Direct answer
There is no single check that proves an image was AI generated; the defensible answer comes from stacking three independent lines of evidence — provenance metadata and watermarks, physical/visual inconsistencies, and a detector result — and weighing them together. Provenance signals (C2PA Content Credentials, EXIF, invisible watermarks) are the strongest when present, but their absence proves nothing because metadata is routinely stripped [1]. Visual artifacts such as garbled text, impossible hands, and inconsistent lighting remain useful but are no longer decisive on their own [2]. Automated detectors add a probability, not a verdict, and their accuracy falls sharply once an image has been compressed, resized, or re-edited [3].
What are the reliable visual signs that an image was AI generated?
The highest-yield place to look is text and hands. Generators still produce lettering that looks plausible at a glance but breaks down on inspection — misshapen glyphs, inconsistent stroke weight, invented characters in signage or labels — and fingers that merge, multiply, or bend the wrong way [2]. Zoom to 200–400% and read every word in the frame; if a sign, badge, or book spine contains characters that are not in any alphabet, that is a strong signal.
Lighting and shadow physics are the second reliable cluster. Check whether highlights, shadows, and reflections all agree on one light source and one direction, and whether a shadow's softness matches the apparent distance and size of the object casting it [2]. AI images frequently show a face lit from the front while the neck shadow falls sideways, or a glossy surface that reflects a scene which is not actually in front of it.
Background coherence and texture are the third. Look for repeated or "melting" detail — identical window patterns, staircases that do not connect, foliage that dissolves into noise — and for skin that is uniformly smooth with no pores, fine hair, or fabric weave at full resolution [2]. These cues are strongest in older or cheaper generators and weaker in the newest models, which is exactly why they must be combined with provenance evidence rather than used alone.
Do AI image detector tools actually work, and how accurate are they?
They work as probabilistic classifiers, not as lie detectors. Detectors look for statistical fingerprints in pixel noise and frequency patterns, and they perform reasonably on untouched output from the generators they were trained on [3]. The problem is generalization: a detector tuned on one model family frequently misclassifies images from a newer or less common generator, so a "100% AI" reading is a confidence score about a pattern, not proof of origin.
Accuracy degrades further under normal real-world handling. Re-compression by social platforms, resizing, screenshotting, colour grading, and cropping all disturb the low-level signals detectors rely on, and reported accuracy can fall substantially on such images [3]. This is why a negative detector result is especially weak evidence: editing an AI image is trivial, and a human photograph that has been heavily processed can also trip a false positive.
The practical conclusion is to treat any single detector as one input among several. Run more than one tool when the stakes are high, record which tools and versions you used, and report the result as "consistent with AI generation" rather than "proven AI" [3]. Where a decision has real consequences — academic misconduct, publication, legal or employment matters — the detector output should support a human judgement, never replace it.
How can students make sure their own images and figures are not flagged as AI?
The strongest protection is evidence of authorship created while you work, not after. Keep original files with intact EXIF, layered source files, raw instrument output, screenshots of intermediate versions, and dated drafts, because these let you demonstrate the process behind a figure if it is ever questioned [4]. Students who build charts from their own spreadsheet data and export them directly have a clear, checkable trail.
Declare AI assistance explicitly wherever your institution permits it. Most academic-integrity policies now distinguish between prohibited generation and permitted assistance such as editing, formatting, or idea generation, and an honest declaration converts a potential violation into a documented, allowed use [4]. Check your module handbook for the specific wording your course requires rather than relying on general guidance.
Finally, understand what institutional tools actually check. Turnitin's AI writing indicator is designed for prose, and similarity reports compare text against source databases — neither produces an "AI image" verdict, so a clean report does not certify that a figure is human-made, and a flagged report is an advisory signal that a human then reviews [4]. If you want to see what a real report looks like on your own draft before you submit, a pre-submission check gives you that view.
Reading a report is one thing; seeing your own document rendered the way your department will see it is what actually removes the doubt. turnitin0 produces the same AI and similarity report format used in institutional systems, so you can inspect the score, the flagged passages, and the match highlights on your draft while there is still time to act on them.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can I tell from the file name or metadata alone?
No. A filename like IMG_4821.png says nothing, and EXIF is easily stripped by messaging apps or forged deliberately, so missing metadata is not evidence of AI generation [1]. Metadata is only useful when it is present and internally consistent.
Why does Turnitin show an asterisk instead of a percentage?
Turnitin displays *% when its AI detection falls below its 20% confidence threshold, meaning the signal is too weak to report as a number. That is a low-confidence result, not a clean bill of health.
Is a single "100% AI" detector score enough to accuse someone?
No. Detector accuracy drops on compressed or edited files and varies by generator, so a single score should be treated as one input in a human-reviewed judgement [3]. Documenting multiple tools and the original file strengthens any conclusion.
Do AI images always contain visible mistakes?
No. Current generators often produce images with no obvious artifacts, and the visual cues described above are least reliable on the newest models [2]. Provenance and detector evidence matter more as artifact quality improves.
Does a clean Turnitin report prove my figures are human-made?
No. Turnitin's AI writing indicator analyses text, and similarity reports compare text against source databases, so images are outside what those reports measure [4]. Keep your own process evidence for figures regardless.