Direct answer
There is no single visual tell that proves a photo is AI generated, so the reliable answer comes from stacking evidence: provenance metadata (C2PA / Content Credentials), reverse image search to find the earliest source, pixel-level inspection at high zoom, and a detector score treated as one signal rather than a verdict. Modern diffusion and GAN models have erased most of the obvious giveaways — hands, text, and background coherence — so any checklist based only on "look closely" will fail on current generators [1]. The practical method is to ask where the file came from before you ask what it looks like.
What are the reliable signs that a photo was AI generated?
The most durable signal is provenance, not appearance. Content Credentials built on the C2PA standard attach a cryptographically signed record of how an image was captured or edited, and that record survives far longer than any visual artifact [1]. When a file carries no provenance data at all, that absence is not proof of anything — most cameras and most social platforms still strip or never write it — but a valid, intact credential is strong positive evidence that the image was not synthetically generated.
Visual inspection still has value, but only as a secondary layer and only at full resolution. Physics that do not resolve — a reflection that contradicts the light source, a shadow with no caster, jewelry that merges into skin, background text that dissolves into pseudo-letters — remains worth checking, but each of these tells has been progressively engineered away by newer models [1]. Treat a single visual anomaly as a prompt to investigate, never as a conclusion, because compression and upscaling produce identical artifacts in genuine photographs.
Structural checks catch what the eye misses. Reading EXIF data tells you the capture device, lens, and timestamp when it survives; a plausible camera profile with consistent timestamps is meaningful, while an empty or contradictory header is merely inconclusive [1]. Reverse image search is often the fastest route to truth: if the earliest indexed instance predates the model that supposedly made it, or if the same face appears across dozens of unrelated identities, the question answers itself.
Finally, calibrate your confidence to the stakes. For a casual profile photo, two independent signals are usually enough; for an academic integrity decision or a newsroom verification, you want provenance plus source tracing plus a second reviewer, because every individual check has a non-trivial failure rate [2]. The mistake people make is not using weak signals — it is treating one weak signal as decisive.
Do AI image detectors actually work, and how accurate are they?
Detectors work well enough to be useful and poorly enough that they must never be the final word. Published accuracy figures are almost always measured on a fixed benchmark against a fixed generator, and performance drops sharply when the same detector meets a model it was not trained on [3]. A tool advertising 99% accuracy is quoting a laboratory number; the real-world number on a re-compressed image from a messaging app is materially lower.
Compression and resizing are the biggest practical enemies. Social platforms re-encode uploads, strip metadata, and downsample aggressively, and detection performance degrades under exactly those transformations [3]. This is why a detector may confidently flag the original file and return "uncertain" on the same image after it has passed through WhatsApp or Instagram — the signal was destroyed in transit, not disproven.
Human judgment degrades in parallel. Studies comparing human raters with automated systems find that people perform near chance on images from generators released after the ones they practiced on, and that confidence does not track accuracy — raters feel certain while being wrong [3]. The lesson is to trust a detector's flag more than your own gut, and to trust neither as proof.
Interpret any score through the tool's own confidence model. Turnitin, for example, reports an overall AI percentage alongside sentence-level flags, and deliberately displays *% instead of a precise figure when its confidence falls below its 20% threshold — an explicit admission that low-confidence detections should not be read as a number [2]. A detector output is an indicator that routes a human to review, not a certificate of authenticity [2].
How can students verify an image is legitimate before submitting it?
Start with the source, because an image you cannot trace is an image you cannot defend. If a figure, photograph, or screenshot is going into an assignment, record where it came from — the database, the repository, the original URL, and the access date — and confirm the page actually hosts that image rather than a repost [4]. When a marker asks about a figure, the citation trail is what settles the question, not a detector screenshot.
Use the same report view your instructor uses, so you see what they see. Turnitin's similarity report highlights matched sources and separates the overall index from the individual matches that produced it, and the index alone tells you very little without reading those matches [4]. Reviewing that output before submission turns an anxious guess into a specific, fixable list of passages and figures that need attribution or replacement [4].
Keep a verification record for anything you did not create yourself. A short note listing the tool used, the prompt, and the date is enough to demonstrate process if a figure is later questioned, and it converts an accusation into a conversation about method [4]. Where an image is genuinely AI-assisted — a generated diagram, a synthetic illustration — label it in the caption, which is almost always acceptable and almost never penalised.
Apply the same discipline to text that you apply to images. If an assignment mixes AI-drafted prose with your own analysis, the AI writing indicator will surface it with sentence-level flags, and understanding that report in advance is far cheaper than disputing it afterwards [2]. Students who preview their own reports consistently report fewer surprises at the marking stage, because the report is a map of what needs work rather than a verdict on their character.
Verifying a single image is manageable; verifying a whole submission — figures, captions, and the prose around them — is where a real report pays for itself. turnitin0 gives you the same Turnitin AI and similarity reports your institution sees, on your own draft, before the deadline rather than after it.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can I tell if a photo is AI generated just by looking at it?
Not reliably. Hands, text, and background coherence were once strong tells, but newer models have largely eliminated them, so visual inspection now works best as a first filter rather than a verdict [1].
Why does an AI detector give a different answer on the same image?
Because compression, resizing, and metadata stripping remove the statistical traces detectors rely on, and performance degrades under exactly those transformations [3]. The original file is always the better input.
What does *% mean on a Turnitin AI report?
It means the detector's confidence fell below its 20% threshold, so Turnitin withholds a precise percentage rather than presenting a low-confidence signal as a number [2].
Is a high AI score proof that a student cheated?
No. A detector result is an indicator that routes a submission to human review, and Turnitin's own documentation frames it that way rather than as a finding of misconduct [2].
What is the single most useful check for a suspect image?
Reverse image search. Finding the earliest indexed instance of an image usually answers the provenance question faster than any pixel-level inspection [1].