The only official tool that should be able to identify the use of ChatGPT it is still unreliable: the disarming admission comes directly from the OpenAI research team, the very ones who developed the most famous artificial intelligence of the moment.
On the official blog of OpenAI a long study has been published on what has been defined as the “classifier trained to distinguish a text written by artificial intelligence from one written by man”. It’s called the Classifier and should help you spot any automated disinformation campaigns, false academic texts and chatbots that pretend to be human.
Il problem is that this tool is still very immature: “Our classifier is not completely reliable”, OpenAI underlined. The response was obtained by subjecting him to various challenges with texts in English. Eventually he was able to correctly identify 26% of text written by ChatGPT (true positives), branding them as “probably written by AI”; however, against a 9% erroneously credited to artificial intelligence instead of man (false positives). In practice, it has such a high margin of error that legally the its value could be undermined by a first-year law student.
However, we must not give up, because reliability seems to improve as text length increases and above all Classifier gets better with constant training. A bit like training a hound puppy that tomorrow will have to deal with an adult prey. All normal.
Does Generative AI Infringe Copyright? The 2 reasons why Getty sues Stable Diffusion
by Emanuele Capone
Weaknesses of Classifier
OpenAI recalls that its classifier cannot be used as a primary decision-making tool, “but as a complement to other methods for determining the source of a piece of text”. The reasons for this caution they are due to the fact that the tool is very unreliable with short texts (less than a thousand characters) and sometimes makes mistakes even with longer ones. Net of these performances, it does even worse with languages other than English and “is unreliable on code”. So forget about using Classifier for Italian texts.
And other problem is that a document written by an artificial intelligence can be modified to evade the classifier: “Classifiers like ours can be updated and re-trained based on successful attacks, but it’s unclear whether detection has a long-term benefit,” the researchers added.
The only consolation is that ChatGPT has proven that training is an essential component to make an artificial intelligence evolve. And so it is expected for Classifier, that is currently available for free just to promote its evolution.
Classifier for professors, but not only
The world of academia, education and research have undoubtedly been the most sensitive to the success of ChatGPT. And indeed the OpenAI team itself admits that this new resource is designed especially for them, although she hopes it will impact other categories: “We are engaging with educators across the United States to learn what they are seeing in their classrooms and to discuss the capabilities and limitations of ChatGPT, and we will continue to expand our reach as we learn,” they explained. Again: “We also welcome any resources that teachers are developing or found useful (course guidelines, code of conduct and policy updates, interactive tools, AI literacy programs, etc.).
What to ask ChatGPT about Cospito and the 41 bis
by Riccardo Luna
The alternative tools
OpenAI is not the only reality that exists working on a detection tool as Classifier, but at the moment it is the most authoritative one. Online announcements are multiplying from professionals, researchers, companies and students who assure that they have found a solution to the problem. As we had already mentioned on Italian Techwe speak for example of GPTZero of a Princeton student, Corrector App of an alleged development company based in London, AI Content Detector and many others.
The problem is not whether or not they are able to function (net of projects carried out by malicious people to steal data or payments), but the certification of the tools. Who proves the reliability?
In Italia recently several companies that operate online have begun to ask employees and their communication agencies not to use ChatGPT, since Google could damage the indexing of online text pages created with competing artificial intelligences. For an e-commerce it would be a economic damage devastating the disappearance of the respective product pages from the search engine results. Yes, but how to make sure that no one uses texts invented by ChatGPT?