How AI text detection works
What each of the seven detectors measures, what it relies on, where its limits are and why their scores are never merged into one number.
Blog
How AI text detection works for Czech and Slovak, how to read the results and what to do with them at school, in editorial work and in your own writing.
The basics
Glossary
What perplexity, burstiness, stylometry, a false positive, humanisation or a watermark mean — each term in one or two sentences.
How it works
Can a detector tell whether ChatGPT, Claude or Gemini wrote a text?
Why detectors only see resemblance to machine writing in general, when a specific model can be named and why a decision should not rest on it.
Why two detectors give a different number for the same text
Why scores differ between tools, what the disagreement says about the text and why averaging the numbers makes no sense.
Half yours, half AI: what a detector shows on mixed text
How the score behaves on a text written together with a model, why it disappears in a long document and what to read instead of the overall number.
What an AI watermark is and why you cannot rely on it in a text
What a model's hidden mark in a text looks like, which models add one today and why you cannot lean on it when judging a text.
Why AI text detectors do worse in Czech and Slovak than in English
Czech and Slovak have free word order, rich inflection and smaller training corpora. What that does to detection and how to read the results.
What the percentage from a detector means and what to do with it
A score of 0 to 100 is not the probability of cheating. How to read it, where the thresholds are and what to do with a result in the middle.
When your own text comes out as AI: what false positives are
Why an honestly written text sometimes ends up with a high score, and how to resolve that situation without accusations.
Perplexity, burstiness and stylometry: three different routes to detection
The three approaches AI text detectors are built on, how they differ and why they give different results on the same text.
Detection against your own hand: what a personal style baseline is
How comparing a text against your own older texts works and why it gives a different answer than a general detector.
How to spot AI text by eye: nine signs to watch for
The specific patterns that give generated text away even without a tool — and why none of them is enough on its own.
How much text a detector needs to make sense
Why short texts cannot be judged reliably, where a reasonable minimum lies and how to split a long text.
Why texts by non-native writers score worse with detectors
Why a simpler vocabulary and careful sentence structure push the score up, who it affects most and how to read such a result.
Why a translated text comes out as machine-made
What translation does to a text — by hand or by machine — and why the score rises afterwards even when a human wrote the original.
Why a detector needs to know what kind of text it is reading
What changes when you pick the text type before a check, why an e-mail has different limits than a term paper and what Type template measures.
Why the same text scores differently the second time
What causes the difference between two checks of the same text and how big a difference is still normal.
Practice
Can a grammar checker or writing assistant raise your AI score?
Which edits in grammar checkers and writing assistants push your own text towards a machine style, which do not, and how to protect the original.
How to set an assignment a model cannot do for the student
What makes an assignment resistant to a model — own data, concrete context, staged deliverables and a short defence.
How to get through thirty term papers in a morning
A procedure for working through a whole batch of student papers without spending half an hour on each one.
Humanising a text vs. rewriting by hand: when each is worth it
The difference between rewriting a sentence automatically and rewriting it yourself, and how to use humanisation so the text does not end up worse.
Checking a master's thesis chapter by chapter
Why not to check a whole thesis at once, where false alarms pile up and when to schedule the check.
How to fit detection into an editorial process with freelancers
Where an AI check belongs in an editorial workflow, how to communicate it to authors and what to do with the result.
Cover letters written by AI: what to do about it in hiring
Why cover letters are now almost useless as a filter, and how to adjust hiring instead of hunting for culprits.
Verification questions instead of accusations
Why a conversation about the text works better than confronting someone with a number, and how to prepare the questions.
Why keep a history of checks and what you can read from it
A one-off check says little. What several checks of the same author over time can tell you.
Checking PDF and DOCX without copying the text
Which formats can be uploaded, what to do with a scanned PDF and why checking the file beats checking an excerpt.
How to check a text you commissioned from an outside writer
What to agree up front, what to read on delivery and how to raise a suspicion without wrecking the working relationship.
Your first text check, step by step
What to do on a first detection, in what order to read the result and what to avoid right at the start.
Can I use AI for my coursework? How to find out and how to do it
Where the rules live, what is usually allowed, how to declare your use and what to avoid in an academic paper.
Choosing
How to choose an AI text detector: seven things worth comparing
What to compare AI text detectors on — Czech and Slovak, sentences instead of a single number, handling false positives, history and how your data is treated.
AI content and SEO: what search engines actually object to
Search engines do not punish AI text as such. What really drags results down, and where detection is useful for content teams.
When not to run a detector: cases where the result settles nothing
Short texts, fixed wording, foreign languages and cases with no chance to ask the author — when a detector will not help and what to do instead.
Is humanising a text cheating? Where the line runs
When rewriting is ordinary editorial work, when it is dodging the rules, and how to tell the two apart.
Rules
Accused of writing with AI: what you can actually show
Which evidence holds up when an honestly written text comes out as machine-made, and how to talk about it with a school or an employer.
Can a detector result be proof?
What a detector result can carry and what it cannot, how to handle it at school and in a newsroom, and what to do instead of accusing.
What happens to the text you put into a detector
What to ask about any tool, where the text is stored in DetekceGPT and why this matters with other people's work.
How to write school rules for using AI
What a rule needs to contain to work — concrete boundaries instead of a ban, and an agreed procedure when there is a suspicion.
Company rules for AI: what has to be written down first
What a company AI policy has to cover — where it is allowed, what must never be pasted into a chat and who is accountable for the output.
AI in academic writing: what publishers ask of authors
Why a model cannot be a co-author, what belongs in an AI use statement and where the line usually sits for language editing.
Is text from AI plagiarism? How it differs from copying
Why generated text passes a similarity check, why it is still usually a breach of the rules and what academic codes tend to say about it.