Paste the Llama draft into the box above and press Humanize. It rewrites every sentence in one pass, up to 12,000 characters or roughly 2,000 words per run, at one credit per fifty words. Run the checker on what comes back for a 0 to 100 score and one of three verdicts: Reads as AI-generated, Borderline, or Reads as human-written. A check costs one credit whatever the length and needs at least 65 words. Ten credits a day are free per network, with no account, no login, no card and no email address to verify. Llama drafts also carry something hosted chat output does not: because the weights are open, you chose how the text was generated, and that decision is part of why it reads the way it does. This page covers the edit and the setting. This site is not affiliated with Meta.
Look at the top and bottom of what you pasted before you think about phrasing. Text copied out of a local runner, a notebook, or a raw completion endpoint often carries the chat format with it. Llama 2 builds wrap turns in [INST] and [/INST], with a system message sitting between <<SYS>> and <</SYS>>. Llama 3 builds use <|begin_of_text|>, <|start_header_id|>assistant<|end_header_id|> and <|eot_id|>. When a front end does not strip those, or when generation stops somewhere odd, they arrive in your document as literal characters.
The quieter version is a bare word assistant on the first line, or an end token printed once at the bottom because the runner did not treat it as a stop. Markdown travels the same way. Asterisks where the bold was, hash marks on headings, and a section rule that becomes three hyphens the moment it lands in a plain text box.
Clear all of it first. The rewrite is priced by word and template debris is words, so cutting it is free money. It also has a habit of surviving into a finished document, which is a worse outcome than any score.
Under the template debris is a set of habits that repeat across almost every Llama draft, whatever the size of the build or who fine-tuned it. Each one takes a single edit.
The writing tells page runs that check in your browser, marks each phrase where it sits, and gives you the reason in one line. It is free and it has no daily limit. It matches on phrasing rather than on where the text came from, so it fires the same way on a small build running on a laptop as it does on anything hosted.
Here is the part only open weights give you. When you run the model yourself you set the decoding: temperature, top-p, top-k, and whatever repetition penalty your runner exposes. Those numbers decide how far the model may stray from the single most likely next word.
Near-deterministic generation, which is what a temperature at or close to zero produces, gives you the most predictable continuation available. Predictability is what a detection score reads. A greedily generated draft is by construction the flattest version of itself: the most ordinary phrasing at every choice, and sentence lengths that settle into a narrow band and stay there for pages. If you configured your runner for code and never changed it back, check that before you blame the model.
Turning the temperature up is not a fix by itself. It buys variety in wording, not specifics, and a looser sampler drifts further from the facts, so names, numbers and citations need checking against a source either way. What it does buy is a draft with more in it to cut. Generating at a normal setting and trimming is easier work than generating at near-zero and trying to put life back in.
Regeneration is also free on your own hardware, which changes the arithmetic. Run the same prompt twice at different settings, check both, and keep the better one. A check costs one credit at any length, so comparing two full drafts costs two credits out of the ten.
Open weights get fine-tuned and republished constantly, under names with no connection to Meta. Much of the instruction data used for that tuning was itself generated by a larger chat assistant, which means the house phrasing of the big hosted models is copied down into small local ones. That is why a small build on a laptop reaches for the same three-item lists and the same negation pivot: not just a tool, but a shift in how the work gets done.
Two things follow. Switching to a different fine-tune does not get you out of the tells, because they came from the training data rather than from the base model, so the fix is editing and not shopping. And if your text came out of a writing product that never says what it runs underneath, this checklist still applies. Nothing in ordinary prose identifies a model, which is also why no score can tell you which one produced a passage.
The examples below were written for this page rather than copied out of a session. The move is the same every time. Cut the chat furniture, then put the specific thing you know where the abstract claim was.
Look at what the rewrites cost. Each one commits to something the original stepped around, and each one is shorter. That is the whole trade, and it is why the second version sounds like a person with an opinion wrote it.
Order matters more than effort. Strip, check, read the findings, fix by hand what is quick, then rewrite the rest. Running the humanizer blind works, but you learn nothing about the habits you will produce again next week.
For anything past 2,000 words, split at your own section headings and run the parts in order, so each run sees a complete argument. The word counter gives you the split points. It is free and unlimited, and so are the proofreader, the readability checker, the AI vocabulary checker, the plagiarism and originality check and the signed drafting record. The writing tells page runs in your browser as well. None of them touch your credits.
One line on coursework and then back to the writing. Rewriting a draft you were permitted to generate is ordinary editing. Where AI is not permitted, that rule covers the draft however it is edited afterwards. Read the policy that applies to you.
Plenty of people arrive here from the other direction. They wrote the piece, a tool disagreed, and somebody assumed a model produced it. Detectors score patterns, and patterns are shared by anyone writing carefully in a formal register. Liang et al. at Stanford HAI found in 2023 that 61% of TOEFL essays written by non-native English speakers were falsely flagged as AI across seven detectors.
In that position the score is not your argument. Your drafting history is. The signed drafting record on this site captures a session as you write it, including timestamps, how much you cut, and whether a large block arrived in a single paste, then exports a signed file using ECDSA P-256 that anyone can check in their own browser without an account and without contacting us. It works on the next piece rather than the finished one, so start it early.
A score is a reading of the prose in front of it and not a statement about the person who typed it. The drafting record is the part that speaks to authorship, which is why starting one on your next piece is worth more than arguing about a number on this one.
No. A score is a judgement about the writing in front of it, not a lookup of which system produced it. Nothing in ordinary prose carries a signature, so no tool can name the model. The checker here reports a number from 0 to 100 and one of three verdicts, and it never names a source.
No. Where the text was generated leaves no trace in the text. A detector reads sentences, and sentences from a small model on your own machine carry the same habits as sentences from anything hosted, largely because much of the instruction data behind both is the same kind of material. Local generation buys you privacy and free regeneration, not a different reading.
It helps, and it is not enough on its own. Generating at or near zero gives you the most predictable phrasing available, which works against you, so a low setting is worth changing. A higher setting buys variety in wording rather than specifics, and it drifts further from the facts. The specifics still have to come from you, which is the part of the work no sampler does.
Yes. The tells match phrasing rather than provenance, so the check behaves the same on any derivative, any quantized build, and any writing product that never tells you what it runs underneath.
Ten credits a day per network, free, with no login, no card and no email address to verify. The checker is one credit per run at any length. The humanizer is one credit per fifty words, so a 400-word draft is eight credits. The proofreader, word counter, readability checker, AI vocabulary checker, originality check and signed drafting record are free and unlimited, and the writing tells page runs in your browser at no credit cost.
It rewrites every sentence, so read the output against your sources before it goes anywhere. Local models invent citations that look correct the same way hosted ones do, so anything factual in the draft needs checking whether or not you rewrite it.