Guide

How to spot a deepfake

Deepfakes are AI-generated videos, images, and audio recordings that can make it look like someone said or did something they never actually did. The technology has become remarkably convincing, but deepfakes still leave telltale signs that you can learn to recognize. This guide walks you through the visual, audio, and contextual clues that reveal deepfakes, plus the tools and techniques you can use to verify any suspicious content you encounter online.

What are deepfakes and why they matter

How deepfake technology works (simplified)

At a basic level, deepfakes use artificial intelligence to swap one person's face onto another person's body in video, clone someone's voice from a short audio sample, or generate entirely new photos and videos of people who do not exist. The AI studies large collections of images or audio to learn what a person looks like or sounds like, then uses that learned representation to create new, fabricated content.

The term "deepfake" originally referred specifically to face-swapped video, but it now encompasses any AI-generated media intended to represent something that did not happen. This includes face swaps, lip-synced videos where the speaker's mouth is altered to match different audio, entirely AI-generated faces and scenes, cloned voices, and AI-written text attributed to real people.

The scale of the deepfake problem in 2026

Deepfake production has accelerated dramatically. The tools for creating deepfakes are freely available, require no technical expertise, and run on consumer hardware. Some mobile apps can generate a face-swapped video in under a minute. This accessibility means that deepfakes are no longer a niche concern for public figures. Anyone with photos or video online (which is nearly everyone) has enough material available for someone to create a convincing deepfake of them.

Who creates deepfakes and why

The motivations behind deepfake creation range from entertainment to serious harm. The largest category by volume is non-consensual intimate imagery, which accounts for roughly 96% of deepfake videos identified online. Beyond that, deepfakes are used for financial fraud (impersonating executives to authorize wire transfers), political manipulation (fabricated statements by politicians), harassment and reputation damage, identity theft, and misinformation campaigns. Understanding the intent helps you evaluate the context when you encounter suspicious content.

Visual signs of deepfake video

Face boundary artifacts and blending edges

The most common visual artifact in face-swapped deepfakes appears at the boundary where the inserted face meets the original head. Look carefully at the edges of the face along the jawline and around the hairline. You may notice a slight color mismatch, a visible seam, or an unnatural softness where the blending occurs. The transition between skin tone on the face and skin tone on the neck or ears is often imperfect.

To check for this, pause the video and zoom in on the face boundary. Compare the skin texture and color on the cheek with the area just outside the face region. In authentic video, these areas blend naturally. In deepfakes, you may spot a subtle but distinct line where the generated face ends and the original footage begins.

Eye reflections and gaze inconsistencies

Human eyes produce specular reflections (the small bright spots you see when light hits the eye). In authentic photos and video, these reflections are consistent between both eyes because both eyes are seeing the same light source. Research has shown that deepfakes frequently produce inconsistent eye reflections, with different shapes, positions, or numbers of reflection points between the left and right eye.

Gaze direction is another tell. In conversation, people naturally shift their gaze, look at the person they are addressing, and blink at regular intervals. Deepfakes sometimes show an unnaturally fixed gaze, limited blinking, or gaze directions that do not match the conversational context (appearing to look slightly past the camera when they should be making eye contact).

Teeth, ears, and hair anomalies

Fine detail is difficult for AI models to generate consistently. Pay close attention to teeth, which should show individual tooth boundaries, consistent sizing, and natural alignment. Deepfakes sometimes produce teeth that appear blurred, merged, or inconsistently shaped across different frames. Earlobes and ear structures are similarly challenging for generators, and may look asymmetric or unnaturally smooth.

Hair presents particular problems for deepfakes. Individual strands, flyaway hairs, and the boundary between hair and background are computationally expensive to generate accurately. If the hair looks unusually smooth, lacks individual strand detail, or shows a suspiciously clean edge against the background, that may indicate AI generation.

Temporal flickering and frame-to-frame inconsistency

Deepfake generation typically processes one frame at a time or in small batches. This can produce subtle inconsistencies between consecutive frames that appear as flickering or "shimmer" in the face region. The face might shift slightly relative to the head, change brightness, or show momentary artifacts when the subject turns or changes expression.

To test for this, slow the video down to quarter speed or step through it frame by frame. Authentic video shows smooth, continuous motion. Deepfakes may reveal momentary glitches, face position jumps, or brief frames where the face generation partially fails before recovering.

Lighting and shadow mismatches

When a face from one source is placed into footage from another source, the lighting conditions often do not match perfectly. The direction, intensity, and color temperature of light on the face should be consistent with the lighting on the body, the environment, and any shadows visible in the scene. A face that appears front-lit in a scene where the background lighting comes from the side, or shadows on the face that fall in a different direction than shadows on the clothing, suggests manipulation.

Background warping and distortion

Face-swap operations can introduce subtle warping in the background around the head. As the generated face adjusts to match head movements, the background pixels near the face boundary may shift, stretch, or compress slightly. This effect is easier to spot in scenes with straight lines or regular patterns near the subject's head, such as door frames, window edges, or striped backgrounds.

How to detect AI-cloned voices

Unnatural breathing and pause patterns

Natural speech includes breathing sounds, micro-pauses, and rhythm variations that correspond to the physical process of speaking. Voice clones often produce speech that sounds technically accurate but lacks the subtle breathing patterns between sentences and phrases. Listen for unusually even spacing between words or sentences, missing breath sounds at natural pause points, and an overall cadence that feels mechanical even if the voice itself sounds correct.

Monotone emotional range

Current voice cloning technology handles neutral, conversational speech well but often struggles with strong emotional expression. If a cloned voice needs to convey anger, excitement, sadness, or surprise, the emotional quality may sound flattened, exaggerated in an artificial way, or disconnected from the words being spoken. Compare the emotional intensity you would expect given the context with what you actually hear.

Background noise inconsistencies

Voice clones are typically generated in a clean digital environment and then overlaid onto a recording. Listen for mismatches between the voice and any background environment. If the voice sounds like it was recorded in a studio but the video shows the person outdoors, or if background noise drops to silence during speech and returns during pauses, the audio may be synthetic. Natural recordings show consistent environmental audio throughout.

Lip-sync and audio-visual mismatch

When AI-generated audio is paired with video, the lip movements may not match the speech precisely. This is particularly noticeable with consonant sounds that produce distinctive mouth shapes: "p," "b," "m" (lips together), "f," "v" (teeth on lip), and "th" (tongue visible). Watch specifically for these sounds and check whether the mouth forms the right shape at the right moment. Even small timing mismatches of 100-200 milliseconds can indicate that the audio was generated separately from the video. For deeper technical details, see our audio forensics research.

Identifying AI-generated images

Hands, fingers, and fine detail errors

Hands have been a well-known weakness of image generation models, and while recent models have improved significantly, they still produce errors. Look for extra or missing fingers, fingers that merge or branch, fingernails that appear in wrong positions, impossible joint angles, and inconsistent hand sizes between the two hands in the same image. Generation models have gotten better at hands in 2025-2026, but complex hand poses and interactions (holding objects, interlaced fingers) still frequently produce errors.

Text and typography artifacts

Text within AI-generated images remains a reliable indicator. Look for garbled letters, inconsistent character spacing, words that are almost English but contain random character substitutions, and text on signs, labels, or documents that dissolves into meaningless shapes when examined closely. While the latest generation models can produce short text strings correctly, longer text passages and multiple text elements in the same image still commonly contain errors.

Symmetry anomalies in faces and objects

AI-generated faces tend toward higher symmetry than real faces. Human faces are naturally asymmetric: one eye slightly different from the other, one ear higher, one side of the mouth slightly different. If a face looks almost perfectly symmetric, that may indicate generation rather than photography. The same principle applies to objects: AI-generated buildings, vehicles, and products sometimes show uncanny bilateral symmetry that manufactured objects rarely achieve in practice.

Texture and pattern repetition

Examine textures in clothing, walls, foliage, and other patterned surfaces. AI-generated images sometimes produce textures that repeat in unnatural ways or break down into incoherent patterns at closer inspection. Fabric weave may not follow a consistent pattern across the garment. Brick walls might show spacing irregularities. Natural foliage may contain shapes that look organic from a distance but lack the structural logic of real leaves and branches when examined closely.

Background coherence issues

The background of an AI-generated image often receives less attention from the model than the main subject. Look at background elements for logical consistency. Do architectural elements make structural sense? Do shadows in the background match the lighting direction on the main subject? Are there objects in the background that appear to merge, float, or lack spatial coherence? Background errors are often the easiest artifacts to spot because most people instinctively focus on the main subject and overlook the surroundings. For technical background on image forensic analysis methods, see our research page.

Step-by-step verification process

When you encounter content that seems suspicious, follow this systematic process to evaluate its authenticity.

Step 1 - Check the source and context

Before examining the content itself, evaluate the source. Where did this content first appear? Is the account that posted it verified or established? Does the content align with what you already know about the person or situation depicted? Is there an obvious motivation for fabrication (political context, financial fraud, personal harassment)? Content that appears suddenly from an anonymous source during a sensitive moment deserves extra scrutiny.

Step 2 - Visual inspection checklist

• check_circle Examine face boundaries, jawline, and hairline for blending artifacts
• check_circle Check eye reflections for consistency between left and right eye
• check_circle Look at teeth, ears, and hair for detail quality and consistency
• check_circle Slow video down to check for frame-to-frame flickering
• check_circle Verify lighting direction matches across face, body, and background
• check_circle Check background for warping, especially near the subject's head
• check_circle For images: examine hands, text, symmetry, and background logic
• check_circle For audio: listen for breathing gaps, emotional flatness, and lip-sync timing

Step 3 - Use reverse image/video search

Reverse image search can reveal whether a photo has been taken from another source and manipulated. Upload the image to Google Images, TinEye, or Yandex image search to find earlier or original versions. For video, capture a key frame and search for that image. If you find the original unaltered version, that confirms manipulation. If the image appears to have no prior existence online, it may be entirely AI-generated.

Step 4 - Run forensic analysis tools

Automated detection tools can identify artifacts that are invisible to human inspection. These tools analyze the mathematical properties of the media, looking for statistical signatures of AI generation, compression inconsistencies, and forensic traces that require computational analysis to detect. No single tool is perfectly accurate, so consider results from multiple tools when available.

Step 5 - Verify with AFIP forensic analysis

AFIP's forensic analysis examines content across multiple modalities simultaneously, combining biological signal analysis, spectral forensics, temporal coherence checks, and statistical modeling into a single comprehensive assessment. Upload suspicious content to the AFIP Verify tool for a confidence score and detailed forensic report that explains which specific evidence supported the determination. Unlike binary detectors, AFIP provides nuanced confidence levels and identifies the specific forensic findings rather than just outputting a label.

Free tools for deepfake detection

AFIP Verify tool

The AFIP Verify tool provides free multi-modal forensic analysis. Upload video, audio, images, or text and receive a detailed confidence assessment with evidence-based findings. The tool analyzes content using multiple independent forensic methods and reports the combined result as a confidence score from 0 to 100, along with an explanation of which forensic signals contributed to the determination.

Other available detection tools

Several other tools are available for deepfake detection. Reverse image search engines (Google Images, TinEye, Yandex) can identify manipulated or stolen images. Browser extensions from some media literacy organizations provide quick-check functionality. Academic tools from research labs at universities including MIT, UC Berkeley, and TU Munich are available for specific use cases. Platform-specific reporting features on social media sites can flag content for internal review.

What to do when tools disagree

Different detection tools use different methods and are trained on different datasets, so they will sometimes produce conflicting results. When tools disagree, consider the confidence level each tool reports (a 52% "fake" score is much less certain than a 95% score), whether the tools examine different aspects of the content, and what the visual and contextual evidence suggests. Tools that provide detailed evidence reports (showing which specific artifacts were found) are generally more trustworthy than tools that only output a binary label. When uncertainty remains, err on the side of caution and do not share or act on the content until you have stronger evidence.

Protecting yourself from deepfakes

Media literacy best practices

The most effective protection against deepfakes is a healthy skepticism toward content that produces strong emotional reactions, appears during sensitive moments (elections, crises, controversies), or comes from unfamiliar sources. Before sharing, reacting to, or making decisions based on media content, take a moment to verify it. The few minutes spent checking a source can prevent you from amplifying misinformation or falling victim to a scam.

Be particularly cautious with content that asks you to take urgent action: send money, share personal information, make an immediate decision, or forward the content to others. Deepfakes designed for fraud typically create artificial time pressure to prevent you from verifying the content before acting on it.

Reporting deepfakes on social platforms

Every major social media platform provides mechanisms for reporting suspected deepfakes. When you encounter content that you believe is a deepfake, use the platform's reporting function and select the category that best matches the violation (typically "misinformation," "manipulated media," or "impersonation"). Reporting helps platforms identify and address harmful synthetic content, and aggregate reports help improve platform detection systems over time.

Legal protections and resources

Legal protections against deepfakes vary by jurisdiction but are expanding. Many U.S. states have enacted or are considering laws specifically targeting non-consensual deepfake intimate imagery, deepfake-related fraud, and deepfake use in elections. The EU AI Act includes provisions requiring transparency for AI-generated content. If you are the target of a malicious deepfake, document the content (take screenshots, save URLs), report it to the hosting platform, and consult with a legal professional about available remedies in your jurisdiction.

Frequently asked questions

Related resources