Direct answer
An online AI detector cannot reliably tell you whether an image was AI-generated by looking at the picture alone. Image classifiers that claim to do this are probabilistic, and they lose accuracy the moment a file is compressed, resized, or passed through a screenshot [1]. More importantly for academic work, the detector your university runs — Turnitin — does not analyse images at all: it evaluates the text layer of a submitted document [1]. So if your real question is "will my essay be flagged?", an image-based checker is the wrong instrument, and a text-based pre-submission check is the right one [1].
Can an AI detector online actually analyse an image and tell you if it was AI-generated?
Partly, and far less reliably than the marketing suggests. Tools that analyse an image fall into two families: classifiers that look for statistical artefacts left behind by generators, and provenance systems that read embedded metadata or watermarks [2]. Classifiers are the ones most often advertised as "AI image detectors," and they are the fragile half of the pair. They are trained on specific generator outputs, so they degrade against newer models and against anything that has been re-encoded [2].
The single biggest practical problem is that screenshots strip the evidence. When an image is screenshotted, resized, cropped, or re-uploaded, the metadata that provenance systems depend on is usually removed, and the compression artefacts that classifiers look for are overwritten [2]. That is why the same picture can come back "AI-generated" from one site and "human" from another within minutes. Detection of synthetic imagery is framed by its own developers as an ongoing adversarial problem rather than a solved binary check [2].
There is also a category error worth naming. A screenshot of a paragraph is an image of text, not an image in the photographic sense. An image classifier asked to judge it is being asked the wrong question — it will report on the pixels, not on whether the sentences were drafted by a language model [2]. If what you actually want to know is whether the writing is AI-assisted, you need a text-based detector, not an image one.
Does Turnitin detect AI in images, screenshots, or scanned documents?
No. The AI Writing Report highlights qualifying text segments and returns an overall percentage, and it is only generated for supported text-based file types [3]. Image-only files, and scanned pages that carry no text layer, fall outside what the report can process, so no AI percentage is produced at all [3]. This is a scope decision, not an oversight: the detector is built to read prose and judge how it was composed.
That has a direct consequence for anyone planning to submit a screenshot. If you upload a picture of your essay, you are not hiding anything — you are simply submitting something the report cannot read, which in most institutional workflows means the submission is rejected or returned rather than scored [3]. Where a text layer does exist and the detected amount is small, the report shows an asterisk-style low-confidence indicator instead of a precise figure rather than implying a confident verdict [3]. Report data is retained in the institutional account and is visible to instructors within the assignment, so a "clean" screenshot proves nothing about the underlying document [3].
The practical takeaway is that Turnitin answers a text question. It will tell an instructor which sentences in a .docx read as machine-composed; it will not tell them whether a JPEG was made by Midjourney, and it will not score a photo of a page [3]. Students who expect an image upload to produce an AI percentage are usually surprised by this, and the surprise tends to arrive at the worst possible moment — after the deadline.
How can you check your own draft for AI and similarity before you submit it?
The reliable route is to review the actual text file, before the deadline, using the same kind of report your instructor will see. Turnitin's own student guidance recommends reviewing similarity and AI indicators before final submission so that flagged passages can be revised rather than discovered after the fact [4]. That draft-level review is presented as a learning step: seeing which sentences were flagged teaches you how your writing reads to a detector, which is far more useful than a pass/fail verdict [4].
The obstacle is access. Institutional accounts frequently restrict self-check features, and many universities disable student previews entirely, which is exactly why students go looking for independent pre-submission checks [4]. When you do find one, submit the source .docx or .pdf — the file with a real text layer — rather than a screenshot or an exported image, because the text layer is the only thing that can be scored [3][4].
Keep the result in proportion. A low AI percentage is not a guarantee of acceptance, and a high one is not a misconduct finding; instructor judgement and the assignment's own rules still decide the outcome [4]. What a preview buys you is time — the chance to rewrite a flagged paragraph on your own terms instead of arguing about it afterwards [4].
If you would rather see the real thing than guess from a screenshot, turnitin0 runs your actual document through a Turnitin check and hands back the AI and similarity reports your instructor would see — the cover, the score, the flagged passages, and the match summary, all in one delivery.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Can I upload a screenshot of my essay to an AI detector?
Not usefully for Turnitin-based checks. The AI Writing Report is generated only for supported text-based file types, and image-only files produce no AI percentage [3]. Submit the original .docx, .pdf, or .txt instead.
Are "AI image detector" websites accurate?
They are probabilistic and inconsistent. Classifiers degrade against newer generators and against re-encoded files, and screenshots usually strip the metadata that more durable provenance checks rely on [2]. Treat any single verdict as a weak signal.
Does Turnitin show an exact AI percentage?
Sometimes it shows an asterisk-style low-confidence indicator instead of a number, when the detected amount falls below the confidence threshold [3]. That marker means the signal is weak, not that the text is clean.
Why does my AI score change between checks?
Detection is a signal rather than proof, and results depend on the exact text submitted, the file's text layer, and the report's confidence threshold [1][3]. Editing the flagged passages — not re-uploading a different file format — is what moves the score.
Can I check my draft before the deadline?
Yes, and Turnitin's own student guidance recommends reviewing similarity and AI indicators before final submission so you can revise rather than discover problems afterwards [4]. Many institutional accounts restrict self-check access, which is why independent previews exist [4].