The Information Technology Industry Council (ITI) recently released a new policy guide for AI. The several page guide specifically addresses foundation models, identifying its technological risks, and suggests potential solutions.
According to a recent press release, ITI released the guidelines which are aimed at managing risks associated with the use and development of “foundation models”, a branch of artificial intelligence. Foundation models, reports the release, are behind many of today's common uses of artificial intelligence, including searching, editing, and chat bot functionality. Although they are very useful and have great potential benefits, ITI also warns of risks associated with the technology, and its new guidelines will reportedly help mitigate them.
As part of the report, ITI points to potential areas of risk associated with this technology, breaking it into three main categories: misinformation, privacy, and security. In terms of misinformation, foundation models are susceptible to corruption because they are only as accurate as those databases from which they are trained, and large data bases often contain misinformation and bias. AI also poses privacy risks, according to the guide, because foundation models could potentially use personal information to “infer even more.” As for security, again foundation models are vulnerable at their source – data sets – which can be “poisoned” or stolen from, actions which could potentially affect any AI systems that use the model.
ITI gave several recommendations for policy makers to address these risks, according to the press release. Among the options, they encourage them to promote transparency among the AI chain, stakeholders and consumers alike, to “invest in measured science” for this technology, and to avoid attempts at extreme overkill restrictions.
ITI’s policy guide also explains the importance of foundation models within AI, stating that they are the focus of “80% of AI research.” Beyond their generally understood capabilities, like summarizing and following instruction, foundation models can also be used in other practical scenarios. For instance, as stated in the guide, a foundation model can scan a medical image and point out abnormalities, it can look for flaws in building structures, and it can analyze and streamline business strategies.
