User Privacy: The Biggest Concern with AI Assistants
As the use of generative AI continues to rise, major technology companies are grappling with the challenge of finding the perfect application for this technology. Initially, online search engines were developed with mixed results. Now, the focus has shifted to AI assistants. Recently, OpenAI, Meta, and Google have launched new features for their AI chatbots, transforming them into personal assistants.
OpenAI’s ChatGPT, for instance, now allows users to have conversations with the chatbot as if they were making a call. It can provide instant answers to questions using a synthetic voice that sounds remarkably realistic. Additionally, ChatGPT users can now browse the internet and access up-to-date information, expanding their knowledge beyond the limitations of 2021.
Google’s Bard, on the other hand, is integrated with various Google services such as Gmail, Docs, YouTube, and Maps. Users can ask Bard questions about their own content, including searching through emails or organizing their calendar. Bard also has the ability to retrieve information from Google Search instantly.
Meta has also joined the AI assistant race by introducing AI capabilities to WhatsApp and Instagram. Users can now ask questions or request the creation of images with a simple phrase. Furthermore, Meta has developed celebrity avatars through artificial intelligence, allowing users to «talk» to their favorite celebrities.
AI Assistants: A Gateway to Scams
Despite the advancements in AI technology, experts warn that AI assistants pose significant risks, particularly in terms of security and privacy. The technology is flawed, often inventing or «hallucinating» information. By granting AI models access to sensitive data like emails, calendars, and private messages, technology companies are exposing millions of users to scams, phishing, and large-scale hacks.
With AI assistants now having access to personal information and the ability to browse the web, they are vulnerable to a type of attack known as fast indirect injection. In this attack, a third party alters a website by adding hidden text to manipulate the AI’s behavior. Attackers can lure users to these websites through social media or email, enabling them to extract private information such as credit card details.
While OpenAI, Google, and Meta acknowledge the dangers, they have yet to find a foolproof solution. Google, for example, is actively researching ways to identify and filter attack attempts using systems like spam filters. They also encourage users to report inaccuracies and «hallucinations» in their assistant’s responses. However, experts caution that relying on users to detect errors may not be sufficient, as people tend to trust computer-generated answers too much.
As technology companies continue to push AI assistants, it seems they accept the risks as inevitable. However, ordinary users may hesitate to adopt technologies that are prone to unpredictable failures. It is only a matter of time before hackers exploit these vulnerabilities. For now, many individuals may choose to wait for advancements in technology and avoid AI assistants snooping through their personal information.