When Brian Hood discovered that the ChatGPT artificial intelligence tool attributed to him a criminal past, the Australian politician found himself facing a dilemma that engineers are working hard to solve: how to teach intelligence to erase mistakes.
The legal option represented by Brian Hood threatening last April to file a defamation lawsuit against the company “OpenAI” (the creator of “ChatGBT”) does not seem an appropriate solution. Also, the solution is not to completely reset the parameters of the artificial intelligence, as training the model again takes a long time and is very expensive.
Specialists believe that the issue of “unlearning,” meaning making artificial intelligence forget some of what it has been taught, will be extremely important in the coming years, especially in light of European data protection legislation.
Lisa Giffen, professor of information sciences at RMIT University in Melbourne, Australia, confirms that “the ability to erase data in learning databases is a very important topic.” But she believes that a great effort is still required in this field, given the current lack of knowledge about how artificial intelligence works.
With the vast amount of data on which AI is trained, engineers are seeking solutions that allow greater specificity, removing false information from the knowledge of AI systems in order to stop its spread.
The topic has gained momentum during the last three or four years, as explained by Miqdad Karmanji, an expert researcher in this field from the British University of Warwick.
Google DeepMind, which specializes in artificial intelligence, worked to address this problem. Last month, experts from the American company published, along with Kurmanji, an algorithm dedicated to erasing data in important linguistic models, such as the “GPT Chat” and “Bard” models (from Google). »).
– Correct some biases
More than a thousand experts who participated in a competition launched by the American company devoted themselves to working between last July and September to improve methods of “unlearning” artificial intelligence.
The method used, similar to what other research has found in this field, is to introduce an algorithm that orders the artificial intelligence not to take some of the acquired information into account, and does not include modifying the database.
Miqdad Karmanji confirms that this process can be a “very important tool” to enable research tools to respond, for example, to deletion requests, in accordance with personal data protection rules.
He confirms that the algorithm that was developed has also proven effective in removing copyright-protected content or correcting some biases.
But others, such as Yann Lucan, head of artificial intelligence at Meta (Facebook and Instagram), seem less convinced of this effectiveness.
Lucan clarifies that he is not saying that this algorithm is “useless, uninteresting, or bad,” but rather that he believes that “there are other priorities.”
Michael Rovatsos, professor of artificial intelligence at the University of Edinburgh, believes that “the technical solution is not a successful solution.”
He explains that “unlearning” will not enable broader questions to be asked, such as how data is collected, who benefits from it, or who should be responsible for the harm caused by artificial intelligence.
Although Brian Hood’s case was dealt with, without explanation, after it received widespread media attention, which resulted in the correction of the data processed by ChatGPT, he believes that the methods that should be used at the present time should remain manual.
“Users should check everything, in cases where chatbots give wrong information,” says the Australian man.
#Experts #seek #artificial #intelligence #erase #mistakes