AI Image Detector Checklist: 12 Checks Before You Trust a Result

Preserve the Best Available File

Start with the highest-quality, least altered version available. A screenshot, social-media download, collage, or heavily compressed copy can change the detector result and remove provenance information.

• Save the original file when available.
• Record the page URL, uploader, caption, and date.
• Note whether the file was cropped, filtered, resized, or screenshotted.
• Do not compare scores from different file versions as though they used identical evidence.

Read the Detector’s Score Definition

Providers do not necessarily use percentages in the same way. One score may represent a model confidence, another may represent distance from a threshold, and another may be a simplified user-facing label.

• Find the provider’s explanation of the output.
• Check whether the score applies to the full image or selected regions.
• Identify the threshold used for “AI” or “human.”
• Check whether the provider publishes known limitations.

Check Which Generators and Image Types the Detector Covers

A detector may perform well on generators included in its training data and poorly on unfamiliar or recently released systems. Performance can also differ across photographs, illustrations, screenshots, diagrams, and heavily edited graphics.

• Look for named generator coverage.
• Check whether the tool was designed for photographs, art, or both.
• Note whether the detector supports partially edited images.
• Avoid assuming “AI detector” means “detects every AI model.”

Review Privacy Before Uploading

A public detector may require you to upload the image to a third-party service. Do not submit confidential, customer, student, employee, medical, financial, identity, or unreleased business material without appropriate authorization.

Use the guide to what not to share with AI tools before uploading sensitive files. Review the detector’s current retention, training, deletion, and account policies separately.

• Remove unnecessary private information.
• Use a redacted copy when the investigation permits it.
• Do not assume a free detector keeps uploads private.
• Avoid uploading evidence that must preserve a formal chain of custody.

Run the First Detector on the Original File

Use the cleanest available file and record the exact output instead of summarizing it from memory.

• Record the detector name and URL.
• Record the date and time of the check.
• Save the exact percentage, label, and explanatory wording.
• Note the file type, dimensions, and whether the tool resized it.

Test a Second Detector Without Averaging the Scores

A second detector can reveal whether the first result is stable across different models. It does not create certainty by itself.

• Use the same file version in both tools.
• Record each tool’s score definition.
• Do not calculate an average from unrelated percentages.
• Do not keep testing until one tool confirms your preferred conclusion.

Check Provider-Specific Provenance Separately

Provenance tools answer a different question from statistical detectors. They look for supported origin signals rather than attempting to classify every image.

OpenAI’s verification tool checks uploaded PNG, JPG, and WebP images for supported C2PA metadata and SynthID watermarks associated with images generated using OpenAI tools. A detected signal supports likely OpenAI origin; no signal does not prove that the image is authentic or non-AI.

Gemini can check SynthID for supported Google AI content and inspect compatible Content Credentials. Its SynthID result is provider-specific: no Google signal does not rule out another AI system.

• Record the provider and exact signal detected.
• Keep “no supported signal” separate from “likely human-created.”
• Review whether metadata may have been removed.
• Do not use provenance to validate the truth of a caption.

Test Whether File Changes Affect the Result

Cropping, compression, screenshots, sharpening, filters, upscaling, and format conversion can change the patterns a detector analyzes.

• Compare the original file with a screenshot only when both are available.
• Record each version as a separate test.
• Do not treat score changes as proof that one version is “more fake.”
• Consider that only one region may have been AI-edited.

Evaluate False-Positive Risk

A false positive occurs when a detector labels an authentic or primarily camera-created image as AI-generated.

Risk may increase when the image contains:

• heavy retouching, denoising, sharpening, or HDR processing;
• unusual lighting, blur, reflections, or compression;
• synthetic-looking products, architecture, or patterns;
• screenshots, game graphics, rendered scenes, or digital art;
• very small dimensions or limited visual detail.

Evaluate False-Negative Risk

A false negative occurs when a detector labels an AI-generated or AI-edited image as likely authentic.

Risk may increase when:

• the generator was not represented in the detector’s training data;
• the image was cropped, compressed, screenshotted, or edited;
• only a small part of a real image was generated;
• the image contains simple or low-detail content;
• the generator was designed to reduce detectable artifacts.

Compare the Score With Non-Detector Evidence

A detector cannot verify the date, location, caption, seller, event, or claim attached to the image. Compare the score with source evidence instead of allowing it to replace an investigation.

Use the AI fact-checking guide to verify names, dates, quotations, product claims, and other factual details through reliable original sources.

• Check the earliest credible source.
• Compare independent reporting or official records.
• Review reverse-search findings and available provenance.
• Ask whether the image is necessary to prove the claim.

Record a Limited, Evidence-Based Verdict

Choose the narrowest conclusion that the combined evidence supports. Do not convert a detector result into a claim about intent, fraud, or authorship without independent evidence.