Meta looks to AI to review harmful content in cost-cutting drive

{"text":[[{"start":9.6,"text":"Meta is racing to replace human moderation with generative artificial intelligence, as it undergoes a broader cost-cutting drive to offset chief executive Mark Zuckerberg’s vast spending on AI. "}],[{"start":23.4,"text":"The $1.4tn tech group has accelerated plans to switch to using large language models to review content and advertising across its platforms, according to four people familiar with the matter, which could help it save billions of dollars annually. "}],[{"start":38,"text":"Already, Meta has replaced about 50 per cent of human review requests with LLMs this year, several of the people said. The company is aiming to reduce that figure even further by the end of the year, the people added, potentially by more than 90 per cent for certain types of content. "}],[{"start":55.1,"text":"Meta has long relied on a mixture of automated systems and human reviewers, including third-party contractors, to assess whether a post or an advert breaches its rules. Appeals by users have typically been handled by human reviewers. "}],[{"start":68.7,"text":"Meta said the shift towards AI moderation was to use a rapidly evolving technology more effectively, rather than for cost savings. Since March, it said, its initial tests have shown that on average, LLMs make 13 per cent fewer mistakes than humans when enforcing against violating content, while finding 10 per cent more actual violations."}],[{"start":89.15,"text":"“The point of this work is to improve our enforcement efforts, and we’re deploying these more advanced AI systems once we’re sure they’re consistently performing better than our current methods of content enforcement,” it added. "}],[{"start":null,"text":"

"}],[{"start":101.5,"text":"The AI moderation push comes as Zuckerberg pours billions of dollars into the talent and infrastructure needed to develop what he dubs “personal superintelligence”, promising hyper-personalised AI products and agents for users."}],[{"start":115.4,"text":"Meta is also attempting to reduce operating costs by using AI to automate internal processes such as coding, several people said. The group has also undergone several rounds of lay-offs and restructurings, severely hitting staff morale."}],[{"start":130.1,"text":"Meta has been using Google’s Gemini large language model for most of its moderation and customer support, but staff have recently been told to switch to Meta’s new foundational model, Muse Spark, the people said. "}],[{"start":142,"text":"Regular AI moderation, which typically relies on traditional machine learning classifiers to flag rule-breaking material, can struggle with nuances such as satire or evolving language. "}],[{"start":153.2,"text":"Meta has argued that rapidly advancing LLMs, trained on large amounts of text and data, have the potential to improve efficiency, accuracy and the speed at which violations are responded to, while also increasing the coverage of more languages. "}],[{"start":167.45,"text":"The company has also previously come under fire for exposing human reviewers to mental health risks, given some of the graphic content they are required to look at. "}],[{"start":176.79999999999998,"text":"In mid-March, Meta said that early tests had been “promising” and that it would be deploying advanced AI moderation “over the next few years”. "}],[{"start":186.99999999999997,"text":"Nevertheless, some staff warn that the rollout of LLM-based moderation is now happening at an even faster rate and that the technology still makes errors and lacks sufficient oversight to be rolled out so swiftly. "}],[{"start":199.89999999999998,"text":"One Meta insider said LLMs moderating its content kept making errors, such as taking down or “shadow-banning” — preventing posts from being recommended in feeds — harmless material. Two people said Meta had not sufficiently established how to measure the technology’s performance, an assertion that the company denied. "}],[{"start":219.2,"text":"Another former employee said Meta had accelerated the shift after gathering substantial data from previous decisions by human reviewers of appeals to help it train and improve the LLMs. The person added that the company would still have a human assess a sample of responses for accuracy. "}],[{"start":237.79999999999998,"text":"Meta said that robust governance processes and continuous evaluation were in place, noting that human reviews also had their flaws."}],[{"start":247.04999999999998,"text":"The LLM moderation shift is already leading to lay-offs, particularly among contractors, several people said. Meta is also likely to cancel future contracts with third-party moderation groups, one person added. "}],[{"start":260.45,"text":"The shift may be particularly sensitive in advertising, where Meta is already under pressure over the volume of scam promotions on its platforms. "}],[{"start":268.7,"text":"Last year, Reuters reported that Meta had forecast internally that in 2024 it would earn about 10 per cent of its annual revenue — or $16bn — from adverts for scams and banned goods."}],[{"start":282.3,"text":"The company is also facing legal challenges over scam ads, including from California’s Santa Clara County and Australian mining magnate Andrew Forrest. "}],[{"start":292.3,"text":"Earlier this month, tens of thousands of accounts on Meta’s photo app Instagram were hijacked by hackers who exploited an AI-powered customer support chatbot, prompting a backlash from some users. "}],[{"start":312.25,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1782384046_1856.mp3"}