AI Tools
Every AI Tools tool on Toolz.dev: 3 free utilities that run entirely in your browser. No signup, no uploads, no downloads.
Three tools sit here and none of them claims to detect AI, because nothing reliably does. What they do is measure and edit the specific, nameable signals that make text read as machine-generated: a vocabulary that reaches for inflated words, sentences of uniform length, hedging where a claim belongs, and in some pipelines, invisible Unicode characters that no reader can see.
The AI writing checker measures those signals and reports them without inventing a percentage. The AI text humanizer rewrites the ones that are mechanical to fix. The invisible character remover strips zero-width and control characters that survive a copy and paste, which is as often a formatting problem as anything to do with AI.
What each AI writing tool actually measures or changes
| Tool | What it does | What it cannot do |
|---|---|---|
| AI Writing Checker | Scores burstiness, overused words, hedging, em dash density | Tell you who or what wrote the text |
| AI Text Humanizer | Removes inflated words, filler and hedge phrases | Add the specifics only the writer knows |
| Invisible Character Remover | Strips zero-width, bidi and tag characters | Remove a statistical watermark in word choice |
Why AI detectors do not work
Detectors estimate how predictable text is to a language model, usually through perplexity and burstiness, and then convert that estimate into a confident-looking percentage. The estimate is real; the confidence is not. The same measurements flag writing by non-native English speakers, technical documentation, legal boilerplate and anything edited into a plain house style, because all of them are more predictable than casual prose. Stanford researchers demonstrated the non-native English bias in 2023, and OpenAI withdrew its own classifier the same year, citing low accuracy.
This matters most where the stakes are highest. An academic misconduct case built on a detector percentage is built on a number with no defensible error rate, and the false positives fall hardest on students writing in a second language. The writing checker is deliberately built the other way: it reports the individual signals it measured so you can judge them, rather than manufacturing a verdict it cannot support.
The tells that are real
Whatever detectors cannot prove, machine-generated prose does have recognisable habits, and they are worth editing out because they make writing worse, not because a detector might notice. A vocabulary of inflated abstractions - delve, leverage, tapestry, testament, landscape - where a plain word would do. Hedging attached to claims that need no hedge. Sentences of uniformly medium length, which is what burstiness measures: human writing varies, alternating long explanations with short flat statements. Symmetrical structure, where every paragraph is three sentences and every list has exactly three items.
The humanizer removes the mechanical half of that list, which is the half a tool can honestly fix. The other half is not automatable: text that says nothing specific reads as generated because it is empty, and the cure is a real example, a real number, a real objection considered. No editing pass adds those. Anything promising to make AI text undetectable is selling the detector fantasy from the other side.
Invisible characters, and what they are not
Text pasted from a chat interface, a PDF or a CMS regularly carries characters that render as nothing: zero-width spaces and joiners, variation selectors, bidirectional controls, and the Unicode tag block. Some arrive from formatting, some from an editor, and some are inserted deliberately as a marker. They break search, string comparison, diffs and word counts, and they are invisible in every editor that does not deliberately show them, which is what makes them so confusing to debug. The invisible character remover finds and strips them.
It is worth being precise about what that does and does not remove. A statistical watermark - the kind that biases a model token choice so a distribution can be recognised later - lives in the words themselves, not in hidden characters, and no character-stripping tool touches it. So this is a text-hygiene tool: it fixes strings that will not match, counts that are wrong and diffs that show a change where none is visible. Everything here runs client-side, so drafts are not uploaded to be analysed.
Frequently asked questions
- Are AI detectors accurate?
- No, not to a standard that should decide anything. They measure how predictable text is and convert that into a confident percentage with no defensible error rate. The same signals flag non-native English writers, technical documentation and heavily edited prose. OpenAI withdrew its own classifier in 2023 for low accuracy, and Stanford researchers documented the bias against non-native writers the same year.
- What makes writing read as AI-generated?
- A vocabulary of inflated abstractions such as delve, leverage and tapestry; hedging on claims that need none; sentences of uniform length, which is what burstiness measures; and symmetrical structure where every paragraph and list is the same shape. The deeper tell is emptiness - text that makes no specific claim - and no tool can fix that one.
- What are the invisible characters in text copied from an AI chat?
- Zero-width spaces and joiners, variation selectors, bidirectional controls and Unicode tag characters, all of which render as nothing. They arrive from formatting, editors and occasionally deliberate insertion, and they break string comparison, search, diffs and word counts while remaining invisible in most editors.
- Does removing invisible characters defeat AI watermarking?
- No. A statistical watermark biases a model choice of words, so it lives in the text itself rather than in hidden characters, and stripping invisible Unicode does not touch it. Treat this as text hygiene, which fixes strings that will not match and counts that are wrong, rather than as a way of concealing anything.
- Does Google penalise AI-generated content?
- Not for being AI-generated. Google guidance targets content produced primarily to manipulate rankings, whoever or whatever wrote it, and asks for evidence of first-hand experience and expertise. Generic content ranks badly because it is generic, which is the same reason it reads as machine-written.
- Is my text uploaded when I use these tools?
- No. All three run client-side, in your browser, so drafts, essays and client work stay on your device. That is not the norm in this category, where the usual arrangement is to post your document to a server to be scored, which means the document has been shared to obtain the score.