Digital Life · Verify
Visual tells come and go as tools improve. Verification habits stay. This guide teaches both, and tells you which to trust more.
The principle
AI-generated images, voices, and video are improving faster than the visual artifacts that give them away. A year ago, AI-generated hands were often distorted. Today, most tools produce realistic hands. The specific tells in this guide reflect what works now, but they will degrade as generation technology advances.
Verification habits, by contrast, work regardless of how good the tools get. Checking whether a second independent source confirms what you are seeing, calling someone back on a number you already have rather than one they just gave you, and pausing before acting on content that triggers a strong emotional reaction: these behaviors do not depend on spotting a visual artifact. They work because they check the source, not the surface.
This page teaches both. Learn the current tells for awareness. Build the habits for durability.
This page contains time-sensitive content. The tells section describes artifacts that are accurate as of mid-2026. Generation tools improve continuously, and some of these tells may no longer apply by the time you read this. The habits section is designed to remain useful regardless. Check the review date at the bottom of this page before relying on any specific tell.
Build these first
These five habits protect you regardless of how convincing AI-generated content becomes. They work because they verify the source and context, not the pixels or audio waveform.
Before sharing or acting on a striking image or video, ask: who published this, and can I find it on an established news outlet or official account? Content that exists only on one social media post, with no confirming coverage anywhere else, deserves scrutiny regardless of how real it looks.
If someone calls claiming to be a family member, your bank, or a government agency, hang up and call back using a number you already have in your contacts or on your paper statement. A cloned voice cannot survive a callback to the real person. This single habit stops most voice-cloning scams.1
A pre-agreed word or phrase that anyone in the family uses to verify an unexpected call asking for money or urgent action. Voice cloning can replicate how someone sounds. It cannot reproduce information the cloner does not have. See The Family Code Word for setup instructions.
AI-generated content is most dangerous when it triggers urgency, outrage, or fear. If a video or call makes you feel you must act immediately, that urgency is the threat. Pause. Verify through a second channel. Real emergencies survive a five-minute delay for confirmation.2
Upload a suspicious image to a reverse image search (Google Images, TinEye). If the image appears nowhere else on the internet, or only on recently created accounts, treat it with caution. Legitimate news photos almost always appear on multiple established outlets.
Current tells · Images
Time-sensitive section. These artifacts reflect mid-2026 generation tools. Some may no longer apply when you read this. The habits above remain reliable regardless.
Current image generation tools (diffusion models like those behind Midjourney, DALL-E, and Stable Diffusion) produce images that look realistic at first glance but often contain subtle inconsistencies in areas the model handles less reliably.
Signs, labels, and text within AI images frequently contain misspellings, inconsistent letter sizes, or characters that blend into each other. Zoom in on any text in the image.
Reflections in glasses may not match the scene. Earrings or jewelry may differ between the left and right side. Symmetrical accessories are a stress point for current models.
Look at edges where the subject meets the background. Blurred transitions, objects that fade into nothing, or architectural features that do not make structural sense (a railing that connects to nothing, a window at an impossible angle) are common.
Current tools handle hands much better than a year ago, but still occasionally produce extra fingers, fused digits, or unnatural grips. This tell is weakening and may not be reliable much longer.
AI-generated faces sometimes look unusually smooth, with pores and fine lines that appear painted rather than organic. The texture may look too uniform across the entire face.
Images with multiple people are harder for current models. Look for people who blend into each other, limbs that belong to no one, or faces in the background that dissolve into abstraction.
Current tells · Voice and video
Time-sensitive section. Voice cloning is the fastest-improving category. The FBI noted in its 2025 IC3 report that AI-generated content is advancing to the point where it is "often difficult to identify."1 Rely on the callback habit, not on your ear alone.
Modern voice cloning can replicate a person's voice from as little as a few seconds of sample audio. The resulting clone can sound convincing enough to fool family members. The FBI reported $893 million in AI-enabled fraud losses in 2025, with voice cloning used in family emergency scams, business impersonation, and romance fraud.3
Flat emotional range. Current clones may sound like the person but struggle with natural emotional variation, laughing, crying, or shifting tone mid-sentence.
Unnatural pauses or pacing. The rhythm of conversation may feel slightly off, with responses that come too quickly (pre-generated) or too slowly (processing delay).
Background silence. A cloned voice layered onto a call may lack the ambient sounds you would expect from where the person says they are calling from.
These tells are becoming less reliable as tools improve. The callback habit and the family code word remain the strongest defenses.
Face-swap and real-time deepfake video tools are now used in video calls, not just pre-recorded content. The FBI's July 2026 advisory documented scammers using AI-generated video of FBI officials to impersonate law enforcement.4
Edge artifacts on head turns. Most real-time face-swap tools lose coherence when the subject turns past about 45 degrees. Watch for blurring, flickering, or warping along the jawline and hairline.
Hand-over-face test. Asking someone on a video call to briefly touch their face or adjust their glasses can destabilize current face-swap tools, which struggle when a foreground object crosses the face plane.
Lighting inconsistencies. The lighting on the face may not match the lighting in the rest of the scene, especially around the neck and shoulders.
The household question
For most households, the highest-risk use of AI-generated content is not a viral fake image on social media. It is a phone call that sounds like a grandchild asking for money, a text that sounds like a spouse asking for a login, or a voicemail that sounds like a boss requesting a wire transfer.
These attacks work because they bypass reasoning and trigger an immediate emotional response: someone I love is in trouble, or someone with authority is asking me to act now. The defense is not better ears or sharper eyes. It is the habit of pausing and verifying through a channel the attacker does not control.
Talk to your household about this. Make sure everyone knows the family code word and the one-call rule: any unexpected request for money or sensitive information gets a callback on a known number before anyone acts.
For how AI-generated content intersects with news and crisis information, see Information Hygiene, which covers verifying public information during emergencies.
Watch for these
Assuming that real-looking means real.
The absence of visible artifacts does not mean the content is genuine. Source verification matters more than visual inspection.
Assuming that fake-looking means AI.
Low-quality video, compression artifacts, and bad lighting exist in genuine content too. A blurry video is not automatically a deepfake. Check the source before concluding.
Trusting a detection tool as a final answer.
Detection accuracy varies by model and media type. Use tools as one input, not the verdict. A tool that says "likely AI" or "likely real" is giving a probability estimate, not a fact.
Memorizing tells without building habits.
Tells expire. Habits compound. If you only learn to spot extra fingers, you are unprotected the moment that artifact disappears. The callback, the code word, and the pause before acting protect you regardless.
Keep going
The single most effective defense against voice-cloning scams. Five minutes to set up, costs nothing.
AI is a tool scammers use, but the underlying patterns (pressure, urgency, secrecy) never change.
Verifying public information during emergencies, when AI-generated content is most dangerous.
This page was last reviewed in July 2026. AI-generated content evolves faster than any other topic on this site. The verification habits in this guide are designed to remain useful regardless of how tools change. The tells sections describe artifacts that were accurate at the time of review, but generation tools improve continuously. If the review date above is more than six months old, treat specific tells with caution and rely on the habits instead. Confirm key claims against the sources above before acting on them.
Enough for now
Revisit the tells section every six months. The habits section should not need updating.
This guide is part of When Something Looks Off — all the guides for this concern in one place.