Artificial intelligence models need more criteria and tests, the researchers say

With the increase in the use of artificial intelligence – benign and hostile – at the speed of dismantling, more responses that are likely to be harmful. These include Hate speechand Violation of copyright or Sexual content.

The researchers told CNBC that the appearance of these unwanted behaviors is exacerbated by the lack of regulations and insufficient testing of artificial intelligence models.

Javier Rando, an artificial intelligence researcher, said that obtaining machine learning models to behave the way it was supposed to do so is too long.

“The answer, after nearly 15 years of research, is, no, we don’t know how to do this, and it does not seem to be improving,” Rando, who focuses on hostile automatic learning, told CNBC.

However, there are some ways to assess risks in artificial intelligence, such as Red team. This practice includes individuals who test artificial intelligence systems and investigate the disclosure of any possible and specific damage – a common way to work in cybersecurity.

Shane Longper, a researcher in artificial intelligence, politics and leadership Data source initiativeNote that there are currently insufficient people working in red teams.

While emerging companies from artificial intelligence now use residents of the first party or the second parties contracted to test their models, opening the test to third parties such as ordinary users, journalists, researchers and moral infiltrators will lead to a more powerful evaluation, according to devices Longper and researchers published by Longper and researchers.

“Some defects in the systems that people found find wanted lawyers, doctors, actual scientists, and actual scientists who are specialized subject experts to see if this is a defect or not, because the common person may not be able or will not have sufficient experience,” Longbury said.

The adoption of reports of “artificial intelligence defect”, incentives and ways to spread information about these “faults” in artificial intelligence systems are some recommendations made in the paper.

Longpre added that this practice has been successfully adopted in other sectors such as software safety, “We need this in artificial intelligence now.”

Rando said that marrying this practice that focuses on the user with governance, politics and other tools would ensure a better understanding of the risks posed by AI and users.

Project Moonshot is one of these methods, as it combines technical solutions and politics. Project Moonshot has been launched by the Singapore Media Development Authority, a large assessment assessment group to evaluate the language model that has been developed with industrial players such as IBM, based in Boston -based IBM Computer robot.

The set of tools merge the standards and red groups and the basic lines test. There is also an evaluation mechanism that allows AI companies startups to ensure that their models can be trusted and no harm to users.

The evaluation is Ongoing operation Kumar, who should be done before and follow the publishing of the models, said.

“I took a lot of startups as a platform because it was Open source, And they began to benefit from that. But I think, as you know, we can do a lot. “

Moving forward, Project Moonshot aims to include customization for the use of specific industry and enable the multi -language and multicultural red team.

Pierre Alcier, a professor of statistics at the Essec College of Business Administration, Asia and the Pacific, said that technology companies are currently Hurry up to release The latest artificial intelligence models without appropriate evaluation.

“When the pharmaceutical company designs a new drug, they need months of tests and very serious evidence that it is useful and harmless before it is approved by the government,” adding that there is a similar process in the aviation sector.

Alquier added that artificial intelligence models need to meet a strict set of conditions before approval. Alquier said that the transformation from the broad AI tools to those designed for more specific tasks will make it easy to expect to abuse and control them.

“LLMS can do many things, but they are not targeted with the tasks specified enough,” he said. As a result, “the number of possible use operations is very large for developers for all of them.”

Such extensive models make determining what is considered safe and safe, according to Search for Rando He was involved.

Rando said that technology companies should avoid excessive transportation that “their defenses are better than they are.”