AI models need more standards and tests, say researchers

The use of artificial intelligence – Benign and controversial – increases in the speed of breakneck, the more potential harmful answers are found. This includes hatred speech, Copyright infringement or sexual content.

The emergence of these undesirable behaviors, informed CNBC that the regulation of undesirable behaviors and unite with the insufficient test of AI models.

Machine learning models also have a tall order to behave the way to do so, but also the researcher Javier Rando in AI.

“Answer, no research for almost 15 years, not, not, and it does not know how to do it,” Rando, Rando, “Rando, told CNBC.

However, there are some ways to evaluate risks in the EU red team. Experience covers the testing of individuals testing and artificial intelligence systems to detect and identify and identify and identify any potential damage or a common fashion operand in cyberecurity circles.

Schayne Longpre, AI and a researcher who led politics and Information Sxous InitiativeHe noted that the lack of people working in red teams.

AI beginnings are now using the second parties concluded to open the test to third parties, such as using the first party appraisers or normal users, journalists, researchers and ethical hackers. A paper published by Longpre and researchers.

“In the systems that people find the lawyers, medical doctors, actual scientists, which are actual scientists, which are, are defective, because the partner could not have enough experience or have sufficient experience,” he said.

Standardized ‘AI defect’ reports, these ‘defects’ in the AI ​​systems, these ‘flaws’ data spread and ways are some recommendations set on paper.

With this experience successfully accepted in other sectors such as software, “We already need it in AI.”

With management, politics and other means, marry this user-centered experience, it will better understand the risks created by AI tools and users.

The project is such an approach by combining moonshot, technical solutions with policy mechanisms. The project launched by the Singapore’s Infocomm Media Development Organization is a wide range of language model evaluation tool assessment tool designed with industrial players such as Moonshot, IBM and Boston based Computer robot.

Toolkit combines benchmarking, red team and trial bases. There is also an assessment mechanism that can be valid for AI’s beginners and moders, Anup Kumar, Aiu, EU information and EU customer engineering, CNBC information and EU information.

Assessment a continuous process He should also be done before and after the placement of the models, noting that the response was mixed in the tools, Kumar said.

“Many beginners took it as a platform Open source, and they began to use them. But I think you know, we can do more. “

Progresses, the project aims to include customization for use in the private industry and create a multilingual and multicultural team.

Essec Business School Statistics Professor, Asia-Pacific, technological companies said Hurts to release them from the correct assessment of their last AI models.

“When a pharmaceutical company prepares a new drug, he needs tests and a very serious proof that this is not useful and harmful before the government is approved,” he said.

AI Models must meet a collection of serious conditions before confirmed. A shift away from extensive AI tools to develop intended for more specific tasks, it will facilitate the easement to expect and control the concepts.

“LLS can do a lot, but not focused on tasks, which are quite specific,” he said. As a result, the number of impossible number of abuse is very great for the expectations to expect all the developers. “

So so much models determine those who specify those who are safe and secure Rando a study was dealt with.

Therefore, technological companies said “Better than they are protected.”