The use of ChatGPT has rapidly become an everyday norm, so it is now more a matter of recognizing which work is actually due to a human and not an AI. AI detectors have been trained to recognize and compare the differences in structure and patterns between AI-generated texts and texts written by humans. In general, AI-generated texts are not complex, are easily predictable and have a non-varying sentence structure and length. This is also referred to as low perplexity and low burstiness.
If a document has 50% AI-generated text after testing, it is considered AI-generated[1]. This is because changes and errors can also be intentionally added to AI-generated texts.
How can AI-generated texts be unmasked?
Of course, there is already a direct solution to the challenge of recognizing AI texts, which is spreading just as rapidly. Who would have thought it: it’s an AI that can detect AI texts. However, the digital world doesn’t wait and so it doesn’t stop at detecting and recognizing AI-generated texts. Most AI detectors modify the texts recognized as AI-generated in such a way that they are no longer recognized as AI-generated.
For the future of AI text recognition, however, work is already underway on a “watermark system”, which should facilitate recognition and enable greater transparency.
Which AI text detectors are available & what can they do?
The selection of AI text detectors is very large. Therefore, 5 are examined in more detail below. For testing, ChatGPT created a text about the company HanseSecure, both in German and in English. The results were very similar, almost identical. The difference was mainly in the detail of the analysis report.
Copyleaks
The first tool to be presented is Copyleaks. It not only offers AI text recognition, but also plagiarism detection. It can also detect AI-generated code as well as plagiarized and modified source code. All of this is possible with Copyleaks in 100 languages, but only for a fee.
In this example, the text passages marked in purple have been detected as AI-generated. The detailed analysis is only available after registering and taking out a paid subscription. Copyleaks can be a very helpful tool if you have the budget for it, as it offers many services for a manageable cost (14$ per month). The free test can be useful to find out if there is a chance that the text is AI-generated. But there are other tools that offer significantly better analyses for AI text detection free of charge.
AIDP (AI Detector Pro)
AI Detector Pro offers the option of checking texts and documents in English, German and Spanish for AI-generated content. In addition, AI-generated texts can be modified so that they no longer have the characteristics of AI-generated texts. The first three scans in a month are free with AIDP. However, registration is required in order to use the tool.
The AIDP analysis report is very detailed and easy to understand. The AI-generated text was also detected here, but this time with a probability of 98%. As this tool offers the option of carrying out three tests per month free of charge, you can get an impression of whether the analysis and results meet your expectations. It can then be decided whether a budget should be set aside for this. Although AIDP does not offer a wide range of different services like other providers, it specializes in AI text detection. This is clearly recognizable by the very detailed report of the analysis, which makes it possible to offer the modification of the detected AI-generated texts.
ZeroGPT
In addition to the function of checking texts for AI-generated content, ZeroGPT also offers an AI translation, a grammar check and a word counter. This AI text detector shows the probability with which a text was captured by AI and marks these parts in yellow. ZeroGPT offers these features free of charge and without registration. However, the limitation here is the length of the text or document. This tool is only free for documents up to 15,000 characters and PDF export is only possible with a paid license. The costs for a license are very reasonable (from $8 per month) and the free features already offer some useful options.
Scribbr
Scribbr is a tool that is particularly useful in the area of plagiarism checking or AI text detection of very long documents. However, the free version is only available for English texts and the detailed analysis is not free.
The tool also recognized the same text as an AI-generated text. However, this was only possible in the English translation. Scribbr does not offer a subscription for the detailed analysis report and German translations, but charges per document. This AI is therefore more useful when the plagiarism check or AI detection of a very long text is required, such as a doctoral thesis. Other providers are therefore more suitable for AI detection of different texts or short documents.
Originality.ai
Originality offers several services in addition to AI text recognition, such as plagiarism detection and a fact checker that checks the accuracy of text content. The readability of a text or document can also be assessed with Originality. So far, only plagiarism detection is possible in several languages; all other checks are currently only available in English. However, none of this is available for free, but only for a paid subscription. This is definitely a disadvantage compared to the other tools, which offer at least the first scans or limited detections free of charge.
Which tool performs best?
The five AI text detectors tested all offer different services and are therefore more suitable for some areas and less suitable for others. For the pure AI text detection of many short documents, however, the free version of ZeroGPT is the best. However, if you have a budget for AI text recognition, you should definitely also consider AIDP and test it with the three free detections per month. These two providers stand out in the field of AI text detection.
But how accurate and reliable are AI text detectors?
In conclusion, it can be said that an enormous amount of time and development work has gone into the AI detectors provided. The potential can be seen in the wide range of providers and the constant further development. So far, however, when using these tools, it must always be borne in mind that they can be detected. False positives are also possible. For example, texts falsely written by humans can be classified as AI-generated by the AI detectors. The idea of working with watermarks in order to achieve a clear, comprehensible and transparent assessment of texts etc. could be a solution. Or does the AI generate watermarks after all? 😉 The journey is just beginning.
[1] Solis, T. (2023, May 15). These are AI text recognizers and how they work. Scribbr. Retrieved July 17, 2024, from https://www.scribbr.de/ki-tools-nutzen/ki-text-erkenner-funktionsweise/
