AI generated vs real photos: how to spot the difference

Most people think they can tell the difference between AI-generated photos and real ones. Most people are wrong. In controlled tests, human accuracy at distinguishing AI-generated images from real photographs hovers around 50-60% — barely better than a coin flip. And that number is dropping as generative models improve.

This guide covers exactly what separates the two — the forensic tells that exist today, why they're disappearing, and what that means for organizations that need reliable answers at scale.

What a real camera actually captures

A photograph is a record of physics. Light bounces off a surface, passes through a lens, hits a sensor, and gets converted into data. That process introduces a specific set of artifacts that are deeply hard to fake:

Sensor noise. Every camera sensor generates random electronic noise, especially in low-light conditions. This noise follows predictable statistical patterns tied to the specific sensor model — a fingerprint of sorts. Real photos contain it. AI images often lack it entirely, or simulate it so uniformly that it looks artificial.

Lens characteristics. Real lenses produce chromatic aberration (color fringing at high-contrast edges), vignetting (darkening at the corners), and distortion specific to their focal length. These aren't bugs — they're signatures.

Depth of field physics. Bokeh — the blur in out-of-focus areas — follows optical laws. The shape of the blur depends on the aperture blades, the lens formula, and the distance to the subject. AI models approximate this, but the approximation breaks down under close inspection.

EXIF metadata. Real photos carry embedded metadata: camera make and model, GPS coordinates, timestamp, shutter speed, ISO, focal length. A photo stripped of metadata is suspicious. A photo with metadata that doesn't match the scene is a red flag.

The most common AI image tells

Even the best current models — Midjourney v6, DALL-E 3, Stable Diffusion XL — leave traces:

Hands and fingers. Human hands are topologically complex. AI models frequently generate hands with six fingers, merged knuckles, fingers that blend into each other, or wrists that don't connect properly. When in doubt, count the fingers.

Text and signage. Ask an AI to generate an image with readable text and it will usually produce something that dissolves into nonsense on close inspection. Letters merge, words are misspelled, and fonts are inconsistently applied.

Symmetry artifacts. Human faces and natural objects are slightly asymmetric. AI models have a bias toward symmetry that produces faces that look almost too balanced.

Skin texture at scale. Zoom into a real portrait and you'll see pores, fine hairs, micro-shadows in skin folds, and subtle color variation. AI skin tends to be smooth in a way that's not quite right.

Background coherence. Real scenes have consistent physics. Shadows fall from a single light source. Reflections are accurate. AI backgrounds often have elements that don't add up: a shadow pointing the wrong direction, a reflection that shows something that isn't there.

Teeth. AI-generated smiles often show too many teeth, teeth that blend into each other, or gum lines that don't follow the anatomy of the jaw.

For a deeper technical breakdown, see our guide on how to tell if a photo is AI generated.

Why the gap is closing — fast

Every tell listed above was more obvious in 2023 than it is today. Models are trained on human feedback specifically targeting their failure modes. Midjourney v6 hands are dramatically better than v4. DALL-E 3 text rendering is legible in a way previous versions weren't.

A 2024 study from the MIT Media Lab found that error rates for human detection of AI faces reached near-parity with real photographs. The Content Authenticity Initiative, a coalition backed by Adobe, BBC, and others, has been working on provenance standards precisely because visual inspection alone is no longer sufficient.

The implication: the forensic tells that work today are a moving target. Checklists of visual artifacts have a shelf life measured in model generations, not years.

Where human inspection fails

Even a trained forensic analyst working with optimal tools will miss some AI-generated images — and will flag some real ones incorrectly:

Volume. A claims adjuster reviewing 500 images a week cannot apply the same scrutiny to each one. Attention degrades.

Adversarial generation. Some AI images are created specifically to pass human inspection. Post-processing pipelines add realistic noise, strip EXIF data, and introduce deliberate imperfections.

Context bias. If a photo arrives with a plausible story — a customer submitting damage photos, a seller listing a property — humans apply motivated reasoning. We want the image to be real.

Compression artifacts. Images shared via messaging apps or social media are re-compressed, which can obscure or mimic AI artifacts in both directions.

This is where the gap between human judgment and algorithmic detection becomes consequential. For insurance use cases, a single missed AI-generated claim can cost tens of thousands of dollars. For e-commerce platforms, fraudulent listings with fabricated product images erode buyer trust.

What automated detection looks at

Modern AI detection systems operate on multiple signal layers simultaneously:

Pixel-level analysis. Convolutional networks trained on millions of real and synthetic images learn to detect statistical patterns in pixel distributions invisible to the human eye.

Semantic consistency checks. Does the lighting direction match across the whole image? Do shadows imply a consistent sun position? Are reflective surfaces physically accurate?

Metadata forensics. Beyond just checking whether EXIF data exists, forensic tools verify internal consistency — does the GPS location match the time zone implied by the timestamp?

Provenance chain analysis. C2PA (Coalition for Content Provenance and Authenticity) standards allow cameras and software to cryptographically sign images at creation. A verified provenance chain is currently the strongest signal of authenticity available.

The Reality AI platform applies all of these layers in combination, returning a confidence score and a breakdown of which signals contributed to the assessment. The platform overview covers how this integrates into existing review workflows via API.

A practical checklist: what to check

When you're manually inspecting an image:

Start with the obvious:

- Count fingers on any visible hands

- Read any text in the image

- Check for unnatural symmetry in faces

Move to physics:

- Identify the implied light source and verify shadows are consistent

- Check reflections in glasses, windows, and wet surfaces

- Look at the background for figures or objects that don't make sense

Zoom in:

- Examine skin texture at 200%+ zoom

- Look at edges where in-focus and out-of-focus regions meet

- Check hair strands at the boundary against the background

Check the file:

- Open EXIF data — note what's there and what's missing

- Compare stated camera settings to the scene

- Look for signs of re-compression artifacts inconsistent with stated origin

This process takes five to ten minutes per image under good conditions. It doesn't scale, and it doesn't catch adversarial generation.

The bottom line

The gap between AI-generated images and real photographs is measurable — for now. The tells exist, they're documentable, and they're useful for building intuition. But they're not a reliable enterprise-grade detection strategy.

The organizations winning on image authenticity aren't relying on analysts to spot bad hands. They're running automated detection at ingestion, flagging anomalies for human review, and building provenance requirements into vendor agreements.

If your business depends on visual evidence being real — claims, listings, compliance, legal — the question isn't whether you can spot the difference. It's whether your system can.

[Book a demo with Reality AI](/book-a-demo) to see how enterprise-grade image authentication works in your workflow.