How to check your essay before you submit it

You used AI somewhere in your process, for an outline or a rough draft or to unstick a paragraph at 2am, and now you have read that assistants embed hidden watermarks and you are not sure what is in your file.

Here is what to actually check, in order, with what each result means. It takes about five minutes.

Step 1: Find out what is literally in the file

This is the only part that produces a definite answer, so start here.

Paste your document into a Unicode checker and look at what comes back. You are looking for:

• Zero-width characters (U+200B, U+200C, U+200D, U+2060, U+FEFF). Invisible, and no reason to be in an essay.
• Tag characters (U+E0000–U+E007F). If you have these, someone encoded a hidden message. Rare, but unambiguous.
• Private-use codepoints. No standard meaning, so no innocent explanation.
• Odd spaces (U+00A0, U+202F, U+2009). Extremely common and mostly meaningless. More on this below.

You can run this check for free, or if you would rather not paste your work anywhere, open the file in a code editor like VS Code with "render whitespace" and "highlight non-ASCII characters" enabled.

What the result means. If it comes back clean, your document contains no invisible carriers, full stop. That is a real fact about your file, verifiable in a hex editor. If it finds something, you now know exactly what and where.

Step 2: Do not panic about odd spaces

The single most common finding is U+202F, the narrow no-break space, followed by U+00A0, the ordinary non-breaking space.

These are almost never a watermark. Web pages use them for typographic control. PDFs produce them. Word produces them. Google Docs produces them. If you copied a quotation from a journal article, you probably brought some along.

They are worth normalising because they can break find-and-replace and cause odd line wrapping. They are not evidence of anything, and a tool that presents them as alarming is manipulating you. Any checker worth using distinguishes between "this character has no legitimate reason to be here" and "this character is everywhere for boring reasons".

Step 3: Check any images separately

If your submission includes generated figures, diagrams or illustrations, the image files are a bigger exposure than the text.

A raw Stable Diffusion PNG contains your entire prompt in a text chunk: prompt, negative prompt, seed, model, sampler. Images from most commercial generators carry a signed C2PA manifest naming the tool. Anything that touched a phone carries EXIF with a timestamp and possibly GPS coordinates.

None of this is visible when you look at the picture. All of it travels with the file, and anyone can read it with free software.

Metadata removal is lossless. It drops the metadata blocks without re-encoding the picture, so there is no quality loss. Worth doing regardless of AI, honestly, just for the GPS.

Step 4: Read your own prose for the obvious tells

This is the part that actually affects how a marker reads your work, and no tool fixes it for you.

Em-dashes. The most cited giveaway of 2025–2026. Models produce them constantly; almost nobody types them, because they are not on a keyboard. If your essay has one every other paragraph and you have never deliberately typed one in your life, that is visible.

Stock phrases. "It's important to note that." "In today's fast-paced world." "Delve into." "A testament to." "Navigating the complexities of." "Ever-evolving landscape." These read as filler to any experienced marker.

Even rhythm. Human drafts vary wildly, with a four-word sentence next to a forty-word one. Model output tends toward consistent mid-length sentences. Read a page aloud; if it is metronomic, it will read as generated even if you wrote it.

Bulleted lists with bolded lead-ins. The "Term — explanation" pattern, repeated. Very recognisable.

Conclusions that restate the introduction. Models default to it. Markers notice.

Step 5: Understand what your university actually runs

This is where most of the anxiety online is misdirected.

Turnitin's AI detection and similar tools work on writing style: sentence structure, word predictability, burstiness. They are not scanning for zero-width characters. So removing invisible characters will most likely not change their verdict at all.

That cuts both ways, and the second direction matters more than people realise. These detectors have a documented false-positive problem, particularly for non-native English speakers and for anyone who writes in a plain, structured style. Several universities have restricted or dropped them as evidence for exactly that reason.

If you are worried about being flagged for work you genuinely wrote, the useful defences are your drafts, your version history, your ability to explain your argument. Not character cleaning.

What this checking does and does not get you

It gets you: a factual inventory of your own document, no exposed prompts in your images, no stray characters causing formatting weirdness, and a clear read on which surface patterns a marker would notice.

It does not get you: a guarantee about any detector's verdict, and it does not touch statistical watermarks like SynthID, which live in word choice and cannot be removed by cleaning a file.

Anyone promising the first thing is overselling. Knowing what is in your own work is a reasonable thing to want, and it is achievable. Everything beyond that gets speculative fast.

The five-minute version

1. Run the text through a checker. Note anything that is not an odd space.
2. Ignore the odd spaces unless they are breaking your formatting.
3. Strip metadata from every image you are submitting.
4. Read a page aloud and listen for the em-dashes and the metronome.
5. Keep your drafts.

That is genuinely all of it.