UnTXT reads any text for AI writing patterns and hands back the passages worth a second look, each one marked and carrying the reasoning that put it there. There is no score anywhere in it. Here is the whole process, from signing in to exporting the result.
Step-by-Step Process
Sign in at untxt.uk
Use a Google account, and there is no separate registration to sit through. A new account arrives carrying a free allocation that covers its first 1,000 words, which is enough to see what the thing does before you decide anything, and no payment information is asked for at any point in that.
Select your role
On the first sign in you choose Student, Teacher or Non academic professional. The interface, the default mode and the features on offer all shift to match the answer, so a marker and a first year are not looking at the same screen. You can change it later, whenever it stops fitting.
Paste or drop in your text
Paste an essay, a report, a dissertation chapter or any document into the main panel. You can also drop a file straight onto the blank canvas, since PDF, Word and plain text are all read the same way that pasted text is.
Run the analysis
Click Analyse, and the text goes to two frontier models that read it separately and then argue about it, with a third settling anything they cannot agree on. Nothing else runs first: there is no quick heuristic pass ahead of it whose answer you might see and mistake for the real one.
Read the marked passages
You are not left staring at a blank screen while that happens. Your own sentences drift past instead, each one tinted by whichever module picked it up, and when the argument finally settles the verdict lands and the marked-up text is sitting there underneath it. SUSPECT means both reviewers read that passage as machine-written. The bits they agreed sound like a person get marked too. That catches people out. They are, though, precisely the bits worth protecting once you start cutting, which is why we show them at all rather than showing you only the damage. There is no overall score anywhere on that screen, and there deliberately is not one. The passages are the point. Click one and a note says what we noticed about it, and that is the whole of it — nothing rewrites the sentence for you, and nothing offers to, because handing somebody a replacement is quietly writing their work for them.
Save or export
Everything you run is saved into Documents on its own, under a title taken from its opening words, alongside the date, the word count, how many passages were flagged and what the run cost you — and clicking any row opens that document straight back into the editor. Teacher and Professional accounts can pull a formatted A4 PDF out of the analysis view, or out of Documents later on when somebody asks for it.
Academic Weapon runs too
Academic Weapon now runs on every analysis, so that it no longer waits to be launched. Two frontier models take opposite sides and argue it out, with Claude Opus 5 defending one reading while GPT-5.6 Sol defends the other, because one model's opinion on its own is not enough to lean on here. They usually settle it inside two to four minutes, and sometimes considerably quicker than that, depending on how long the passage is and how far apart the two of them started. But not always, and to be clear that is rather the point of bothering to run two of them. When they cannot agree, a third model called Claude Sonnet 5 comes in and breaks the tie. Results land as highlights, gold where a passage reads AUTHENTIC and sky blue where it reads SUSPECT, and a reasoning panel shows what each model actually argued. It is the most confident signal that the system has to offer.
The 11 Detection Modules + Academic Weapon
Each module targets a distinct characteristic of AI-generated writing:
SPIKE
AI buzzwords, the overused phrasing that LLM output leans on. Replace them, or rephrase around them entirely.
SYNTAX
Repetitive sentence structures that AI models favour. Restructure them once you notice the pattern.
FORMAL
Hedging language that is more cautious than the point being made really warrants, and editing it down almost always helps the sentence.
HUMAN
Genuine writing markers, meaning your informal expressions and your personal voice. Keep every one of them.
TEMPLATE
Boilerplate phrasing lifted from AI templates, which reads badly enough on the page that it should either come out altogether or be rewritten from scratch in your own words.
FRONTED
Fronted adverbials that AI systems overuse. A major red flag, so rephrase.
FORENSIC
Style that sits inconsistently against the rest of the document, which is always worth investigating before you conclude anything about it.
RHETORIC
AI rhetorical structures, such as serves as, not just X but Y, rule of three triplets and heavy em dash use. Restructure them.
PUFF
Inflated language. Significance puffery, promotional vocabulary, vague attribution of the experts argue kind, and editorial modifiers ending in ing all count, and the repair in every case is to remove the phrase or replace it with something that says what you mean.
FORMULA
The rigid AI structural template, that challenges and then looking ahead arc so typical of machine written essays. Rewrite the conclusion around a genuine argument.
LEXICAL
AI characteristic discourse scaffolding. It shows as discourse marker density, uniform sentence length, and a type token ratio that drifts, so vary your lengths and swap the scaffolding out.
ADVERSARIAL ⚔
Multi model consensus by way of Academic Weapon. A highlight appears when Claude Opus 5 and GPT-5.6 Sol independently agree, or when Claude Sonnet 5 breaks a tie between the two of them, and because that is the highest confidence signal in the system it repays reading the per model reasoning in the side panel before you decide anything at all.
How Credits Work
Usage is counted in CVT, and one of them covers about 75 words. On a monthly plan an allowance lands each billing month, and anything past it is metered at the same rate as buying credit outright, so going over costs what it would have cost anyway. Credit packs are for organisations of three or more. Your first 1,000 words cost nothing.
Invite & Earn
Your account comes with a referral link, which lives in the account menu, and if somebody signs up through it and then runs their first check, the pair of you pick up bonus CVT for it. Nobody has to buy anything at either end, and it counts on the free tier.
Confidence Guard for Reviewers
If you are reading somebody else's work, Confidence Guard raises the bar a passage has to clear before it gets marked at all, because a false positive costs far more when you are marking a student's essay than when you are editing your own — so Teacher and Professional accounts see fewer marks and can lean harder on the ones they do get. Students keep the standard setting, which catches more, and that is the right trade when the work is your own.