face, while background object shadows suggest a different angle. Similarly, specular highlights on skin, glasses, and metallic surfaces may not reflect a consistent environment.

This check is more reliable for indoor scenes with distinct light sources than outdoor scenes under diffuse natural light, where the constraint is less obvious. It is most useful for images where multiple objects with different surface properties appear near each other.

5. Skin Texture at High Zoom (Moderate Reliability)

AI generators render skin with a characteristic over-smoothness at certain zoom levels — a porcelain-like quality that lacks the fine hairs, pores, capillary detail, and minor blemishes of real skin photographed under high magnification. This tell is most apparent on faces photographed at close range in high-resolution images. At lower resolutions or with heavy post-processing, it becomes unreliable.

Current models have improved significantly on this. Midjourney v6 in particular produces convincing skin texture in many conditions. Use this check in combination with others, not as a standalone signal.

6. Background Edge Coherence (Lower Reliability, Still Useful)

Zoom to the edges where a foreground subject meets the background. Real photographs show consistent focus fall-off governed by depth of field physics. AI images often show abrupt transitions from sharp foreground detail to blurred background that do not follow the geometry of real depth of field — the blur radius is wrong relative to the focal length, or the bokeh pattern is characteristic of AI softening rather than real lens behavior. Background objects may dissolve into abstract texture at the edges rather than remaining geometrically coherent.

Metadata and Technical Signals

EXIF Data Inspection

EXIF metadata is embedded by cameras at the point of capture: camera make and model, lens identifier, focal length, aperture (f-stop), shutter speed, ISO setting, timestamp, and optionally GPS coordinates. AI-generated images have no camera — so they typically either contain no EXIF data at all, or contain metadata that was manually added after the fact (which may list impossible hardware configurations).

On Windows: right-click any JPEG, select "Properties" → "Details" tab. On Mac: open in Preview, "Tools" → "Show Inspector" → "Exif". Online: tools like Jeffrey's Exif Viewer or ExifTool (command-line) show complete metadata.

The critical caveat: do not treat absent EXIF data as evidence of AI generation. Instagram, X, Facebook, and most other social media platforms automatically strip EXIF data from images on upload. A genuine photograph shared on social media will typically have no EXIF metadata. The signal here is positive identification (complete, physically plausible EXIF from a real camera model), not negative inference.

Reverse Image Search

Google Reverse Image Search, TinEye, and Bing Visual Search index billions of images and can identify whether the same image — or visually similar images — appear elsewhere on the web with different claimed origins. This check does not directly detect AI generation, but it catches misrepresentation: an image presented as an original photograph that actually appears on AI art galleries, stock generator sites, or multiple unrelated contexts with different claimed dates and authors.

Automated Tool Comparison: What They Catch and Miss

Compiled from: Springer Nature Journal 2025 (arXiv:2025.154), NTIRE 2026 Challenge benchmarks, GIJN Reporter's Guide 2026, ScienceDirect meta-analysis (2024)

The Complete Verification Workflow

Professional fact-checkers at AFP, Reuters, and the European journalism organization GIJN use structured workflows rather than relying on any single signal. The GIJN Reporter's Guide to Detecting AI-Generated Content (2026 edition) recommends the following sequence:

1. Source the highest-resolution original. If you received the image via social media, find the originating post. Download, do not screenshot. Screenshots add a JPEG compression layer that degrades frequency-domain signals substantially.
2. Check EXIF before anything else. Takes 30 seconds. Open file properties. Complete, physically plausible camera metadata is a strong authenticity signal. Move forward based on the metadata result before running any detector.
3. Run one automated detector. Use EyeSift's image analysis tool or Hive Moderation if you have API access. Note the confidence score — not just the binary verdict.
4. Run a reverse image search. Google Images, TinEye. Note every context in which the image appears online, including dates and claimed origins.
5. Apply visual inspection to three specific areas: eyes (catchlights), any visible text, and hands if present. Zoom to at least 200% on each. Document what you observe.
6. If the first detector returns 40–80% confidence, run a second tool with a different underlying method. Convergence at that level increases confidence; divergence means the result is genuinely ambiguous.
7. Make a final judgment call. Treat everything below 75% automated confidence as requiring additional investigation or human expert review. For consequential decisions — news publication, legal proceedings, academic fraud cases — document your full verification process.

What Has Improved vs. What Has Not Changed

It is worth being explicit about which tells have been meaningfully patched by current models and which remain exploitable — because checklist-based guides age quickly in this domain.

Largely improved in Midjourney v6, DALL-E 3, Flux 1.1, and equivalent 2025–2026 models: basic finger counts, foreground face coherence, obvious skin texture over-smoothing, large foreground text (when prompted explicitly), and ear anatomy.

Still reliably detectable in 2026: subtle catchlight geometry mismatches in eyes, background crowd anatomy at medium zoom, complex hand articulations (gripping, pointing), background text and signage, shadow direction inconsistencies in complex indoor scenes, and frequency-domain statistical signatures (only detectable with tools, not naked eye).

The practical takeaway: the 2022 checklist of "look for six fingers and strange hands" is outdated as a standalone method. The 2026 approach requires more targeted zooming, context-specific checks, and mandatory tool augmentation. Visual inspection alone has never been reliable; in 2026 it is even less so without a systematic protocol.

For editors and educators who need to understand the detection technology itself at a deeper level, our technical guide on how AI detectors work covers perplexity scoring, neural classifier architecture, and watermarking methodology in detail.

Frequently Asked Questions