How to tell what's real and what's AI-generated on social media

01The current state of AI-generated content

The explosion of generative AI tools has flooded social media with synthetic images, voice clones, and deepfake videos at an unprecedented scale. What once required Hollywood-level budgets and specialized teams can now be produced by anyone with a laptop and a subscription. AI-generated content is no longer a novelty — it is a daily reality on every major platform, from Facebook and X to TikTok and Instagram.

According to research compiled by Wikipedia contributors tracking synthetic media, the volume of AI-generated content online doubled between 2023 and 2025, with deepfake videos alone growing by over 550% year over year. The TODAY segment that anchors this article walks through real examples viewers encountered on their own feeds, demonstrating just how convincing — and how widespread — these fakes have become.

The core problem is not just that fakes exist, but that they spread faster than corrections. A fabricated image of an explosion near the Pentagon in 2023 briefly caused stock market dips before it was debunked. Political deepfakes have appeared in election cycles worldwide. The gap between creation and detection is narrowing, but it has not closed.

02Visual artifacts that reveal AI images

AI image generators — Midjourney, DALL-E, Stable Diffusion, and others — still leave detectable fingerprints if you know where to look. The most common artifacts appear in hands, teeth, eyes, and background text. AI models frequently produce six fingers, merged teeth, mismatched earrings, or gibberish text on signs and labels.

Lighting and shadow inconsistencies are another tell. A subject's face may be lit from one direction while the shadow falls the opposite way. Reflections in water, glass, or mirrors often fail to match the scene. Skin textures can look waxy or overly smooth, lacking the pores, blemishes, and asymmetry of real photographs.

Backgrounds are where AI struggles most. Distant crowds may have melted faces. Architecture can feature impossible geometry — windows that lead nowhere, staircases that defy physics. Zooming into an image and scanning edges and fine detail often reveals these inconsistencies before they are visible at full size.

03Voice cloning detection

Voice cloning technology has reached the point where a few seconds of audio can produce a convincing replica of someone's voice. This has enabled a surge in voice-cloning scams, including fake kidnapping calls and fraudulent messages impersonating executives. The threat is not theoretical — families have paid ransoms after hearing what sounded like their child's voice crying for help.

Detection of cloned voices is harder than image detection because the artifacts are acoustic, not visual. Cloned speech often lacks natural breathing patterns, has slightly off cadence, or produces unusual prosody on emotional words. Background noise can sound artificially consistent — real phone calls have variable audio quality, while cloned calls may be suspiciously clean.

The best defense is verification through a secondary channel. If someone calls asking for money or sensitive information, hang up and call them back on a known number. Voice alone should never be treated as proof of identity in high-stakes situations.

04Deepfake video tells

Deepfake videos present the highest-stakes challenge because they combine visual and audio manipulation. The TODAY segment highlights several detection methods viewers can apply at home. Key deepfake tells include unnatural blinking patterns, mismatched lip-sync, boundary artifacts around the face, and inconsistent skin tone between the face and neck.

Older deepfakes struggled with blinking — subjects either never blinked or blinked at unnatural intervals. While newer models have improved, micro-expressions around the eyes and mouth still lag behind real human movement. Hair, teeth, and jewelry often show blending artifacts at their edges where the synthetic face meets the original footage.

Audio-video sync is another strong indicator. If the audio track drifts even slightly from the mouth movements, or if the room tone changes when the speaker is on-screen versus off, the video may be manipulated. Slow-motion playback can reveal frame-level inconsistencies that are invisible at normal speed.

05Reverse image and source verification

One of the most powerful tools for verifying any image — AI-generated or not — is reverse image search. Google Images, TinEye, and Yandex can trace where an image first appeared online. If a viral photo has no prior history or first appeared on an anonymous account with no track record, that itself is a red flag.

For videos, checking the source account's history matters. Does the uploader have a pattern of original reporting, or is the account new with limited activity? Cross-referencing with established news outlets is critical — if no reputable organization is reporting the event shown in a viral video, skepticism is warranted.

Metadata can also help. While metadata can be stripped, EXIF data, upload timestamps, and geolocation tags — when present — can corroborate or contradict the narrative being presented. Tools like InVID and the FotoForensics platform provide forensic analysis for journalists and curious users alike.

06Tools for detecting AI content

A growing ecosystem of detection tools has emerged to counter the flood of synthetic media. AI image detectors like Hive Moderation, Sensity, and the Deepfake Detector analyze images and videos for synthetic artifacts, with accuracy rates that vary by content type. Detection is an arms race — as detectors improve, so do generators.

Platform-level tools are also being deployed. Meta labels AI-generated content on Facebook and Instagram when it is detected or disclosed by the creator. Google's SynthID watermarking embeds imperceptible signals in AI-generated images and audio that can be checked later. However, these tools are opt-in for creators and not universally applied.

For the average user, browser extensions like FakeGuard and the NewsGuard extension provide real-time flags for questionable content. The Reality Defender platform offers API access for organizations that need to screen large volumes of media. No single tool is perfect — the best approach combines multiple detection methods with human judgment.

07Why media literacy matters more than ever

Technical detection tools will always be a step behind generative AI. This makes media literacy the most durable defense against synthetic media. Understanding how AI fakes are made, recognizing their common tells, and applying critical thinking before sharing content are skills that scale with the threat.

Schools, libraries, and community organizations have begun incorporating synthetic media awareness into digital literacy curricula. The Media Literacy Now initiative tracks legislation across US states requiring media literacy education in K-12 schools. Internationally, organizations like UNESCO have published guidelines for teaching citizens to navigate the synthetic media landscape.

The shift from trust-by-default to verify-by-default is not cynicism — it is the natural adaptation to a media environment where reality itself can be fabricated. The TODAY segment closes with a practical message: slow down, verify, and think before you share. In an age of synthetic reality, that may be the most important habit of all.