With the ChatGPT boom, many are already misusing and resorting to plagiarism to write texts. However, some tools are already being developed to detect it.
We are still in the early days of LLMs, but we are already seeing the influence they are having in our world today. ChatGPT, without going any further, which we have already talked about in other articles, is now facing major bans by the media, schools, and institutions.
And it is that some are using the tool to build texts or to use its utilities without paying attention to the false information or lack of grammar that can sometimes be seen in its results.
For example, Stack Overflow, a question-and-answer website for programmers, has banned answers generated by ChatGPT due to its propensity to provide false information to technical questions. On the other hand, New York has done the same, affecting 1,800 public schools.
In short, the fight against AI plagiarism will be one of the big topics of 2023, and who knows what will happen when GPT-4 is released? The good news is that great efforts are being made to detect the results generated by this and other tools.
So you can know if a text is written by a person or by a machine
1. GPT-2 Output Detector: This application detects if a text has been generated by GPT-2. It uses the RoBERTa OpenAI Detector base model. You can try it here.
Allows you to paste the text into a box and immediately see the probability that the text was typed by AI. According to OpenAI research, the tool has a relatively high detection rate.
2. GPTZero: Created by Edward Tian, a student at Princeton University, it seeks to determine if the text has been written by humans. We leave you the application linked.
If the text is very complex, it is more likely that it is written by humans. The lower this parameter that you state as “bewilderment”, the more likely it is that it is generated by artificial intelligence.
And it is that humans tend to write longer or more complex sentences and shorter sentences, which causes peaks of perplexity in human writing. Machine-generated phrases are more uniform and coherent. Including small typos and grammatical errors on purpose increases your puzzlement score, making your content more “human.”
Finally, it should be noted that there is an application called Watermarking (watermark) that is still under development but Scott Aaronson, a researcher, has explained that a “pseudorandom cryptographic function” would be used so that any text that GPT generates is indicated with a hidden watermark that only OpenAI will know about.
Knowing this, ChatGPT will still generate text but OpenAI will be able to discover a watermark and thus identify if it is being used for plagiarism. With all this, it is expected to deal with these misuses that have been born with the new boom of this tool.