Is This Image AI Generated? How to Tell With Free Tools

A peer-reviewed study published in Science in 2025 found that human accuracy at distinguishing AI-generated photographs from real ones had fallen to 38% — below the 50% level expected from random guessing. The generators have simply gotten better faster than human perception has adapted. A 2024 meta-analysis of 56 studies in ScienceDirect confirmed the same pattern across modalities: average human deepfake detection accuracy is 55.54%, and for high-quality deepfake videos specifically, humans are correct only 24.5% of the time.

This is not a failure of attention or intelligence. It is a technical reality: modern diffusion models like Midjourney V7 and GPT Image 1.5 generate images by optimizing for human perceptual realism at the pixel level. They are better at fooling human eyes than human eyes are at catching them. The solution is methodical: check EXIF metadata, run multiple detectors, perform reverse image search, and reserve visual inspection for targeted forensic markers — not as a first-pass filter.

Step 1: Check the EXIF Metadata Before Anything Else

EXIF (Exchangeable Image File Format) metadata is embedded in image files by cameras and phones at the moment of capture. It contains the camera manufacturer, model, lens focal length, aperture, shutter speed, ISO, timestamp, and often GPS coordinates. When an image is generated by an AI tool, no camera fired a shutter — so there is no authentic camera metadata to embed.

This makes EXIF inspection the fastest first filter available:

• Completely absent EXIF data — strong indicator, though note that WhatsApp and some social platforms strip all metadata on upload
• EXIF listing a camera model that does not exist — conclusive indicator of post-hoc manipulation
• Impossible parameter combinations — a lens with a focal length range incompatible with the listed body, or ISO values outside any consumer camera's range
• Software tags showing image editors — Photoshop or GIMP in the EXIF chain does not confirm AI, but indicates post-processing worth scrutinizing further

How to check: on Windows, right-click the image → Properties → Details tab. On Mac, open in Preview → Tools → Show Inspector → (i) tab. Free online tools including Jeffrey's Exif Viewer and ExifTool handle any format including HEIC, WebP, and RAW. Note that EXIF can be manually injected using ExifTool, so its presence is not a guarantee of authenticity — absence is the more reliable signal.

Step 2: Run a Free AI Image Detector

AI image detectors work by analyzing patterns in the pixel data that are invisible to human eyes but statistically characteristic of AI generation. The primary methods are:

• Frequency-domain analysis — converting the image using a 2D Discrete Fourier Transform reveals spectral artifacts left by GAN upsampling layers and diffusion model denoising processes. Research published in arXiv (2510.19840) showed a ResNet50 trained on frequency transforms achieving 92.8% accuracy and AUC of 0.95 on GAN detection.
• Neural classifiers — deep learning models trained on millions of labeled real and AI-generated images to detect subtle texture, edge, and noise distribution patterns at the feature level.
• Semantic inconsistency detection — identifying local regions where AI models generated plausible-looking but globally inconsistent content: shadows without correct geometry, reflections that do not match the scene, objects that violate depth occlusion rules.

Sources: DDIY independent testing 2026; Undetectable.ai benchmark 2026; Keepnet Labs 2026; AU10TIX tool review 2026

Step 3: Perform a Reverse Image Search

Reverse image search is underused by non-journalists but consistently effective. Upload or paste the image URL into Google Reverse Image Search, TinEye, or Bing Visual Search. You are looking for two things:

1. The image appearing in a context that contradicts the claimed origin — a "personal photograph" that shows up on a stock image site, a news archive from five years ago, or a foreign language social media account is a definitive red flag.
2. Near-duplicate variants suggesting a generator was used — AI generation often produces multiple near-identical outputs from the same prompt. Finding several similar-but-not-identical images clustered together strongly suggests AI origin.

For professional fact-checking workflows, Bing Visual Search often surfaces more results than Google on synthetically generated imagery because Microsoft has been more aggressive about indexing AI art platforms. TinEye is better for tracking image repurposing on traditional web content.

Step 4: Apply Targeted Visual Forensics

Visual inspection should come last, not first — it is a supplementary signal, not a primary filter. But used in combination with detector results, these targeted forensic markers remain useful:

Hands and Fingers

Despite significant improvement in recent model versions, hand anatomy remains the most consistently imperfect element in AI images. Look closely at the full count of digits, the knuckle geometry, and whether fingernails are present and anatomically correct. Do this at 100–200% zoom — the errors are often invisible at normal viewing size. Midjourney V7 handles hands markedly better than V5, but asymmetric finger counts, fused joints, and impossible hand orientations still appear in a significant minority of outputs.

Text and Lettering Within the Image

Diffusion models generate text character-by-character as a visual texture, not as semantic content — they do not "know" what letters mean. The result is text that looks like text from a distance but dissolves into meaningless shapes at zoom. Check signs, labels, book covers, license plates, and name tags in any image. DALL-E 3 and GPT Image 1.5 have substantially improved text rendering, but Midjourney and Stable Diffusion variants still struggle.

Eye Reflections and Light Sources

In real photographs, reflections in eyes and glasses correspond to the actual light sources in the scene — windows, overhead lights, sky. AI models generate these reflections as visual texture that looks plausible locally but does not match the rest of the scene's lighting environment. Check whether the catchlight (reflection of the main light source) in each eye is symmetric and whether it matches the apparent light direction in the rest of the image. Mismatches are strong indicators.

Background-to-Foreground Transition

AI images generate foreground and background semi-independently, which creates characteristic edge artifacts at their boundary. Look for: background elements that "bleed through" foreground objects at edges, unnatural blur gradients that do not correspond to any real depth-of-field, background elements that violate perspective (objects in the distance that are implausibly large or small), and subject hair that does not separate cleanly from background. The hair boundary remains one of the most reliable visual tell-signs in 2026.

Where Detectors Consistently Fail: The Accuracy Gap

The accuracy figures in the comparison table above are best-case results from controlled test sets. Real-world performance is consistently lower. Three specific conditions break every detector currently available:

For a deep technical explanation of how detectors handle these failure modes, our guide on AI detection false positives covers the accuracy-versus-robustness tradeoff in detail.

Interpreting Confidence Scores: What the Numbers Actually Mean

Most detectors return a probability score: "87% AI-generated" or "31% AI probability." These numbers are frequently misread. The key principles:

• Scores above 85% AI — strong signal worth acting on, but verify with a second tool before high-stakes decisions
• Scores between 40–75% AI — genuinely inconclusive. The detector is uncertain. Do not treat these as confirmation of authenticity. Require additional investigation.
• Scores below 20% AI — probable authentic, but not conclusive. For critical decisions, apply EXIF check and reverse image search regardless of the score.
• Divergent scores across tools (e.g., 82% AI on one tool, 23% AI on another) — the image falls in the genuinely ambiguous zone. Document the divergence and apply additional forensic steps before making any determination.

The AI content authentication market — encompassing image, video, and text detection — was valued at approximately $1.8 billion in 2026 and is growing at a 42% CAGR (market research via OpenPR 2026), which means new tools are entering rapidly. The detection landscape in late 2026 will look meaningfully different from today. The workflow above — EXIF first, two detectors, reverse image search, targeted visual inspection — remains valid regardless of which specific tools dominate the market.

Industry Use Cases: Who Needs This and Why

HR and Recruitment Teams

A 2024 academic analysis of 15 million Twitter/X profile pictures identified approximately 7,723 confirmed AI-generated profiles used for disinformation and scam coordination (arXiv:2401.02627). On professional platforms including LinkedIn, AI-generated profile photos are a documented and growing problem for background-check workflows. KYC (Know Your Customer) platforms including AU10TIX added AI face-generation detection to their identity verification APIs in direct response. For HR teams without enterprise KYC access, running profile photos through the free detection workflow above takes under five minutes per candidate.

Publishers and Editorial Teams

The AFP integrated automated AI image detection into its fact-checking workflow through WeVerify. Reuters released formal AI detection guidance for its visual newsroom in 2024. The operational standard at major outlets: automated detection is a first-pass screening layer only. Borderline cases require senior photo editor review plus multiple independent tool checks. For regional and independent publishers without AFP-scale infrastructure, the free workflow described here approximates the same methodology.

Fraud Investigators and Legal Teams

Keepnet Labs documented deepfake fraud attempts surging 2,137% in three years, with romance fraud and business email compromise (BEC) as the primary attack vectors. For legal proceedings, automated detection results should be supported by EXIF documentation, reverse image search records, and where possible, cryptographic provenance data (C2PA manifests). A single tool's verdict is insufficient for evidentiary purposes — convergent multi-tool results with documented methodology are more defensible.

The C2PA Standard: Provenance at Creation Time

All the detection methods above are reactive — they analyze an image after the fact. The Coalition for Content Provenance and Authenticity (C2PA) takes a fundamentally different approach: embedding a cryptographically signed manifest at the point of image creation, recording what tools were used, whether AI was involved, and the editing history.

Over 5,000 organizations have joined Adobe's Content Authenticity Initiative. All five major camera manufacturers — Sony, Canon, Nikon, Fujifilm, and Leica — are C2PA members. For journalists using C2PA-enabled cameras (Sony Alpha series with C2PA firmware) and Adobe Lightroom, the provenance chain from shutter click to publication is cryptographically verifiable without any detection tool. C2PA-aware platforms display content credentials as a small "cr" badge. The limitation is that bad actors operating outside C2PA tools generate no provenance signal — reactive detection remains necessary for adversarial content.

For a comprehensive technical breakdown of both C2PA and the detection algorithms covered here, our AI generated image detector guide covers the full technical stack.

Frequently Asked Questions