Is This Image AI-Generated? A Verification Workflow That Works

A polished image can be synthetic, but a strange looking image can also be genuine, compressed, edited, or photographed under unusual conditions. Visual inspection alone cannot settle the question. Modern verification works by testing a specific claim about an image: who published it, what it supposedly depicts, where and when the event occurred, and whether independent evidence supports that account. The useful outcome is not always a binary verdict of real or fake. A responsible result may be verified for this claim, contradicted by evidence, or still unverified. That distinction prevents a weak clue, such as an odd hand or missing metadata, from becoming a confident accusation. Step 1: write down the claim before examining the pixels Separate the file from the story attached to it. Record the exact claim, the account or page making it, the post URL, the publication time shown, and any named place, date, person, or event. Save the best available copy without repeatedly converting it. Screenshots are useful for preserving a post, but they can remove metadata and crop away context, so also preserve the original URL and downloadable file when lawful and safe. Decide what must be true for the caption to be accurate. An image presented as today's flood in one city creates testable questions about date, location, weather, landmarks, and earlier publication. An image presented merely as an illustration makes a different claim. AI involvement is only one possibility; an authentic photograph reused with a false caption can be equally misleading. Step 2: trace the earliest discoverable source Run more than one visual search when the stakes justify it. Google Lens and About this image may show visually similar versions, pages using the image, and occasions when Google first found a similar version. Google describes that date as an approximate discovery signal, not the capture date. Crop and search distinctive regions separately: a storefront, vehicle, sign, skyline, uniform, or unusual object may retrieve a source that a full image search misses. Sort results by what they establish, not by rank. Look for an earlier upload, a higher resolution file, an uncropped frame, a photographer credit, or an event page. The WITNESS verification methodology similarly recommends reverse searching thumbnails and visible details, sorting results chronologically, and checking whether weather, time of day, and independent reports fit the claim. The earliest result you can find is not automatically the original. Search indexes are incomplete, private posts are invisible, timestamps can reflect reposting, and a high resolution copy may still be a derivative. Treat source tracing as a chain: current post, earlier appearances, likely publisher, and, where possible, the person or organization that captured or created the image. Step 3: evaluate the publisher and obtain the source account Inspect the account outside the single viral post. Does it identify an accountable person or organization? Does its history fit the claimed access? Are there links to an official site, original gallery, assignment, or contact method? Search independently for the name rather than trusting profile badges, screenshots, or links supplied in replies. If the image matters for reporting or a consequential decision, contact the apparent creator through a separately verified channel. Ask for the original file, when and where it was made, what device or tool was used, what edits occurred, and whether they authorize the stated use. Answers are evidence to test, not automatic proof. Preserve the response and compare it with the file, public timeline, and other reporting. An anonymous source can still provide authentic material, and a familiar account can publish a mistake. Source reputation changes the weight of evidence; it does not replace verification. Step 4: test location, time, and physical context List visible anchors before interpreting them. Signs, road markings, transit equipment, architecture, terrain, utility poles, vegetation, shadows, weather, clothing, and emergency vehicles can support or contradict a claimed setting. Search each anchor independently and compare it with maps, official imagery, weather records, event schedules, or reporting from the alleged location. Do not overread visual anomalies. Compression can deform text and edges. Panorama stitching, portrait mode blur, rolling shutters, computational photography, reflections, motion, and ordinary editing can create features that resemble generation errors. Conversely, a synthetic image can avoid familiar errors. Pixel level clues should produce questions for further testing, not a standalone verdict. Check internal consistency at the scale available. Does a reflected object exist in the scene? Do shadows broadly agree? Is a sign's language plausible for the place? Could the claimed camera position exist? A contradiction needs an alternative explanation test. If compression, cropping, or perspective reasonably explains it, label the clue inconclusive. Step 5: inspect provenance and metadata without overclaiming Use a Content Credentials aware inspector when the original file or a supported URL is available. The Content Authenticity Initiative's Verify service can display Content Credentials and a recorded history when present. A valid credential can help establish that provenance information is cryptographically associated with the asset and has not been altered outside the recorded workflow. It may identify a signing tool or organization, actions, ingredients, or an AI related digital source type. That is a provenance signal, not proof that the depicted event occurred. The C2PA's harms guidance clearly says valid manifests do not make an asset true; an image with valid credentials can still carry misinformation. Read who or what signed the credential, what was actually asserted, whether validation produced warnings, and where the recorded history begins. Absence proves little. Content Credentials are opt in, adoption is incomplete, and copying, screenshots, or unsupported processing can separate a file from embedded data. Ordinary EXIF metadata can also be removed or edited. Missing provenance should be recorded as unavailable, not converted into evidence of AI generation. Step 6: corroborate the event, not just the image Search for the alleged event using names, landmarks, quoted text, and local language terms. Look for independent images from different positions, official notices, live streams, local reporting, or public records that establish the same place and time. Independence matters: ten posts that all copy one unverified source are one evidence chain, not ten confirmations. Compare details across sources. Matching smoke, vehicle positions, weather, damage, or crowd movement can strengthen a timeline when the sources are genuinely separate. Conflicts may reveal that an old image was relabeled, a real scene was composited, or the caption exaggerated what the frame shows. If the claim concerns a person, disaster, conflict, crime, election, health issue, or financial event, increase the threshold. Do not publish identity or intent from resemblance alone. Follow the relevant editorial, privacy, safety, and legal process. Step 7: record a bounded conclusion Write the conclusion so another reviewer can reproduce it. State the claim tested, files and URLs examined, tools used, strongest supporting and contradicting evidence, unresolved gaps, and decision time. Use calibrated language: Verified for the stated claim: multiple independent evidence layers support source, context, and integrity. Miscontextualized or contradicted: earlier use or scene evidence conflicts with the caption. Provenance verified, event not verified: the file history checks out, but the real world claim remains unsupported. Unverified: available evidence cannot justify publication as fact. Do not translate an AI detector score into certainty. Detection systems and platform labels can be additional leads, but the publishable finding should rest on traceable evidence. When evidence remains incomplete, withholding or clearly qualifying the image is a valid result. A compact pre publication checkpoint Before sharing, confirm that you have preserved the source, searched earlier uses and meaningful crops, evaluated the publisher, tested location and time, inspected available provenance, sought independent corroboration, and documented uncertainty. If one critical step cannot be completed, adjust the caption or do not publish the claim. The central question is broader than whether AI touched the pixels. Verification asks whether this particular file, from this source, supports this particular account of the world. Sources Learn more about an image https://support.google.com/websearch/answer/14177408?hl=en WITNESS Media Lab: Date methodology https://lab.witness.org/portfolio page/date methodology/ Content Credentials Verify https://verify.contentauthenticity.org/inspect C2PA Harms Modelling https://spec.c2pa.org/specifications/specifications/2.4/security/Harms Modelling.html Cover image credit Cover image by ViktorDFC , made available under CC BY SA 4.0 . WIKIVISE cropped and converted the source image.