Table of Contents
Here’s an uncomfortable finding to start with. MIT researchers ran five preregistered experiments with over 2,200 participants, testing whether people could identify state-of-the-art political deepfakes. The result: humans performed no better than chance. Not “slightly worse than experts.” Coin-flip territory. And audio fakes fooled people even more reliably than video.
So let me be honest about what this guide is and isn’t. Most “how to spot AI images” articles are recycling tells from 2022-era generators, hands with seven fingers, garbled text, and teaching you those today is teaching you false confidence. The current answer has three layers: a shrinking set of visual tells that still work sometimes, detector tools whose accuracy problem nobody mentions, and the method that actually holds up, which isn’t about pixels at all. Plus a 60-second routine that puts it all together before you share anything.
The Old Tells That Don’t Work Anymore
Retire these from your mental checklist, because modern generators have largely fixed them: mangled hands and extra fingers, garbled text on signs, dead plastic eyes, obviously warped backgrounds, and for video, the classic advice about unnatural blinking and blurry face edges. Those were artifacts of early tools, and the polished output of 2026’s generators simply doesn’t reliably produce them. Worse, believing in dead tells cuts both ways: it makes you certain a sophisticated fake is real because “it doesn’t have the AI look,” and lets bad actors dismiss authentic photos as fake because they’re grainy or oddly lit. Both failure modes are exactly what misinformation feeds on.
What Still Gives AI Images Away (Sometimes)
A few flaws persist in 2026, and they reward zooming to 200 to 400% rather than glancing:
Jewelry, accessories, and mechanical detail. Still among the most frequent slip-ups: watch faces with impossible hands, earrings that don’t match each other, rings that melt into skin, glasses whose frames don’t quite continue behind the ear. Generators understand faces deeply and objects-attached-to-faces less so.
Skin and texture at close zoom. Either poreless and airbrushed or oddly plastic, with a uniformity real skin doesn’t have.
Backgrounds painted with one brush. A uniform, smeared blur that treats everything behind the subject as a single texture, unlike real lens blur, which varies with distance.
Physics and consistency. Shadows disagreeing about where the light is, reflections that don’t match the scene, patterns (fabric, tiles, crowds) that repeat a little too rhythmically.
Treat every one of these as a lead, not a verdict. Their absence proves nothing, and that asymmetry is why the next two sections matter more.
AI Detector Tools: The Accuracy Problem Nobody Mentions
Paste an image into an “AI detector,” get a percentage, done? I wish. Here’s what the numbers actually look like: tools advertise 96 to 99% accuracy, measured on their own benchmark datasets. Independent testing against current commercial generators tells a different story, with accuracy collapsing to as low as 18 to 30% on images from the latest models. Even state-of-the-art academic detectors that score 91 to 92% on standard benchmarks drop to roughly 60% on fakes from methods they weren’t trained on, and new generators are increasingly tuned specifically to defeat existing detectors.
The practical translation: a detector score is a first signal, never a verdict. A “92% real” badge on a fake, or “87% AI” on a genuine photo of a real event, are both routine outcomes, and people have been wrongly accused and wrongly reassured by both. Use detectors the way a doctor uses one ambiguous symptom: noted, and never diagnosed alone.
The Provenance Method: How Verification Actually Works Now
Here’s the shift that matters. The question “does this look real?” is a fight your eyes will increasingly lose. The question that still has good answers is “where did this come from, and can that be verified?” Three tools do the heavy lifting, and they’ve matured dramatically this year.
Reverse image search, still undefeated. Google Lens or TinEye, ten seconds, free. It catches the most common deception of all, which isn’t generated imagery but real images recycled out of context: an old photo relabeled as breaking news, a scene from one country captioned as another. It also surfaces whether reputable sources have already debunked the image. This is the single highest-value habit in this entire article.
Content Credentials (C2PA). An industry standard where cameras and creation tools cryptographically sign media with a tamper-evident record of how it was made and edited. Adobe, Google, OpenAI, Microsoft, and camera makers like Sony and Leica are on board, and you can inspect any file at contentcredentials.org. Unlike a detector’s guess, this is cryptographic: when credentials are present, they tell you the creation tool and whether AI was involved, full stop.
SynthID and Chrome’s new check. Google’s invisible watermark is baked into the pixels of content from participating generators, surviving cropping, compression, and screenshots, with over 100 billion pieces of content marked so far. The practical upgrade landed this year: Chrome can now check an image for it from a right-click, and detection is rolling into Search. OpenAI, Nvidia, and ElevenLabs have joined the standard.
One caveat governs all three, and it’s the honest heart of provenance: absence proves nothing. Open-source generators carry no watermark, credentials can be stripped, and most real photos ever taken have neither. Presence settles questions; absence just means inconclusive. Also worth knowing: from next month, the EU’s AI Act begins requiring AI-generated content to be labeled, which will make provenance signals steadily more common, and their absence on slick viral content steadily more suspicious.
Video and Voice: The Harder Problem
Video deepfakes add motion tells worth knowing: lip movements slightly desynced from audio, faces that stay eerily stable while heads turn, and above all the cut-length tell, since manipulated clips are so often short, cropped, and context-free precisely because surrounding footage would expose them. For anything consequential, ask where the full video is and who else has it.
Voice is the frontier where humans test worst of all, and it’s already weaponized in family-emergency scams using clones built from seconds of social media audio. The defenses aren’t audio analysis, they’re protocol: verify through a channel you initiate, and agree on a family code word, exactly as we laid out in our WhatsApp scams guide. A clone can copy a voice perfectly; it can’t know the word your family never posted.
The 60-Second Check Before You Share
For any image or clip that’s shocking, infuriating, or too perfect:
- Source first (10 seconds). Who posted this, and is any credible outlet carrying it? An earth-shaking event existing in exactly one anonymous account’s post is its own answer.
- Reverse image search (15 seconds). Google Lens the frame. Recycled context and existing debunks surface here.
- Provenance check (15 seconds). Right-click in Chrome for the AI check, or run it through contentcredentials.org. Presence of a mark or credential settles it; absence means continue.
- Zoom pass (15 seconds). Jewelry, hands-meet-objects, shadows, background texture, at real magnification.
- The feeling check (5 seconds). If the content is engineered to make you furious or vindicated, that’s precisely when fakes spread, and precisely the mechanism your feed’s ranking system rewards, as we’ve covered elsewhere. Strong emotion plus single source equals wait.
Nothing conclusive after all five? Then the honest position is “unverified,” and unverified things don’t get shared with your name on them.
Quick Answers
- Can you reliably spot AI images by eye in 2026?
No. Controlled studies show people perform at chance level against current fakes. Visual tells are leads at best; verification comes from source and provenance checks. - Are AI image detector tools accurate?
Against the newest generators, often not: independent tests show accuracy falling as low as 18 to 30%, despite high advertised figures. Treat scores as one signal, never proof. - What’s the most reliable way to verify an image?
Reverse image search plus provenance: Content Credentials at contentcredentials.org and SynthID checks now built into Chrome. When credentials or watermarks are present, they’re cryptographic evidence, not guesses. - If an image has no AI watermark, is it real?
Not necessarily. Open-source generators embed nothing and metadata can be stripped. Absence is inconclusive; presence is what settles questions. - How do I protect my family from voice deepfakes?
Verify by calling back on a number you already have, and set a family code word that’s never been posted online. Protocol beats ear.
The Bottom Line
The pixel-inspection era is ending, and pretending otherwise is how people get fooled twice: by fakes that look clean and by real images they’ve learned to doubt. What replaces it is already in your hands: a reverse image search, a right-click provenance check, a zoom pass for the tells that still linger, and the discipline to treat single-source outrage as unverified by default. The generators will keep improving and the detectors will keep chasing them, but “where did this come from?” is a question that doesn’t age. Ask it for sixty seconds before you share, and you’re ahead of almost everyone your misinformation will ever come from.
