Guide to Spotting Deepfakes in Images and Videos

Deepfakes are increasingly being used in scams, misinformation campaigns, and fraud. For example, fraudsters created a deepfake video of YouTube’s CEO to trick creators into believing false policy changes and steal their credentials. Fortunately, many of these fakes are successfully caught by detection tools.

In 2024, the nonprofit TrueMedia flagged a fake video of a Ukrainian official as “highly suspicious,” with 100 percent confidence that it was AI-generated. Similarly, speech forensics tools like Pindrop have been used to detect deepfake audio, such as fraudulent robocalls impersonating U.S. leaders. These examples underscore the urgent need for reliable tools to verify suspicious media before it is shared or believed. Detection not only helps prevent deception, but also plays a critical role in maintaining trust in authentic content.

How Deepfake Detection Tools Work

AI deepfake detectors use a combination of computer vision, forensic analysis, and machine learning to spot signs of manipulation. These tools look for subtle artifacts that often go unnoticed by the human eye, such as:

• Facial inconsistencies: For example, abnormal eye movements or blinking, mouth shapes that don’t match the spoken phonemes, and distorted facial expressions. These can be signs of a face swap or lip sync deepfake.
• Digital artifacts: Small glitches in pixels, blurring around the edges of a face, or inconsistent lighting and shadows. For example, deepfakes may show inconsistent skin texture or unusual color tones around the face compared to the background.
• Biometric signs: Anomalies in natural human signals. An advanced method detects pulse or heart rate from video. Some detectors analyze color changes in skin pixels caused by blood flow. Real faces show tiny pulse-induced color shifts, while fake faces don’t. This technique can help catch even highly realistic fake video in controlled tests.
• Audio-visual mismatches: Tools analyze whether the spoken words (phonemes) match the lip movements (visemes). For example, if the lips form an “O” sound but the actual sound is an “E,” that’s a red flag. Researchers have developed detectors that focus on these mismatches to identify video deepfakes.
• Metadata and fingerprints: Some tools examine metadata or embed digital watermarks in authentic content. For example, OpenAI uses hidden metadata in generated images as a cryptographic signature (following the C2PA standard). Detection software reads these tags to determine if the content is AI-generated. Other tools, such as Attestiv, create a secure digital fingerprint of a video on a blockchain. If any pixel is altered, the mismatch in the fingerprint will reveal the change.

No single method can catch all deepfakes, so modern detectors often combine multiple techniques. For example, an AI tool might perform both pixel-level analysis and deep learning classification of the face. The bottom line: The best detectors examine details that humans often overlook, from invisible pixel glitches to biological signals, in order to determine whether media is real or fake.

Comparing Top Deepfake Detection Tools (Free vs. Enterprise)

In 2025, there are several deepfake detection tools available, from free web tools for everyday users to more advanced enterprise platforms. Below is a comparison of some of the top AI tools for detecting fake images and videos. This comparison covers features, ease of use, accuracy, platform support, and cost:

Free & Open-Source Deepfake Detection Tools

These tools are accessible to the public, often through web platforms or apps, and do not require paid licenses. Although they may have usage limits, they are perfect for individuals and journalists on a budget.

Tool & Source	Media Types	Key Features	Platform	Accuracy / Notes	Cost
TrueMedia Detector (nonprofit)	Images, Video, Audio	Combines multiple AI detectors; flags suspect content with a probability score. Provides a simple verdict (“highly suspicious” if likely fake). Developed for journalists by an AI research team.	Web (browser) – upload or URL analysis	~90% overall accuracy across media types (extremely high for known deepfakes). Example: correctly flagged a fake Ukraine official video with 100% confidence.	Free (open to journalists & fact-checkers; non-partisan initiative)
Microsoft Video Authenticator	Images & Videos	AI model analyzing visuals for subtle blending boundaries & pixel-level anomalies. Outputs confidence score per frame (in real-time for video).	Windows app / Web (released to media partners; some versions as browser tool)	Microsoft hasn’t publicly disclosed exact accuracy; frame-by-frame scoring identifies face swaps with high confidence.	Free (for democracy preservation initiatives)
Deepware Scanner (Deepware AI)	Videos (and audio)	User-friendly web scanner for deepfake videos. Upload a video or input a URL, runs a cloud AI model for deepfake clues. API/SDK available for developers.	Web app (browser); Mobile app	Continuously trained on new deepfakes. Open-source project noted for combining several detection algorithms.	Free (no cost for web tool; API requires registration)
InVID / WeVerify Plugin (EU project)	Images & Videos	Browser extension suite for fact-checkers. Breaks a video into key frames; includes error-level analysis, reverse-image search, and a deepfake detector built on academic models.	Browser extension (Chrome, Firefox)	Widely used by journalists. Leverages public fake-video database. Relies on research-based detectors, accuracy improves with updates.	Free & Open-Source (EU-funded)
Illuminarty AI Detector	Images	Online tool specializing in AI-generated image detection. Analyzes pixels/patterns to estimate probability of AI creation (e.g. Midjourney, DALL·E). Provides percentage likelihood and model guess.	Web (also has an API)	Correctly identifies popular AI-generated images even when metadata is stripped. Can be fooled by heavy post-processing in rare cases.	Free (API may have a free tier)
FaceForensics++	Images & Videos (faces)	Open-source database & model suite for deepfake detection. 1.8M+ fake images, hundreds of videos for training detectors. Includes pre-trained models for those with ML expertise.	Code library (requires Python/ML skills)	A research benchmark widely used by detection algorithms. General public might not use it directly, but it powers many other tools.	Free & Open-Source (non-commercial use)

As mentioned earlier, free tools can be very effective. For example, TrueMedia’s web tool provides an all-in-one solution for analyzing images, videos, and audio with high accuracy. On the other hand, browser plugins like InVID offer a more detailed forensic approach by breaking down suspicious videos frame by frame.

Ease of use varies across tools. Deepware and Illuminarty make it easy by simply requiring users to upload a file or paste a link, while FaceForensics++ is intended for experts to build custom models. Most free tools provide a straightforward yes/no result or a probability score, which is easy to understand. However, users should stay cautious—false positives are possible, and a “clean” result doesn’t always guarantee the content is authentic.

Nonetheless, these accessible tools provide essential detection capabilities at no cost, making it easier for everyday users to spot deepfakes.

Enterprise-Grade Detection Solutions

For organizations, governments, and users who need advanced detection capabilities, several enterprise-grade platforms offer more comprehensive solutions. These platforms often combine multiple AI models, large datasets, and real-time monitoring. They typically include APIs, dashboards, and dedicated support services, though they come at a cost. Below is a comparison of some notable enterprise-level deepfake detection tools available in 2025:

Tool / Platform	Features & Capabilities	Accuracy & Scale	Supported Platforms	Usage & Cost
Hive AI Detection API	Multi-modal detection for images & videos (focus on faces). Automatically identifies faces, labels “deepfake” with confidence score. Real-time content moderation: screens videos at upload. Trained on vast dataset (incl. deepfake porn & misinfo videos).	High accuracy for spotting AI-manipulated faces. Backed by U.S. Defense Dept ($2.4M investment). Used in production by social networks to quickly catch banned content.	Cloud API and SDK (integrates into apps). Web dashboard for analysis.	Commercial (Enterprise API license, priced by usage). Some free detections for trial; Chrome extension for AI image detection.
Sensity AI Platform	All-in-one deepfake detection hub (video, images, audio, text). Multimodal scanning for face swaps, voice clones, synthetic text. Monitors 9,000+ sources continuously & alerts clients of new deepfakes. SDK for live identity verification.	~95–98% accuracy on known deepfakes. Detected 35,000+ malicious deepfakes in the wild. Scalable for bulk real-time processing.	Web app with user-friendly UI, API, and on-prem options. Suitable for law enforcement & major enterprises.	Commercial (Subscription/license). Demo/trials available for qualified organizations.
Reality Defender	Ensemble detection across multiple AI models + proprietary algorithms. Flags deepfakes in video, images, audio, documents. Focus on fraud prevention: used by banks for ID & voice verification. Real-time screening + threat intelligence feed of known fakes.	Widely adopted by govt agencies & broadcasters in Asia. Identified deepfake voice scams & forgeries missed by simpler checks. RSAC 2024 finalist, $15M funding.	Cloud-based (API & web portal). Integrates with existing security systems (call centers, media DBs).	Commercial (Enterprise pricing). Focus on large organizations, customized service.
Intel FakeCatcher	Real-time video deepfake detector analyzing blood flow signals (PPG) in faces. Checks authenticity signals instead of errors. Can run 72 video streams simultaneously on specialized Intel hardware. Ideal for live video screening (social media, broadcast).	~96% accuracy in controlled tests (~91% in real-world videos). Rapid analysis in milliseconds. Requires good video quality for best results.	On-prem server or cloud instance (optimized for Intel hardware). Web-based interface for uploads/streams.	Commercial (contact Intel for solutions). Not a direct consumer app; enterprise/broadcast focus.
Attestiv (Enterprise Suite)	Digital content authentication platform (AI forensic analysis for videos/images). Generates “Overall Suspicion Rating” from 0–100. Secure fingerprinting on a ledger to detect any pixel-level tampering. Context analysis (metadata, transcripts).	Produces detailed video reports (timeline of suspicious edits). High precision for video manipulations. Might not catch all manual (non-AI) edits.	Web portal & API; mobile SDK for on-device capture verification. Scalable with encryption & compliance features.	Freemium / Commercial (Free tier: 5 video scans/month). Paid plans for bulk scanning & enterprise usage.
DuckDuckGoose “DeepDetector” (Startup)	Explainable deepfake detection (frame-by-frame analysis). Highlights manipulated regions, used for identity protection in finance & brand integrity. Real-time alerts when a fake involving clients appears online. Dashboard called “Phocus” with detailed AI explanation.	Claims ~99% accuracy in internal tests, “seconds” detection. Used by banks in fraud prevention. Award-winning (Best AI Company 2024).	Cloud platform with web dashboard + API integration. Clients can review flagged content & see how it was identified.	Commercial (Subscription, custom pricing). Aimed at enterprise; demo requests available.

Enterprise solutions provide more comprehensive coverage and support. For example, platforms like Sensity and Reality Defender handle not only videos but also fake audio and text, offering an all-in-one solution for organizations managing various types of synthetic media. These tools integrate smoothly into workflows, enabling automatic scanning of content uploads or generating forensic reports for investigators. Accuracy rates for these tools are reported to be very high (often above 95% for known deepfake types), though it’s important to keep in mind that these results come from controlled test environments. In real-world use, even enterprise-grade tools can be bypassed by new deepfake techniques or simple tactics, like adding noise or resizing images.

A big advantage of enterprise tools is their ability to scale and provide real-time monitoring. Platforms like Hive and DuckDuckGoose can scan thousands of social media videos in real time and automatically take action, such as alerting moderators, to prevent a fake from going viral. This is especially important during events like elections, where rapid detection is critical. These platforms also offer easy-to-use dashboards with visual explanations. For example, DuckDuckGoose not only flags an image as fake, but also shows exactly where and how the manipulation was detected, highlighting the altered areas, which helps build confidence in the results.

Finally, enterprise solutions usually come with dedicated customer support and customization options. For example, a company could train a detector specifically to identify subtle deepfakes of their CEO’s face for greater accuracy. While these enterprise tools are incredibly powerful, they do require significant resources. For general public use, free tools remain the go-to option, though it’s important to note that advanced detection technology often appears first in enterprise solutions before becoming available to the wider public.

Step-by-Step: Verifying an Image’s Authenticity

Even without deep technical knowledge, you can take simple steps to verify a suspicious image. Below is a beginner-friendly guide that combines common sense checks with AI tools:

Inspect the Image Closely

Start by using your own eyes. Deepfakes or AI-generated images often contain subtle, odd details:

• Look for inconsistencies: Does the person have asymmetrical or missing earrings? Are the hands or fingers distorted, or are there too many? Count elements like teeth or limbs — older AI models often produced extra fingers or teeth, though newer versions have improved.
• Check textures and lighting: Is the skin too smooth or too perfect, lacking natural pores or wrinkles? Does the lighting on the face match the shadows in the background? For example, an AI-created face might have a flat, airbrushed look or lighting that doesn’t match the environment’s light sources.
• Look at text and background details: AI often struggles with legible text on signs or clothing. Blurry or gibberish text (like an unreadable stop sign or nonsensical license plates) is a red flag. Background people or objects might also appear distorted or unnatural. If anything seems off upon close inspection, remain skeptical.

Check the Image Metadata

If you have the image file, right-click and view its properties or use a metadata viewer. Authentic photos (especially from cameras or smartphones) typically contain EXIF data, such as camera model, timestamp, and GPS coordinates. Deepfakes or AI-generated images might have stripped or odd metadata, like a software name such as “StableDiffusion” as the creator. However, the absence of metadata alone isn’t proof of manipulation—it’s just a clue.

Use Reverse Image Search

Perform a reverse image search on platforms like Google Images or TinEye. The goal is to see if the image (or parts of it) have appeared before. Often, deepfake images are derived from real ones. For example, a fake image of a celebrity in an unusual scene might be based on an original photo with a swapped face. If reverse search returns the original (with a different face or context), you’ve caught the fake. If no result is found, it could be a new or AI-generated image — proceed to specialized tools.

Leverage AI Image Detectors

• TrueMedia (web tool): Upload the image to get a probability score indicating whether the image is AI-manipulated. If it’s flagged as “highly suspicious” with a high confidence rate (e.g., 90–100%), it strongly suggests a deepfake.
• Illuminarty or Hive’s AI Detector: Upload the image to Illuminarty or use Hive’s AI-or-Not tool. These tools will return verdicts like “98% likely AI-generated” or “Likely authentic.” Compare results across tools: if both say it’s AI-made, it probably is. If results differ, continue investigating.
• Error Level Analysis (ELA) tools: Tools like FotoForensics or InVID’s ELA feature show image compression levels. If parts of the image, like the face, show different error levels compared to the rest, it may indicate manipulation.

Cross-Verify with Trusted Sources

If the image is newsworthy (e.g., a politician doing something shocking), check reputable news outlets or fact-checkers like AFP, Reuters, or Snopes. They may have already debunked the image. Even the best AI tools can produce false results, so it’s smart to check if human experts have weighed in.

Interpret the Results Cautiously

If multiple tools flag the image as fake, it’s likely a deepfake. However, context still matters. A “real” result from a detector doesn’t guarantee authenticity—it simply means the tool didn’t find known markers of fakeness. Consider the source: Who posted the image? Is it plausible? When in doubt, seek expert opinion or avoid sharing it as fact.

By combining visual inspection, metadata analysis, reverse searches, and AI tools, you can verify images more confidently. For instance, the viral fake image of the Pope in a designer puffer jacket was exposed using these steps: strange hand details, reverse image search findings, media reports, and AI detectors all pointed to it being AI-generated. Using multiple steps together is the best way to reach a reliable conclusion.

Step-by-Step: Verifying a Video’s Authenticity

Detecting deepfake videos can be more challenging than still images, as there’s motion, audio, and multiple frames to consider. However, the core principles remain the same. Here’s a step-by-step guide to help you spot deepfake videos:

1. Assess the Video Source and Context

Before diving into frame analysis, consider the source of the video. Is it from a random social media post or an official channel? If it’s supposedly from a known figure (e.g., a politician or celebrity), check their official pages to see if the video is posted there. Many deepfakes spread via untrustworthy sources. A lack of a reputable origin is a red flag.

2. Watch Closely for Visual Artifacts

Play the video and watch for any details that feel “off.” Key indicators include:

• Face and Expression: Do facial features stay stable during motion, or do you see flickering or blurred edges? Does the lip-sync match the speech?
• Eyes and Gaze: Are blinking patterns natural? Do the eyes move realistically? Look out for “glassy” eyes or inconsistent reflections.
• Body and Hands: Is the body language typical for the person? Are there size mismatches or strange proportions?
• Lighting and Shadows: Do lighting and color match across the face, neck, and background? Are shadow movements consistent with the environment?
• Audio cues: Does the emotional tone of the voice align with the facial expression? Are there robotic sounds or mismatched accents?

3. Break the Video into Frames

Use tools like the InVID plugin to extract frames or pause the video at key moments. Some deepfakes look fine in motion but glitch when viewed frame-by-frame. Slowing playback (e.g., to 0.25×) can reveal flickers or blending issues that aren’t noticeable at full speed.

4. Run the Video Through AI Detection Tools

• Deepware Scanner: Upload the video or enter a URL to get a verdict, such as “deepfake detected” along with a confidence percentage.
• Attestiv: This platform scores video segments with suspicion levels (e.g., 85/100), highlighting issues like mismatched lip-sync or face anomalies.
• Hive Moderator / Sensity: Enterprise tools that scan faces frame-by-frame, labeling each as real or fake. Available via trial or API access.
• TrueMedia’s Video Analyzer: Upload a clip to receive a composite verdict from multiple detectors, including a statement like “Highly likely AI-generated imagery in video.”

5. Examine the Detector Output

If a tool flags the video as a deepfake with high confidence, trust the assessment. These tools are designed to minimize false positives. If results are unclear, use a second tool for confirmation. Pay attention to time-stamped sections where problems were detected, and review those parts manually.

Also, consider cheapfakes—edited real videos that AI detectors might miss. Compare the video to known originals to spot manual tampering or out-of-context edits.

6. Cross-Check with Trusted Sources

If the video is newsworthy (e.g., a shocking moment involving a public figure), check reputable fact-checking outlets like Reuters, AFP, Snopes, or others. They may have already debunked the video. While AI tools are powerful, human expertise still plays a crucial role in verification.

7. Use Common-Sense Judgment

After gathering results from tools and considering the context, make an informed decision. If the evidence points toward it being a fake, treat the video as untrustworthy. Even if a tool says it’s real, remain cautious. AI detection in 2025 is highly advanced, but not flawless. Context, consistency, and credibility remain essential factors in determining authenticity.

Example in Action:

Take, for example, the case of a European mayor tricked into a deepfake video call. The impostor’s face was generated in real-time. Tools like FakeCatcher or Hive would likely flag the unnatural gaze and absence of biometric cues like micro pulse changes. The key takeaway: inspect carefully, validate with tools, and confirm with trustworthy sources.

Summary: Deepfake video detection is within reach for everyone using this method:

• Check the source
• Inspect visuals and audio closely
• Pause or extract frames for detailed analysis
• Use multiple AI detectors for verification
• Cross-check with official news or fact-checkers
• Use critical thinking — and when in doubt, don’t share