Direct answer
Detecting an AI photo means checking three layers in order: provenance data (C2PA / Content Credentials and EXIF), visible artifacts (hands, rendered text, reflections, lighting and shadow logic), and a detector model that scores the file itself. No single layer is conclusive — provenance is the strongest signal when present, artifacts are the weakest, and detector scores should be read as one indicator among several rather than as proof [1]. If you are verifying academic work rather than an image, note that Turnitin's own AI detection indicator is built for written prose and does not analyse photographs, so it answers a different question than an image detector does [1].
How Can You Tell If a Photo Was Created By AI?
Start with the file, not the picture. Original photographs carry metadata: camera make and model, lens, ISO, shutter speed, and often GPS coordinates. AI generators usually strip this or leave a suspiciously empty EXIF block, and provenance standards such as C2PA Content Credentials can record whether an image was captured, edited, or generated — a signed manifest is far more informative than any visual hunch [2]. If provenance data is present and says "generated," you have your answer without guessing.
Then look at the pixels. Current generators still slip on a predictable set of details: fingers that merge or multiply, jewellery that dissolves into skin, text that looks like writing but is not readable, reflections that do not match the subject, and shadows that fall in two directions at once [2]. These artifacts are real but they are also the fastest thing for model developers to fix, so their absence proves nothing — a clean image with no provenance data is ambiguous, not innocent.
Finally, work backwards through the web. A reverse image search can surface the earliest known version of a picture, and if the trail leads to a generator's gallery, a stock composite, or a different context than the one claimed, that history is stronger evidence than any artifact [2]. Combine the three layers and record what each one showed, because a documented chain — no EXIF, no provenance, three artifact clusters, reverse search returning a generator page — is defensible, while "it looks fake to me" is not.
Do AI Image Detectors Actually Work, And What Are Their Limits?
Partially, and the limits matter more than the headline accuracy figures. Detector performance drops sharply once an image has been resized, screenshotted, re-compressed, or run through a social platform's upload pipeline, because those steps destroy the low-level frequency traces many models rely on [3]. A detector that performs well on clean PNG output from one generator can fall close to chance on a WhatsApp-forwarded JPEG from another [3].
Accuracy also varies by generator and by training data. Models are typically trained on outputs from a known set of systems, so a newer or less common generator can produce images the detector has effectively never seen, and the resulting score is unreliable in both directions — false positives on real photos and false negatives on synthetic ones [3]. This is why serious practitioners never treat an image detector score as standalone evidence of misconduct or fabrication.
The practical rule is triangulation with disclosure. Report the detector you used, the version or date, the file as received, and the score, then place it alongside provenance findings and artifact observations rather than above them [3]. Where the stakes are high — academic integrity panels, editorial decisions, legal matters — the detector output should be described as an indicator that prompted further review, not as a verdict [3].
How Do You Verify An Assignment's Authenticity Before You Submit It?
The verification question changes completely when the artifact is a document rather than a photograph. Turnitin's AI detection indicator is trained on prose, reports a percentage of qualifying text flagged as AI-generated, and deliberately shows an asterisk (*%) instead of an exact figure when its confidence falls below its 20% threshold — those asterisk results are low-confidence signals, not clean bills of health [1]. Students who understand that banding read their reports far more accurately than students who only look for "0%".
The reason to check before submitting rather than after is timing. A pre-submission check surfaces AI indicators, similarity matches, and quotation or citation problems while you can still revise, instead of after a grade or an integrity referral has already been triggered [4]. Turnitin0 delivers a Turnitin AI detection report and a similarity report together in a single order, so the AI indicator and the plagiarism picture are read side by side rather than in isolation [4].
Students also care about what a check does to their record. A non-repository check compares the file without adding it to Turnitin's student paper database, and reports are not shared with third-party databases, so a draft check does not create the very match it is meant to detect [4]. Turnitin0 is an independent service and is not affiliated with Turnitin, LLC; it exists so students can see the same style of AI and similarity reports their institutions use before final submission, with 100,000+ reports delivered to 20,000+ students and a 4.9/5.0 satisfaction rating.
Image detectors will help you judge a photograph, but they will not tell you what an instructor sees when your written work goes through Turnitin. If you have used ChatGPT, Claude, or Gemini anywhere in a draft, the score that matters is the AI writing indicator on your prose — and turnitin0 lets you preview that report, alongside your similarity report, before you commit the file.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does Turnitin detect AI photos?
No. Turnitin's AI detection indicator is designed for written prose and reports the share of qualifying text it flags as AI-generated; it does not analyse photographs or images [1]. If your concern is an image, you need an image detector plus provenance checks, not a writing report.
What is the single most reliable sign that a photo is AI-generated?
Signed provenance data. C2PA Content Credentials and intact EXIF can record whether an image was captured, edited, or generated, which is far more decisive than visual artifacts [2]. Artifacts like malformed hands and unreadable text are useful clues but are being fixed quickly, so their absence proves nothing [2].
Why do AI image detectors give different answers on the same picture?
Resizing, screenshots, and platform re-compression strip the low-level traces many detectors depend on, and most models are trained on a limited set of generators [3]. A file that scores confidently in its original form can score near chance after a single round of social-media compression [3].
Can I check my written assignment before I submit it?
Yes. A pre-submission check surfaces AI indicators and similarity matches while you can still revise, rather than after an integrity referral [4]. Turnitin0 returns a Turnitin AI detection report and a similarity report together, and its non-repository check does not add your file to the student paper database [4].
What does an asterisk (*%) AI score mean?
It means Turnitin's confidence was below its 20% threshold, so it displays *% instead of an exact percentage — a low-confidence signal rather than a confirmed result [1]. Read the flagged passages and the rest of the report before drawing conclusions [1].