Documents and sources: Meta sees a path to recovery using AI tools that boost views of Reels, improve ad targeting even with less available user data, and more
which Meta processed without a valid legal basis in the EU. And now it appears its unlawful processing is being leveraged to develop newly invasive ad targeting AIs. When will the law catch up with this? https://twitter.com/... https://twitter.com/... Dan Treasure / @codetreasurer : @powerbottomdad1 If we've learned anything from this it's that salt looks terrible on everyone, even the especially talented like LeCun Yann LeCun / @ylecun : @DiogoSnows ... FAIR is already open. We just open source stuff so that *anyone* can build on top of our technology. @value_invest12 : The death of $META has been greatly exaggerated. It was the peak bullshit act of the 2022 growth stock massacre. https://twitter.com/... Yann LeCun / @ylecun : I came up with the example above during my interview with Alex Kantrowitz for his podcast. ChatGPT's answer sounds good, but it's exactly backwards. Funny thing is that Alex read the response quickly to me during the interview and didn't immediately realize it was wrong! 2/ @bobkohn : [Image of ChatGPT answering the question:"Can you draw a diagonal line in a triangle?"] @realkunalashah : They were always gonna have the best way to address ATT vs the other platforms apart from Amazon. I think Meta always did have the ability to mitigate most of the damage on a relative basis even if absolute targeting was over https://twitter.com/... @attention_by : @powerbottomdad1 Meta has mostly been a charity, but between google and deepmind id be willing to bet you will see some properly insane shit soon Yann LeCun / @ylecun : @bobkohn Yes, questions requiring elementary geometric intuition is another failure mode. @pandurevich : @powerbottomdad1 And they geolocked it, boo https://twitter.com/... Erik Bugge / @buggeerik : FTC managed to order deletion of models trained with illegally obtained data, so why not @EU_EDPB? https://twitter.com/... Yann LeCun / @ylecun : @DataChaz ... You are making my point. Ask yourself why Galactica was pilloried and crucified while chatGPT (which has similar flaws) was welcome as the second coming of the Messiah? Jpeg Jimbo / @based_coded : @powerbottomdad1 lmao didnt even realize meta released a chatbot @diogosnows : @ylecun ... Ask Meta to spin-off a semi-independent section of FAIR and release under a non-profit setup with exclusive deal to Meta (profit) That's how Microsoft is working around it. Maybe name it OpenAIR 😉 @powerbottomdad1 : long openai / microsoft until the heat death of the universe GPT to the fuckin moon @sonicfields : @powerbottomdad1 Yeah that meta chatbot sucks. It's going down a different route than how chatGPT is built though. But lambda is kinda good, wish that big G would release it @powerbottomdad1 : yann lecun yesterday: open ai is simply just not ahead of google and meta. they can easily make a product like chatgpt a conversation with metas new chatbot today: [Screenshots of answers from Meta's chatbot] @datachaz : 🚨 Correcting @ylecun here: @Meta/@facebook *did* release a #ChatGPT-like thing, the highly-anticipated #Galactica, but it was withdrawn just three days after its release, following a deluge of trolling and accusations of bias! The internet remembers: https://techunwrapped.com/... https://twitter.com/... Shira Ovide / @shiraovide : I've said it before: Photos of powerful people with silly name tags at Sun Valley will haunt you forever. https://www.wsj.com/... https://twitter.com/... @borrowed_ideas : Some interesting quotes from WSJ piece on $META: “The damage (ATT related) had diminished from the more-than 8% hit to revenue early in the year to just 2.5%, and likely would disappear entirely in the fourth quarter” Laura Landry Forman / @doubleljsquared : How Meta (formerly Facebook) is using AI to circumvent Apple's stranglehold on external user tracking toward more organic data collection. Wed's 4Q22 print could show some early payoff. Great piece by @JeffHorwitz @sal19 @WSJ https://www.wsj.com/... via @WSJ Alex Morris / @tsoh_investing : “At the talk, Mr. Schultz said Meta had already absorbed the worst from Apple's tracking changes. The damage had diminished from the more-than 8% hit to revenue early in the year to just 2.5%, and likely would disappear entirely in the fourth quarter.” https://www.wsj.com/... Kelsey Hightower / @kelseyhightower : Educated skeptics are so important. Yann highlights a limitation of current LLMs and tools like ChatGPT. Some might consider this as being a hater, but I consider this as the only way to understand how this stuff actually works. https://twitter.com/... Jason Kint / @jason_kint : Now let me tell you about “Project Lighthouse.” 📞 me. https://twitter.com/... Thanks: @sal19 See also Mediagazer
Context & Ripple Effects
Meta's recovery story has been building all year: after Apple's ATT framework stripped away mobile ad signals, the company spent 2023 rebuilding its ad engine around AI, and by mid-year the numbers validated it — a Q2 beat with revenue up 11% YoY sent META up double digits. The WSJ reporting now adds the mechanism behind that arc: AI tools lifting Reels views and restoring targeting accuracy even with less available user data, while the ATT-related revenue hit reportedly shrank from over 8% toward 2.5%.
The same documents carry a legal sting: they allege Meta processed EU user data without a valid legal basis, and that this processing is now feeding newly invasive ad-targeting models. That tension runs through Meta's whole AI posture — the same company that pulled Galactica within days over toxic output has been deliberately cautious on chatbots while pushing hard on the monetization side of AI.
First-order effects
- Advertisers regain effective targeting on Facebook and Instagram even with diminished signal, because Meta's AI models reconstruct audience inference from first-party content and engagement rather than cross-app identifiers.
- Meta's P&L directly improves: as the ATT-related revenue drag fades from over 8% toward roughly 2.5%, the AI-driven Reels and targeting gains convert into the revenue growth investors rewarded in Q2.
Second-order effects
- Apple's ATT framework loses leverage as a competitive moat: if platform-scale AI can recover most of the lost signal from owned data, the privacy feature stops translating into durable ad-market disadvantage for Meta against Apple's ecosystem.
- EU regulators face a sharper enforcement question — documents alleging unlawful processing as the training substrate for ad-targeting AI turn a past GDPR violation claim into an argument for halting or unwinding the models built on it.
Third-order effects
- If the pattern holds, large platforms' proprietary content becomes the compensating asset for signal loss industry-wide — every post, like, and watch minute doubles as inference input for ad models, entrenching scale advantages over smaller ad networks.
- Privacy law is being forced to catch up with model training rather than data collection: regulators will increasingly have to decide whether consent frameworks written for tracking apply to AI systems trained on already-collected behavior.
The trend: Consumer-internet platforms are rebuilding ad economics on AI trained from their own behavioral data, turning privacy-signal loss into a moat while regulation trails the modeling layer.