You paste AI-generated text into a detector, get a percentage, and still cannot tell which sentence made it look machine-written. Detector scores tell you AI, but they rarely show the parts that triggered the flag, because most of the signal hides in characters you cannot see.
AI systems like ChatGPT watermark output with zero-width spaces and Unicode joiners. Those invisible markers survive rewrites and copy-paste, so an edited draft can keep flagging as AI long after you think you fixed it.
Within a few minutes you can inspect a piece of AI-generated text and see what the detector was reacting to, then strip the leftover markers with AI Watermark Remover.
Which tool shows what is hidden in AI-generated text?
The most direct way to see what was removed from AI-generated text is to run it through AI Watermark Remover, a free tool that detects the invisible zero-width spaces, Unicode joiners, and hidden marks that ChatGPT, Claude, Gemini, and Grok embed. Revealing them shows exactly which invisible elements a remover strips.
If you only need to know which sentences read as machine-written, free detectors like GPTZero or QuillBot highlight flagged passages instead, but they cannot reveal those hidden characters.
What you need before you start checking
To see what was removed from AI-generated text, you need the original output, not a retyped copy. Invisible characters like zero-width spaces survive copy-paste, but a word processor that normalizes fonts erases them before you check. If the text has been reformatted, the advice changes: compare the output against your original prompt instead.
What you'll need ready before step 1:
- The untouched AI output. Copy it straight from ChatGPT, Claude, Gemini, or Grok into a plain text file before editing.
- A current browser. Detection runs in your browser, so use a recent Chrome, Firefox, Safari, or Edge rather than an old build that hides zero-width characters.
- Selectable text. A screenshot or scanned PDF cannot be checked for invisible characters.
- Your original prompt. When no hidden marks appear, comparing output to prompt shows whether the model rewrote the text or the marks were lost in transit.
- A free AI detector as a second opinion. Grammarly's detector flags the sentences that read as machine-written when invisible marks are gone.
- A plain text editor tab. Word processors silently replace special characters; a plain editor preserves them exactly.
3. Compare the visual patterns
Once you flag a candidate passage, zoom out and look at the whole page. AI-generated text often fails to vary sentence length naturally. Human writing mixes short punchy sentences with longer clauses. Machine output tends to clump: a string of identically-structured sentences, each running 12-18 words, with the same subject-verb-object pattern repeating.
Run a quick scan on paragraph length, too. A human writer will let paragraphs breathe: some run seven lines, some are a single sentence. AI output frequently produces paragraphs within a narrow band of three to five lines, all roughly the same visual weight on the page.
4. Check for lexical tics
I keep a mental list of words that flag AI writing. These are terms models lean on because they sit in the middle of the probability distribution for many contexts: "delve," "navigate," "landscape," "leverage," "robust," "testament," "underscores," "foster," "arguably," "ultimately." A single instance means nothing. Three or more in a 500-word passage is a strong signal.
Human writing has signature vocabulary: the writer's pet words, their favorite transition, their verbal tics. AI writing has no signature. It uses the safest possible word in every position. If you read a paragraph and every word feels generically correct with no personality, that is evidence.
5. Watch for the bland conclusion
AI models love summary. After making a point, the model almost always appends a sentence that restates what it just said in slightly different terms: "This shows that…" or "In summary…" or "These factors together illustrate why…"
Humans, by contrast, tend to end a point and move on without the signpost.
So when you reach the last sentence of a paragraph, check if it is telling you something you already read. If yes, the passage likely came from a model. Cut those paragraphs out and judge the remaining text on its own merit.
6. Test the logical depth
This is the hardest trick to automate, but easy for a human reader. AI writing is shallow in a specific way: it lists supporting reasons rather than building a chain of reasoning. A human writer might say "Strategy X failed because the feedback loop was too slow, which meant engineers shipped changes faster than they could validate them." That is one causal link with a mechanism attached.
An AI will write "Strategy X failed for three reasons: it lacked sufficient testing, communication was poor, and timelines were unrealistic." Every bullet is a category, not a mechanism. The text names the problem but never explains how it actually happened.
When you see a paragraph that names multiple factors without connecting them causally, flag it. That is surface-level analysis, which is what language models produce by default.
7. Disambiguate with a context-specific prompt
Sometimes the text is genuinely mixed: human-written sections with AI-generated patches, or an AI draft that was lightly edited. For those cases, I use a targeted test. I pick a narrow technical detail and ask a question the text should answer if it came from real experience.
For example, if the text claims "running this query in PostgreSQL requires an index on both columns," I stop and ask: Which index type? A B-tree on the ordering column and a hash on the equality column? Or a single composite B-tree with the right column order? The real practitioner either knows or describes enough context for someone else to know.
The AI cannot produce the wrong-specificity answer: it will either be generically correct ("you should add an index") or detailed in a way that matches exactly what the documentation says, never the nuanced "it depends" that comes from having actually hit the edge case.
I call this the specificity gap. If you read a passage and think "this sounds like someone who read the docs but never ran the query," you are probably right. Trust that instinct.
*Next step in the workflow: after you identify AI-generated sections, the real question is whether that matters for your use case. A student's essay? High stakes. A draft technical memo that was machine-written and reviewed by a human afterward?
Probably fine. The detection is only as valuable as your decision about what to do with the finding.*
8. Check for circular definitions
AI text has a habit of explaining a term by restating it in slightly different words. A human will give you an example or a concrete property. An AI writes "agile methodology is a flexible approach to software development that emphasizes adaptability." That tells you nothing.
The human version: "agile means you ship working software every two weeks and adjust the backlog based on what you learn."
When you read a sentence and feel like you already knew everything it said before you started the sentence, that is the model filling space. Circle those sentences. The text around them is also suspect.
9. Look for the missing second word
Human writers reach for the precise second word in a pair. "Beer and pretzels" instead of "beer and snacks." "Budget and timeline" instead of "budget and resources." The specific word carries texture.
AI writing defaults to the generic second term because it lowers prediction risk.
Run a find for "and" then check the word that follows it. If you see "and more," "and other factors," "and related topics," you are reading placeholder language. A real writer would have thought of the actual second item or left the list at one item.
The AI-generated text inspection checklist
Before you finalize any cleaned AI text, run through this quick verification table. Each task is something you can confirm in under a minute.
| Task | Why it matters | Done |
|---|---|---|
| Paste the raw AI output into a text editor that shows invisible characters (like Notepad++ or VS Code with "Render Whitespace" on) | Zero-width spaces and joiners are invisible by design; you cannot see them in a normal browser or word processor | |
| Compare the character count before and after running a watermark removal tool | A drop of 5-30 characters (depending on text length) confirms hidden Unicode marks were actually stripped | |
| Copy the cleaned text into an AI detector like GPTZero or Copyleaks and check the "AI probability" score | If the score drops significantly, the invisible watermark was influencing detection; if it stays high, the text still reads as AI-generated | |
| Search the raw text for the Unicode characters `\u200B` (zero-width space), `\u200C` (zero-width non-joiner), and `\u200D` (zero-width joiner) using a find tool | These are the most common watermarking characters; finding them proves the watermark exists and removal was needed | |
| Re-run the cleaned text through the same AI tool that generated it (e.g., ChatGPT) and ask it to "check if this text contains any invisible watermarking characters" | The AI itself can often confirm whether the hidden marks are gone, giving you a second opinion | |
| Read the cleaned text aloud or have someone else read it | Watermark removal should not change the visible words; if you spot missing punctuation or garbled phrases, the tool may have removed visible characters by mistake | |
| Search for the Unicode characters `\u200E` (left-to-right mark) and `\u200F` (right-to-left mark) in the cleaned text | If your text mixes English with Arabic, Hebrew, or Urdu, a cleaning tool that removes every zero-width character will also delete legitimate bidi marks, causing the cleaned text to render in the wrong reading order | |
| Search for Cyrillic lookalikes such as `а` (U+0430, Cyrillic small a) when your document should be pure English | Some watermarking methods encode bits by swapping visually identical letters, for example Cyrillic `а` for Latin `a`; the words look normal, but a detector that reads Unicode values will see the difference | |
| Open the cleaned file in a hex editor such as HxD on Windows or Hex Fiend on Mac and confirm the byte order mark is unchanged | Some cleaning tools rewrite the file as plain UTF-8 and drop the byte order mark, which breaks strict XML parsers and older publishing pipelines that expect a UTF-16 byte order mark | |
| Search for variation selectors `\uFE0E` (text style) and `\uFE0F` (emoji style) | Newer watermarking schemes hide data in these selectors, which tell a device whether a symbol renders as text or as an emoji; if any remain after cleaning, the watermark is not fully removed | |
| Run the cleaned text through a different detector, for example Winston AI or Writer, and compare its score with the first detector | Detectors rarely agree on borderline text; two independent low scores are stronger evidence than one, and a high second score means the visible wording, not the watermark, is still the problem |
One final check: if you are working with a document that was copy-pasted from a web interface, also strip formatting (Ctrl+Shift+V in most editors) before inspecting. Some platforms embed hidden characters in the clipboard data that are not part of the AI watermark at all. For a deeper walkthrough of how invisible watermarks behave in real documents, see our case study on text watermark removal.
The flagged sentence is often not the real problem
When a detector flags a sentence in AI-generated text that reads perfectly naturally, the usual advice is to rewrite it. That advice only holds when the detector is reacting to your word choice. With text that came straight from ChatGPT, Claude, or Gemini, the highlight can be caused by invisible characters sitting between the visible letters: zero-width spaces and Unicode joiners that carry the model's watermark.
Those marks shift the statistical pattern the detector measures, and rewriting the visible words leaves them in place, so the same spot keeps getting flagged.
Tests that compared 30 detection tools found that highlights and accuracy scores vary between tools, so a sentence flagged by one detector may pass another. The rule of thumb is simple: if a clean, natural sentence stays highlighted after you reword it, stop editing and check for hidden characters first. That is the case where stripping the watermark, not rewriting the prose, is the actual fix.
Frequently asked questions about AI-generated text
How can I tell if text was written by AI?
Look for a combination of signals rather than any single one. AI text tends to use uniform sentence lengths, predictable transitions, and filler phrases like "delve" or "in conclusion." It also avoids typos and personal anecdotes. For a faster check, paste the text into a detection tool like GPTZero or QuillBot's AI detector, which highlight the specific sentences they flag as machine-written.
What are invisible watermarks in AI text?
Invisible watermarks are hidden Unicode characters, such as zero-width spaces and joiners, that some AI systems embed in the text they generate. You cannot see them in a normal editor, but they survive copy-paste and can mark the text as AI-produced. Tools like AI Watermark Remover scan for these characters and strip them out while leaving your visible words untouched.
Why would I want to see what was removed from AI text?
Knowing what was removed tells you whether a detector flagged your text because of hidden characters or because of the writing style itself. If invisible watermarks were the cause, removing them may change the result. If the style is the issue, you need to rewrite the content instead. That distinction saves you from editing the wrong thing.
Do AI detectors rely on invisible watermarks?
No, most detectors do not. Tools like Copyleaks and Grammarly's AI detector analyze word choice, sentence structure, and predictability using language models. They do not look for hidden characters. That means removing watermarks will not fool a detector that flags your text for stylistic reasons.
The two problems are separate and need separate fixes.
Can I remove AI watermarks without changing my text?
Yes, if the watermark is made of invisible Unicode characters. A tool that targets those characters, such as AI Watermark Remover, removes them without altering a single visible word. Your sentences, formatting, and meaning stay exactly the same. The catch is that this only helps with hidden characters, not with the detectable writing patterns that most AI detectors actually use.
Conclusion: confirm the text is clean before you publish
Run the final draft through AI Watermark Remover one more time and confirm the output carries no zero-width spaces, joiners, or hidden Unicode marks. That closes the loop this article opened: detect the invisible marks, strip them, verify they are gone, then publish.
Choose differently only when your problem is style, not watermarks. A detector with sentence-level highlights, like GPTZero or Scribbr, will show which sentences read as AI-generated, but that points to a rewriting task, not a removal one.
Next, read our guide to getting more out of seeing what was removed for repeat runs and routine checks. Processing happens on your device, so nothing you paste leaves the browser; the privacy policy spells out the details.
