Most AI text detector tools that detect text in an image are based on AI outputs, offering only one output to you- say, 87% AI; take it or leave it! That figure alone doesn’t give you an explanation as to what the text is saying, and it certainly doesn’t clarify how to edit the text. That is why I made Veritype (pun intended), a program that analyses a text and visualises the actual signals that cause the score, as well as the rewrites in an alternate mode, in which the same signals are applied to a natural human text so it fits the same rhythm and can be understood without difficulty on its own.

This article will cover all aspects of the tool, what each score represents, and how to utilize both the analyzer and humanizer to produce the text desired. 

What Veritype Actually Measures

Veritype does not look at AI detection as one out of ten; it decomposes the score into four separate and explainable parts, each of which has a weighted component of the score.

One feature of Burstiness is the variability of sentence length throughout the text. If all of the text is machine-generated, it will tend to have a more level writing span than a human writing a text normally does. Machine-generated text has a narrower, more uniform range compared to human writing in a text. Vocabulary diversity (using a type-token ratio of the moving average type) is a measure of whether the text continues to expand into a larger vocabulary versus using common, everyday words with ease.

Using a surprisal-based proxy, word predictability alerts text to the tendency towards polished, “thesaurus-style” word usage, over basic, familiar vocabulary. Compares the use of those floatier phrases or AI-typical phrases to a big list of stock AI phrasing, such as ‘in today’s fast-paced world’ or ‘it is important to note that’, and looks at how many times each phrase appears every 100 words.

They all receive a score, a weight in your final mark, and a narrative that explains your score in plain terms… for example, rather than using “72/100,” each of these four signals receives its own score, its own weight in the final mark, and its own written explanation of why the mark is what it is. 

How to Use the Analyzer

Drag and drop a .txt, .docx, or .pdf file into the main box or paste your text into the main box on the Analyze Text tab — nothing is uploaded anywhere – it’s all done in your browser! It can be helpful to tweak Detection sensitivity between Strict, Balanced, and Lenient: Strict will raise allot more writing style flags that are right on the edge of being categorized as AI, while Lenient will log only the most blatant and most obvious AI patterns. The included sample texts enable you to download an AI-generated marketing paragraph, a regular human blog snippet, or a mixed sample with academic texts with a single click simply to see how it will work out.

Analyze Text executes the entire engine in about 20ms, all locally. The results begin with an overall AI-typicality gauge and conclude with either a verdict (Likely human-written, Mixed signals, or Reads as AI-typical), along with a confidence percentage, which we refer to as “confidence weight. Sample size, degree of similarity of the four components, and the distance from a decision boundary are taken into account in the calculation of confidence, and it is more accurate to report lower confidence on a smaller sample than to sound confident when really you aren’t. 

Reading the Detailed Breakdown

The confidence breakdown is below the main verdict and gives the scores for every component, the weights, and the one-line technical item in short form (e.g., variation index from the sentence length or measured MATTR percentage). This is followed by an explanation section written in plain text, which parses the numbers into sentences that can be used by a non-technical reader, such as: vocabulary remained unusually diverse throughout, which is a common indicator of AI.

This sometimes gets practical as the sentence heatmap indicates the colours: red for flagged sentences with AI-typical characteristics and green for text-literate sentences — and hovering over a sentence reveals the precise reason: Is it received as a pattern that is too long, or is it an exact hit on a repeated AI-typical sentence, or does it consist almost entirely of ordinary common words. This same approach is also applied at the paragraph level—the paragraph scores—which can be particularly helpful when editing longer documents to find out which paragraphs are in greater need of editing than others and to handle them accordingly.

And the phrases that are AI-typical appear in the section list; each has two or three natural alternatives that can be swapped in manually if you’d rather edit by hand than go through the entire humanizer. 

The Writing Quality Checker

Additionally, Veritype features a Hemingway-style writing quality score, independent of the AI-detection score itself, simply checking whether something seems clear to write. It identifies passive voice constructions, counts words that aren’t needed in -ly on your manuscript, alerts you to filler and hedge words (basically, sort of, it is worth noting that…), identifies words that should not be in your manuscript and would be hard to read (passive voice) based on the Flesch-Kincaid grade score, and flags words that are somewhat repeated in your work that feel like a tic.

This will run automatically after each analysis and will be assigned a quality score between 0 and 100; a text could be very “humane” and still be poorly written, and vice versa — it is not the same thing. 

How the Humanizer Actually Works

Unlike many rewriting tools that run a text through just one AI model and provide any output, the Humanize tab operates a bit differently. The Humanize tab is a bit different from a lot of rewriting tools that take a text through one AI model and hope for a good result. Veritype instead generates several rewrite candidates internally: two in case of Light intensity, 4 in case of Medium intensity, and 6 in case of Heavy intensity, then re-analyzes all rewrite candidates with the same detection engine on the Analyze tab, and retains the one that scores the lowest.

It does this for several iterations until a candidate crosses the line between human readability or a set number of iterations – but it does not overwrite texts that are near that line.

Use it by pasting your text into it, setting the Intensity slider (lower values better suit casual and medium levels and leave more phrases intact; higher values will cause a greater variation in sentence rhythm and are suitable for formal, academic, journalistic and SEO text, and some others), selecting the Tone (lower values sound more casual, higher values are more formal, academic, journalistic, SEO etc., and leave a bigger variety of contractions and connector words intact) and finally, if necessary, entering Keep terms at the bottom separated by commas (you don’t want to keep changing terms), brand names, product titles etc.).

Humanize does the entire best-of-N process and returns a result as a side-by-side, side-line duo-gauge, with the AI Score before and after, as well as the number of stock phrases and word choices changed on the way. 

Reviewing What Changed

Either after a rewrite has appeared, you can switch from a line-by-line display to a complete view of the rewritten text, which you can edit directly in the output’s textarea; or you may switch back to a line-by-line view so you can see each sentence and easily extract the corresponding AI-or-human tag, and then see exactly which sentences remain machine-typical and which do not.

A Show changes allows diff to show the original and the rewritten sentence line by line; a button Re-check result lets you re-run the analyzer on your rewritten copy to make sure that you didn’t accidentally reintroduce AI-typical patterns. All rewrites you create appear in Version history, so you can revert to an earlier version or view two versions at once. 

Batch Tools for Multiple Texts

For cases that have more than one piece of text to be checked or rewritten, the Batch Tools tab supports up to 40 separate samples that are divided by a line terminated with –––, and allows a choice between Batch Analyze (reveals the average score and a summary found in the column How many texts came back human-like? as well as How many texts read as if AI wrote them? applies the same intensity and tone settings to all the samples) and Batch Rewrite (runs the Complete passage humanizer over every sample using the same intensity and tone settings as chosen).

All the results come back on separate cards containing their own score chips with the single-text Analyzer plus a separate writing output box that can be edited, so you can analyze a batch of ten blog drafts or ten product descriptions; the same card can be used anytime you need to retest it. 

The Dashboard

All your usage metrics are stored in the browser’s local storage, and you can keep an eye on them from the Dashboard tab: number of scans performed, number of words analysed, the average AI score out of every check performed, number of texts that you “humanized”, and a little sparkline chart that represents the performance of your AI score on the timeline of each check you made, letting you discover if your writing goes in a trend of becoming more or less human. Here, there is no sending out any of this; it is truly a personal record, and a Clear data button obliterates it totally when necessary. 

Real-Life Examples

A blogger or anyone who is doing blogging work knows that you can paste a draft into the Analyzer, find that the content has a high score because of its “AI-typical phrase density” section, and immediately start at the “phrase list” section in your hand to edit out a few of the more rehashed phrases instead of having to rewrite it.

If the student is checking the assignment before they submit it, they may see a high score in two paragraphs and wonder why: It is likely due to their use of unusually uniform sentence lengths within the paragraphs, not to anything in the actual content.

One freelance writer working on 10 client articles can take the 10 articles and submit them to “Batch Analyze,” where it specifies the 3 that are AI; and then humanize the 3 that were flagged as AI, versus the 10 that were not.

A copywriter in the unique company voice can use Keep Terms to make certain the brand voice doesn’t change and can use Heavy intensity to generate the maximum amount of rhythm variation with none of the frozen terms being changed. 

Why the Scores Are Trustworthy — and Their Real Limits

All the scores in this tool were computed by statistical patterns in texts, assessed directly against the text submitted, such as sentence length, vocabulary overlap, word rarity, and matching words to real-world text documents of AI-produced writing, instead of a black-box model making a strange guess. All the transparency is for: if anyone hasn’t hit the correct signal, you can see which one you hit and, for your own peace of mind, re-read the flagged sentence! Again, this specific AI text detector detects regularities, not authorship; by extension, so will any AI text detector.

A (very formal, very repetitive) human can get as AI-typical as AI, and AI-assisted text that a person has massaged can read as human. Veritype doesn’t claim to be proof of someone’s authorship, and it doesn’t claim that at all; it’s there to improve and give stylistic feedback on texts. 

Frequently Asked Questions

Is this AI text detector actually accurate? 

Takes into account real statistical patterns – such as sentence rhythm, vocabulary range, word predictability, and stock-phrase density – and provides details on which patterns led to the score rather than just one arcane number. It is like all detectors: just a qualitative guess and not a proof of authorship. 

Does humanizing a text guarantee it will pass every AI detector? 

No tool will promise against a 3rd party detector. What it does do is measurably decrease the same statistical signals this and most detectors are looking for — sentence uniformity, synonym churn, and stock phrasing. 

Is my text uploaded anywhere? 

No — analysis and rewriting both are done within your browser, not on the server, including reading the file that you have uploaded — from .txt or .docx or .pdf. No data is pushed to a server.No data is pushed to a server. 

What’s the difference between the AI score and the writing quality score? 

The AI rating reflects the nature of the writing patterns in order to determine how typical a machine would exhibit these patterns. The quality of the writing score is a separate Hemingway-style check on passive voice, weak words, and complex sentences; it’s possible to be high on one and low on the other. 

Conclusion

A useful AI Text detector must supply its explanation, and a useful AI Text humanizer needs to apply its explanation to solve the problem, instead of merely feeding a text through a black box twice. The humanizing side has a rewriting process that is best-of-several-candidates, and with nothing uploaded to the cloud, the rewriting process, as well as transparent component-based scoring, are located on the analysis side of Veritype.

We’re polishing our own blog post, or we’re dealing with several client drafts — and it’s ultimately about the reasons behind how something reads, then making it read how we want it to read.