- AI image detectors give you a probability score, not definitive proof—combine them with reverse image search and visual checks for better accuracy.
- Look for unnatural hands, garbled text, impossible reflections, and odd jewelry details, but know that newer generators like Midjourney and ChatGPT Images 2.5 fix many of these tells.
- Use Google Lens to trace an image back to its original source and posting date—the context often matters more than the pixels.
- Content Credentials (C2PA) and Google’s SynthID watermark are emerging standards, but only work when the creator or platform embeds them.
An AI image detector analyzes a photo and returns a percentage estimate of whether it was made by a generator like Midjourney, DALL-E, or Stable Diffusion. The number is a guess based on patterns the detector has seen before, not a forensic verdict. A detector might flag a real photo as AI or miss an AI image entirely, especially one made with a newer model it has never studied.
That is why no single method is enough. To judge whether a viral photo is AI-generated, you combine detector results with reverse image searches, visual spot-checks, and source context. This guide walks through all four approaches so you can make an informed call without a computer-science degree.
What AI image detectors do and why they are not definitive
An AI image detector is a piece of software trained on thousands of real and AI-generated images. It learns to recognize statistical fingerprints—compression artifacts, color gradients, texture patterns—that generators tend to leave behind. When you upload a photo, the detector compares it to those learned patterns and outputs a score: “85% likely AI-generated” or “20% confidence.”
The problem is twofold. First, each detector is only as good as its training set. If it has never seen images from the latest AI image generators, it will miss their tells. Second, real photos can trip the same statistical patterns—heavy editing, certain camera sensors, and JPEG compression all introduce artifacts that look “AI-like” to the algorithm.
That means a high score is a red flag worth investigating, not a verdict. And a low score does not prove the image is real. Always treat detector results as one data point in a larger picture.
Visual clues you can check yourself
Before AI generators became mainstream, spotting a fake was straightforward: look for mangled hands, nonsense text, and physics-defying reflections. Those tells are still useful, but they come with a shrinking window of reliability.
Hands and fingers
Early diffusion models struggled with hands because fingers overlap, bend, and foreshorten in ways that confuse pixel-level training. You would see six fingers, fused knuckles, or thumbs sprouting from wrists. By mid-2025, models like Midjourney v7 and ChatGPT Images 2.5 learned to render hands correctly most of the time. A perfect hand no longer proves the image is real, and a broken hand no longer proves it is fake—some real photos catch fingers mid-motion or partially obscured.
Still worth checking: zoom in on any visible hands. Count fingers. Trace each finger to its base. If the geometry is off or a finger seems to vanish into a sleeve with no clear endpoint, that is a point against authenticity.
Text and signage
Generators used to produce garbled, alphabet-soup text on storefronts, book covers, and T-shirts because they treat letters as visual textures rather than readable language. Recent models have improved, but text remains a weak spot. Look for:
- Nonsense words that almost look real (“COFFAE” instead of “COFFEE”)
- Letters that change style mid-word
- Curved or wavy baselines where text should sit straight
- Repeated glyphs in a pattern that makes no semantic sense
If the text is tiny, blurred, or angled away, even a real photo might look garbled. Context matters.
Reflections and shadows
Physics is hard for diffusion models. A person standing in front of a mirror should appear in the reflection with the same outfit, pose, and lighting. AI generators sometimes render a different pose, swap clothing colors, or omit the reflection entirely. Shadows should fall in a direction consistent with the light source; AI-generated shadows sometimes point in multiple directions or float detached from their objects.
Check any glass surface, puddle, or shiny floor. Does the reflection make physical sense? Is the lighting consistent across the scene?
Jewelry and fine details
Necklaces, earrings, and chains are another trouble spot. Generators struggle with thin, overlapping geometry. You might see a necklace clasp that connects to nothing, earrings that pass through the earlobe instead of hanging from it, or chain links that fuse into a solid mass. Teeth, eyeglasses, and hair strands show similar issues—look for transitions that defy anatomy or optics.
Backgrounds and edges
AI-generated images sometimes show “melting” at the boundary between subject and background. A shoulder might blend into wallpaper, or hair strands might merge with the sky. This happens because the model’s attention mechanism blurs the distinction between foreground and context. It is subtle, but zooming in on edges can reveal it.
Backgrounds themselves can be too perfect—no stray cables, no clutter, everything symmetrically arranged in a way real rooms rarely are. Or they can be incoherent: a bookshelf where every spine is nonsense, a cityscape where buildings have impossible perspectives.
Why these clues are fading
Every tell listed above was a reliable red flag in 2023 and early 2024. By late 2025 and into 2026, the top-tier generators have closed most of these gaps. Midjourney v7 renders anatomically correct hands. ChatGPT Images 2.5 can produce legible signage. Flux and Stable Diffusion 4.0 handle reflections and fine jewelry far better than their predecessors.
That does not make visual checks useless—it makes them probabilistic. A broken hand is still evidence. A perfect hand is no longer proof of authenticity. You need more data points.
Content Credentials and SynthID: the emerging standards
Two technical standards aim to label AI-generated images at the point of creation, so you do not have to guess later.
Content Credentials (C2PA)
Content Credentials is a specification maintained by the Coalition for Content Provenance and Authenticity (C2PA). When a camera, editing app, or AI generator supports C2PA, it embeds a cryptographically signed manifest into the image file. That manifest records the tool that created it, the date, and whether any edits were made. Adobe, Microsoft, and Leica have adopted it; some Canon and Nikon cameras ship with C2PA support.
To check an image for Content Credentials, upload it to a compatible viewer or use a browser extension. If the manifest is present and intact, you will see a provenance chain: “Created by Adobe Firefly on 2026-10-12, no edits.” If the image was edited after creation, the chain shows each step.
The catch: Content Credentials only work if the creator enabled them. Most AI generators do not embed C2PA by default. Social media platforms often strip metadata when you upload a photo, breaking the chain. And a bad actor can simply screenshot an AI image and re-save it, shedding the manifest. Content Credentials are a strong positive signal when present, but their absence proves nothing.
Google SynthID
SynthID is Google’s watermarking system for AI-generated images and text. Instead of adding a visible mark, it subtly alters pixel values in a way that is invisible to the human eye but detectable by SynthID’s decoder. Google embeds SynthID in images created through its Gemini image generator and has made the technology available to other platforms.
SynthID survives moderate editing, cropping, and compression—techniques that erase traditional watermarks. But it is not foolproof. Heavy filtering, converting the image to a different format multiple times, or printing and re-scanning can degrade the watermark below detection thresholds.
Like Content Credentials, SynthID is opt-in. Only images created by participating generators carry it, and there is no public SynthID decoder you can use at home. It is a tool for platforms and investigators, not a first line of defense for everyday users.
Reverse image search: finding the source
One of the most reliable ways to judge an image is to find out where it first appeared and in what context. Google Lens makes this simple.
How to use Google Lens
On a smartphone, open the Google app or download Google Lens. Tap the camera icon, then either take a photo of the image on another screen or upload it from your gallery. Lens will show you visually similar images from across the web, along with pages where the exact image appears.
On a desktop, go to lens.google.com. Click the upload icon and select the image file. Lens returns a list of matches with dates and sources.
What to look for
If the image has been online for months or years, it is more likely to be a real photo making the rounds again. If it only appears on Reddit threads from the past 48 hours with captions like “Is this real?”, that is a red flag.
Check whether the earliest source is a credible news outlet, a verified social account, or a stock-photo library. If the trail leads back to an AI art showcase or a “generated images” tag on a forum, you have your answer.
Reverse search also catches recycled fakes—an AI image from 2024 passed off as a breaking news photo in 2026. The timestamp and context matter as much as the pixels.
Using the image-to-prompt tool
If you suspect an image is AI-generated, Convly’s free image-to-prompt tool can help you understand what prompt might have created it. Upload the image, and the tool reverse-engineers a text description—”a photorealistic portrait of a woman in a red dress, studio lighting, shallow depth of field.”
This does not prove the image is AI-generated, but it gives you the vocabulary to search for similar known-AI images. Paste the reconstructed prompt into an image search or an AI art database, and see if you find matches. If you do, that is strong circumstantial evidence.
Why you should not rely on detector websites alone
Dozens of websites offer free AI image detection: upload a photo, wait a few seconds, get a percentage. They are built on the same principle—trained classifiers looking for statistical tells—but their accuracy varies wildly, and none publish their false-positive or false-negative rates.
A 2025 independent study tested eight popular detectors on a mix of real photos and AI images from Midjourney, DALL-E, and Stable Diffusion. The best detector achieved 78% accuracy on its training-era models but dropped to 54% (barely better than a coin flip) on images from generators released after its training cutoff. The worst detector flagged 40% of real photos as AI.
Detectors also fail in predictable ways:
- Heavily edited real photos score high because the editing software introduces artifacts the detector associates with generation.
- Simple AI images—flat illustrations, minimal backgrounds—score low because they lack the complex textures the detector learned to recognize.
- Adversarial tweaks: adding a thin layer of noise or running an image through a specific filter can fool the detector without changing how the image looks to you.
Convly does not offer an AI image detector. If you choose to use one, treat the result as a starting point. A high score means “investigate further,” not “case closed.”
Deepfakes and video: a related problem
AI-generated images are one piece of a larger puzzle. Deepfake videos—face-swapped clips, AI-generated voices, synthetic news anchors—raise the same questions of authenticity but move at 30 frames per second. The techniques for checking them overlap with still-image detection (visual artifacts, source checks, metadata), but video adds motion analysis and audio forensics to the mix.
For a deeper look at how deepfakes work, why they are hard to detect, and what tools exist to spot them, read Convly’s guide on deepfakes: the threat and how to detect them.
Putting it all together: a checklist
When you encounter an image you want to verify, work through these steps in order:
- Run a reverse image search with Google Lens. Find the earliest source and check the context. If it traces to a credible outlet or predates the current AI generation era, it is likely real.
- Look for Content Credentials or SynthID. If present, they are strong evidence (though absence proves nothing).
- Inspect visual details: hands, text, reflections, jewelry, background coherence. One flaw is not proof, but a cluster of issues raises the probability.
- Upload to an AI image detector if you still have doubt. A very high score (above 90%) is worth noting; a middling score (40–70%) is noise.
- Check the image-to-prompt tool to see if the reconstructed prompt matches known AI art styles or tropes.
- Weigh the context: who shared it, when, and with what claim? A breathless “this just happened” attached to an image with zero prior web presence is a red flag all by itself.
No single test is bulletproof. Together, they give you enough signal to make a reasoned judgment.
Frequently asked questions
Can I trust a 95% AI detection score?
A very high score means the detector found patterns it strongly associates with AI generation, so it is worth investigating further. But detectors have false positives—heavily edited real photos, certain camera sensors, and compressed images can all score high. Use the score as one input, not the final word.
Why do some AI images look completely real?
The top AI image generators in 2026—Midjourney v7, ChatGPT Images 2.5, and Flux—have been trained on billions of high-resolution photos and have learned to mimic lighting, texture, and anatomy at a level that fools the human eye in many cases. They still make mistakes, but those mistakes are rarer and subtler than they were two years ago. As models improve, visual spot-checks become less reliable, which is why reverse search and metadata checks matter more.
What is the most reliable way to check if an image is real?
Reverse image search with Google Lens is the single most useful tool because it tells you where the image came from, when it first appeared, and in what context. An image that has been online for years or that traces to a credible news outlet is far more likely to be real than one with no prior web presence. Combine that with visual checks and detector results for the strongest verdict.
Do social media platforms label AI images?
Some do, inconsistently. Meta (Facebook and Instagram) began adding “Imagined with AI” labels in 2024 to images with C2PA Content Credentials, but only if the metadata survived upload. X (formerly Twitter) does not label AI images. TikTok requires creators to disclose AI-generated content but relies on self-reporting. No platform has a perfect system, so treat labels as helpful but not exhaustive.
Can I remove an AI watermark to pass off an image as real?
Stripping metadata or cropping a visible watermark is technically possible, but doing so to deceive others—especially in news, advertising, or academic contexts—is unethical and often illegal. Many jurisdictions now have laws against deepfake fraud and misleading synthetic media. Beyond the legal risk, it erodes trust in visual information for everyone. If you generated an image, disclose it.
Will AI image detection get better?
Detection and generation are in an arms race. Each time detectors learn a new tell, generators are trained to avoid it. The long-term solution is not better detectors but better provenance systems—Content Credentials, SynthID, and blockchain-based media registries that tag authenticity at creation. Until those systems are universal, detection will remain probabilistic and imperfect.