Despite recent concerns regarding its reliance on artificial intelligence systems, Meta is reportedly moving forward with plans to replace staff with AI tools, with the company planning to use LLMs to review the vast majority of its content and ads by the end of this year.
As reported by the Financial Times, Meta is accelerating its plans to implement AI models in place of humans for content review. Meta said 50% of these tasks are now being managed by AI tools, and it's looking to increase that to 90% by the end of 2026.
Which means that, soon, virtually all of all the content decisions related to posts and ads will be made by AI across Facebook, Instagram and its other platforms.
This comes after about 20,000 Instagram accounts were broken into by hackers in a recent exploit, in which the hackers tricked Meta’s AI support bot, which then gave them access to accounts they did not own.
Though “hackers” is probably not the best term to use here, because the exploit didn’t require any coding skills or system knowledge. The account thieves simply asked Meta’s AI support system to send account verification codes to a different email address.
So it’s a clumsy vulnerability, and one Meta said has been rectified. But the incident highlights a significant concern with AI tools, in that there will be vulnerabilities like this due to the very nature of their systems, which are designed to empower users to pose queries in natural language.
The challenge here is that because people can ask for actions in an almost infinite number of ways, it will be very difficult for Meta's models, or indeed any AI-powered system, to solidify its systems against misuse.
The same thing has happened with other AI models. When they’re asked in a specific way, even if certain query types are blocked, they will still undertake tasks for the user. And because there are so many ways to pose queries, it’s impossible to block all of them. That means that the more that Meta implements AI agents that are empowered to undertake sensitive tasks, the bigger the risk that they will be misused and tricked into giving users access.
Yet, Meta also needs to empower its AI tools with more tasks like this, in order to demonstrate their value for other organizations.
Meta is investing hundreds of billions, maybe even trillions of dollars, into AI development, in the hopes that AI will become the transformative technology of its time, able to undertake a wide range of tasks that, up until now, have only been possible through human labor.
The business benefit for Meta, then, is that it will eventually be able to sell these AI models to other organizations and help other companies reduce staff costs, and rely on AI to do the work instead. But if Meta can’t demonstrate that use case within its own business, that weakens this premise, which could impact the future earnings potential of its AI projects.
As such, Meta is keen to cut staff, and replace people with AI, as a practical example of the value of its AI systems.
But could Meta be moving too fast as a result, and opening up more avenues to potential exploits?
