The two ways AI text gets flagged
People talk about "AI detection" as if it were one thing. It is really two very different mechanisms, and they fail in different ways.
Detectors that guess
A statistical or stylistic tool reads the finished text and estimates how likely a machine wrote it. It looks at things like word predictability and sentence rhythm, or runs the text through a classifier model trained on human and AI samples. It has no inside knowledge. It produces a probability, not proof.
Watermarks that are embedded
The company that made the model weaves a hidden, machine-readable signal into the output at the moment it is generated. A matching detector can later check for that signal. This is about provenance, where the text came from, and it only works if the maker added the mark and you hold the matching key or tool.
The practical upshot: a detector is judging your prose, so cleaner prose changes the guess. A watermark is a fingerprint left by the generator, so editing the surface style does not reliably remove it. Keep the two apart and most confusion disappears.
Perplexity and burstiness, in plain terms
Two words come up constantly with detectors. Perplexity measures how surprising each word is to a language model. AI text tends to pick the "safe," highly probable next word, so its perplexity is low and even. Burstiness measures the variation in sentence length and complexity. Human writing tends to be bursty (a long, winding sentence, then a short one), while machine writing often marches along at a uniform pace. Low perplexity plus low burstiness is the classic pattern detectors chase.
Signs text may be AI-written
These are the stylistic tells detectors look for, and the same signals the Fetch It AI checker surfaces in your own text. If you arrived here from the Tips button in the tool, this is the full reference. None of them is proof on its own. The honest read is cumulative: one or two mean nothing, a dozen stacking up is a strong hint.
Overused "elevated" vocabulary
Certain words show up far more often in AI prose than in everyday human writing.
Stock phrases and hedging filler
Whole sentence openers recur so often they read like a template.
"It is important to note that..." "In today's fast-paced world..." "That being said..." / "When it comes to..." "plays a crucial role" / "a wide range of factors"Uniform structure
- Sentences of very similar length, paragraph after paragraph (the low burstiness above).
- The rule of three everywhere: "clear, concise, and compelling" triads used compulsively.
- List obsession, even when plain prose would be more natural.
- The summary reflex: an "In conclusion" paragraph restating what was just said.
Punctuation and tone
- Heavy em dash use (the reason our tool has a one-click Remove Em Dashes button).
- Curly quotes and apostrophes where straight ones would be expected.
- Very few contractions, giving a stiff, formal voice.
- Relentlessly neutral tone, over-explaining the obvious, and vague specifics (confident claims with no concrete names, numbers, or dates).
Good human writers use em dashes and the rule of three all the time. Careful prompting erases most of these tells. That is precisely why a detector reading style alone can only ever guess, and why it guesses wrong often enough to matter.
The watermarking landscape in 2026
Watermarking is a fast-moving, uneven area. What follows is measured on purpose: coverage varies by provider and model, and the picture is changing month to month. Treat specifics as "as reported," not as settled fact.
SynthID and SynthID-Text (Google DeepMind)
Google DeepMind's SynthID embeds a statistical signal directly into a model's token choices during generation. In simple terms, when the model picks among several plausible next words, SynthID nudges those choices in a subtle pattern that a matching detector can later score. It is reportedly designed to survive light editing and paraphrasing, and Google has published and open-sourced parts of the text approach (SynthID-Text). It applies where the provider has enabled it, not to text from every tool.
The EU AI Act, Article 50
Article 50 of the EU AI Act sets transparency obligations for AI systems. As reported, those provisions became applicable on August 2, 2026. In broad strokes they push providers of generative systems to mark synthetic content in a machine-readable way where technically feasible, and to disclose when people are interacting with AI. This regulatory pressure is a big reason the industry is moving toward standardized, machine-readable marking.
Industry coordination and Content Credentials
- Google reportedly signed the EU's Code of Practice on transparency in July 2026, and is reportedly coordinating with partners, including OpenAI, on more interoperable watermarking approaches.
- C2PA "Content Credentials" adds cryptographic provenance metadata: signed information about where a file came from and how it was edited. Today this is used mainly for images and audio. Because metadata can be stripped, it is provenance you can verify when it is present, not a guarantee that it is always there.
As of August 2026, OpenAI had not publicly shipped a user-facing text watermark in ChatGPT (it has reportedly researched one without releasing it broadly). Across the industry, text watermarking remains partial and inconsistent: strong in some products, absent in others, and evolving quickly. No claim here should be read as "vendor X watermarks all of its text." Where we name a mechanism, we are describing what has been reported, not certifying its coverage.
| Property | Style detectors | Provenance watermarks |
|---|---|---|
| What it checks | The finished words and their patterns | A signal embedded at generation time |
| Needs the maker's help | No | Yes, the provider must add it |
| Output | A probability or score | Present or not present (with a matching detector) |
| Affected by editing style | Yes, cleaner prose shifts the guess | Often designed to resist light edits |
| Coverage in 2026 | Many third-party tools, varied quality | Partial, varies by provider and model |
Why detectors are unreliable
Style detectors are useful as a hint and dangerous as a verdict. A few reasons to be careful:
- They output probabilities, not proof. A number like "82% likely AI" is an estimate about patterns, not evidence about who wrote the text.
- Documented bias against non-native English writers. Research has found detectors disproportionately flag essays by non-native speakers as AI, because simpler, more predictable phrasing looks "machine-like" to them.
- Plain, formal, or formulaic human prose trips them too. Technical writing, legal boilerplate, and careful academic prose can all score as AI.
- Even the makers have backed off. OpenAI retired its own AI text classifier in 2023, citing low accuracy. That is a telling admission about the whole category.
The takeaway is not that detectors are useless. It is that acting on a single detector score, especially in a high-stakes setting like a grade or a job, is a mistake. Treat it as one weak signal among many.
Where Fetch It AI fits (honestly)
Fetch It AI works on the first mechanism only: the stylistic signals. It reads your own text, shows you the tells above (overused words, stock phrases, uniform rhythm, em dash density, low contraction rate), and cleans the specific ones you choose, such as stripping em dashes and re-punctuating cleanly. You stay in control of your own writing, and you can see exactly what a stylistic reader would notice.
Fetch It AI is not a watermark remover, and cannot promise you will pass any detector.
It does not touch cryptographic or statistical watermarks like SynthID or C2PA Content Credentials. Those are embedded by the AI maker, and no surface edit reliably removes them. Fetch It AI also cannot guarantee that any text will pass any detector, because detectors are probabilistic and change without notice.
What it honestly does: it makes the stylistic signals in your writing visible, and cleans the ones you decide to clean. That is the whole promise, and we will not pretend otherwise.
Sources and further reading
Plain-text references for the claims above. Look these up directly from their publishers.
- EU AI Act, Article 50: transparency obligations for providers and deployers of AI systems, including machine-readable marking of synthetic content and disclosure of AI interaction.
- Google DeepMind, SynthID and SynthID-Text: technical overview of embedding a detectable statistical signal into generated content, including text.
- C2PA, Content Credentials: the Coalition for Content Provenance and Authenticity standard for cryptographically signed provenance metadata.