If you load excerpts from the Bible or the US Constitution into an AI text detection program, it will almost certainly tell you that those texts were written by a neural network. A similar inspection result has been obtained many times. Each time, the screenshots went viral on the internet and created a wave of jokes. And if James Madison isn’t a time traveler, there’s only one reason: the device is defective.

wrong activation

This issue has been discussed lively on the English Internet after a wave of accusations from teachers to students and students. They claimed to have found signs of ChatGPT use in the trial.

In the US and some European countries there is a very strong tradition of relying on essays as a tool to assess students’ learning, and teachers are reluctant to abandon it. But they also don’t want to allow the possibility of fraud with the help of neural networks, so they use artificial intelligence tamper detection tools. The problem is that experience has already shown their unreliability. Some experts believe that due to the large number of false positives, detectors such as GPTZero, ZeroGPT, and OpenAI’s Text Classifier cannot be used to detect text generated by large language models (LLMs) such as ChatGPT.

To explain why these tools make such obvious mistakes, you first need to understand how they work.

• First, it is useful to understand that artificial intelligence works on the principle of compilation. It has a specific set of data that it is trained on and produces its answers by actually gluing together different parts of the text, making predictions about the probability that a given word will come after the previous one. So artificial intelligence does not invent anything new on its own. It will simply combine all existing snippets to create new ones that will likely be typical and cliché.
• Second, at the core of at least some AI detection tools is the principle of surprise or astonishment, a measure of how much a piece of text the AI ​​model deviates from what it learns during training. It assumes that natural human language is chaotic and unpredictable. However, the dataset on which the AI ​​is trained is limited and often builds on a variety of literature, articles, and other structured, well-organized and predictable texts. So if the system detects this “surprise”, it will assume that the text was written by a human. On the contrary, if it sees matches (if it has access to) millions of other queries on Google, it may think that the result was generated by a computer based on memorized information or analysis of the Internet. Because these detection programs know the principles of artificial intelligence, they know the principles of information gathering, and they know that the computer will never invent things for you that are not taught during training.

Here is a simple example. Let’s say you guessed the next word in “I want a cup…”. Most people will fill the void with “water”, “coffee” or “tea”. A language model trained on a large number of texts will do the same, because these phrases are often found in works. Any of these three results would not be surprising, as the estimate is fairly accurate. But if it were for example “I would like a glass of spider”, then both man and machine would be very surprised by this sentence. At the time of writing, no search results were found for this phrase. However, if you look for options with coffee, water or tea, you’ll get millions of them. This is a very crude example, but it illustrates the logic of reconnaissance tools.

This brings us to the interesting case of the US Constitution and the Bible. They The texts are so embedded in these AI models that detectors classify them as artificially created.creates false positives.

The US Constitution is a text that has been repeatedly introduced into the curriculum of many mainstream language models. As a result, many of these major language models have learned to produce texts similar to the Constitution and other commonly used educational texts. GPTZero predicts text likely to be produced by major language models, and so this fascinating thing happens.
– says GPTZero creator Edward Tian.

people are not original

The problem is people can also create content of low complexity. Lyrics can be predictable. Writers may pass on the sentences they heard or read somewhere, without knowing it, thinking that they were made up by themselves. However, as they often say, everything was invented for us a long time ago. There are not many ways to say the same sentence. This deeply undermines the reliability of AI typing detectors.

Another point that GPTZero and similar ones pay attention to is the variability of sentence lengths and structures in the text. People often exhibit a dynamic writing style that results in texts of varying sentence length and structure.. For example, we can write a short simple sentence followed by a long complex sentence, or we can use many adjectives in one sentence and none in the other. This variability is a natural consequence of human creativity and spontaneity. Instead, AI tends to be more consistent and uniform (at least for now). Language models that are still in their infancy produce sentences of more regular length and structure. The lack of variability can result in a low discontinuity score, indicating that the text may have been created by artificial intelligence.

But even this indicator is not reliable because people have different writing styles. That’s why we like some writers more and some less. Everyone has their own way of creating text. In addition, the AI ​​model can be trained to mimic more human-like variability in the length and structure of sentences, making it difficult to classify posts according to this indicator. Research shows that their writing is becoming more and more human-like as AI language models evolve.

After all, there is no magic formula that can always tell the difference between human-written text and machine-generated text. AI write detectors can make a strong guess, but margin of error is too large to rely on them for an accurate result and making sure that some text was created dishonestly. The cost of such false accusations can be very high in some places.