AI presents businesses with real opportunities to improve efficiency. It also brings new challenges, which is why businesses are increasingly looking to use a specific form of AI against AI: AI detectors.
There are legitimate reasons to use AI detectors: protecting brand voice, fraud detection, spotting plagiarism, getting value for money from contractors, and more.
But AI detectors aren’t infallible, and for most businesses the risks of using them outweigh the benefits. Here’s what you need to know before you use one.
How do AI detectors work?
AI detectors are specialized software that mathematically analyze patterns to estimate the probability that text was machine-generated. Popular detectors include GPTZero, Copyleaks, Originality.ai, Sapling, Turnitin, and Grammarly AI Detector.
What AI detectors look for, in technical terms, comes down to two metrics:
- Perplexity measures how predictable a sequence of words is. Lower predictability means less likelihood of AI involvement, because AI models are trained to guess and use the most statistically probable words. Humans make more varied and eccentric word choices.
- Burstiness measures variation in sentence length and complexity. Human writing “bursts” with a mixture of sentence-types, whereas probability-seeking AI tends to write in a regular, smooth, almost mathematical way. Low burstiness writing strongly indicates AI-involvement to a detector.
Some AI detectors also use stylometric analysis, evaluating vocabulary choice and grammar, hunting down suspiciously rigid and formal structures.
A handful of AI models are starting to use watermarking, which involves embedding invisible and detectable patterns into AI-generated text at the point of creation. But because most models don’t currently leave a trail, you can’t yet rely on this solution.
Are AI detectors reliable?
AI detectors are never 100% reliable, and they’re often significantly less reliable than that.
Accuracy varies significantly between tools, content type, and (critically) exactly how AI was used to create the analyzed content, if it was used at all.
When detectors work
Paste unchanged (or ‘raw’) AI output into any AI detector and the detector will detect it. Raw output is easy for the detector to identify because the statistically probable pattern it’s looking for hasn’t been broken.
For example, we pasted text generated directly from Claude Sonnet 4.6 into Sapling. Sapling had 99.5% confidence that this AI-generated text was AI-generated.
Originality (another leading AI detector) flagged the same paragraph as AI with even greater confidence: 100%.
Detectors can also identify AI-writing when the text submitted isn’t 100% AI generated, although significantly less reliably, especially when the text is short (less than 300 words) and the patterns detectors are looking for don’t have time to establish themselves.
When detectors fail
It’s well-documented that AI detectors find false positives: signals of AI use in text entirely written by humans. One recent peer-reviewed study of academic writing(nyt vindue) rated the overall accuracy of Originality and Turnitin when identifying AI-generated academic texts as only 69% and 61% respectively.
One issue is that humans often naturally write like AI, particularly if they’re non-native language speakers, who (like LLMs) use simple, highly grammatical phrasing. A Stanford study(nyt vindue) had seven popular AI detectors analyze Test of English as a Foreign Language essays by Chinese students. The average false positive rate was 61.22%.
It’s also hard for detectors to spot when humans deliberately disrupt the statistical patterns they’re looking for. AI-assisted authors don’t have to do much to break those patterns: rewrite a few sentences, vary sentence structure, use less predictable words.
If this seems like too much work, they can get it done for them by one of the increasing array of ‘humanizer’ tools that have inevitably arisen to outsmart detectors.
They can also simply stick to their chosen LLM and use smarter prompts. A 2025 study of 12 AI detectors(nyt vindue) found that “minimal polishing with GPT-4o can lead to detection rates ranging from 10% to 75%, depending on the detector”.
We asked Claude Sonnet 4.6 to write a 300 word essay about AI detector accuracy, and asked it to write like a human, with varied sentence lengths, direct opinion, and unexpected word choices. Sapling scored the text as only 12% likely to be generated by AI.
Originality analyzed the same text and declared that it was 100% confident that the text was AI generated.
Originality looks to be the superior tool based on this test, but it’s important to recognize that both tools were equally confident in their analysis. If you wrote the analyzed text, you’ll know which detector is correct. If you didn’t, then you can never be sure.
How AI detectors endanger your data
The bigger issue with AI detectors is one that many businesses overlook entirely: the risks they present to the privacy and security of your data. These are similar risks to those posed by AI in general:
- Compliance risks: Detection models inherit biases from training data, creating potential for discriminatory outcomes and breaking equal opportunity laws. Many detectors are “black boxes”, and frameworks like the EU AI Act(nyt vindue) and the NIST AI Risk Management Framework(nyt vindue) demand organizations must provide meaningful logic for AI outputs.
- IP and training data risks: Many AI detectors have Terms of Service (ToS) that grant the provider the right to store, analyze, and use submitted data to train future models.This can mean that your provider has the legal right to use whatever product roadmaps, financial projections, and proprietary research you’ve submitted with no obligation to guarantee its privacy or prevent it surfacing in a competitor’s output.
- Data security risks: AI detectors are vulnerable to cyberattacks, and submitting content to a third-party detector extends your attack surface to include your provider’s. If your provider suffers a breach, your data will be in it. Free, unvetted tools — the very type employees can be tempted to reach for — won’t let you audit a provider’s security practices in advance.
- Data sovereignty risks: Submitting a document to a detector can mean transferring text to foreign servers, potentially stripping it of local protections and exposing it to extraterritorial government surveillance under the CLOUD act(nyt vindue).
For most businesses, the question isn’t whether AI detectors work: it’s whether the risks of using them are worth taking. If you’re working with any sensitive, original, or confidential data, the answer is no.
FAQ on AI detectors
How do AI detectors work?
AI detectors scan text looking for patterns they’ve been trained to associate with AI-generated text. These patterns appear because AI generates text by algorithmically predicting and using the most probable next word.
What do AI detectors look for?
The main indicators of AI-generated text that detectors look for are low levels of ‘perplexity’ (unpredictable word choices) and ‘burstiness’ (variation in sentence length and structure).
Some AI models are starting to experiment with watermarking, which means embedding invisible (but detectable) patterns into text at the point of generation. However, since only a few models have started to do this, watermarking doesn’t yet provide a foolproof test for AI-generated text.
Do AI detectors work?
Accuracy depends on the detector used, the type of text being analyzed, and how (if at all) AI has been used to generate the text. Detectors perform well on raw, unedited AI output. On anything more complex — edited content, human writing by non-native speakers, AI output that’s been lightly paraphrased — accuracy drops significantly.
Accuracy aside, the bigger problem with AI detectors is the threat that using them can pose to your data security. Submitting content to a third-party detector means transferring it to external servers, potentially exposing it to cyberattacks, data harvesting, extraterritorial surveillance, and handing a provider legal rights to your submitted content.
For most businesses, especially those working with any sensitive, original, or confidential content, the data risks that come with third-party AI detection outweigh the benefits.
