Meta AI

Meta AI is the AI research and product division of Meta Platforms (formerly Facebook). It is known primarily for the Llama series of open-weight models, and in 2024–2026 was the most prominent example of a Big Tech company publicly arguing for open-source AI before moving toward closed-weight frontier models. Through 2025 and 2026 Meta consolidated frontier work under Meta Superintelligence Labs (MSL), raised its AI capital-expenditure guidance to $125–145 billion for 2026, executed its largest layoff round of the year while reassigning thousands of staff to AI teams, and faced copyright, youth-harm, privacy, and EU regulatory actions.

Snapshot

Financials

Capex and capital raising

Compute and silicon

Workforce

Models

The series began with LLaMA, described in a paper submitted February 27, 2023 covering four dense models at 6.7B, 13.0B, 32.5B and 65.2B parameters, trained on 1.0–1.4T tokens drawn only from publicly available sources. Its stated results were that "LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B," and its design followed from an argument about inference rather than training cost: that "given a target level of performance, the preferred model is not the fastest to train but the fastest at inference," which led Meta to train past the compute-optimal point. Meta framed the release as democratizing access "since it can be run on a single GPU," and reported LLaMA-13B running on a single V100 at inference (LLaMA: Open and Efficient Foundation Language Models (Touvron et al., Meta AI, February 2023)). The training mixture included "the Books3 section of ThePile," the component later at issue in copyright litigation over training corpora; see AI Copyright.

Meta's open-weight Llama series has become a foundation for the open-source AI ecosystem and for Chinese labs doing fine-tuning and distillation. Open-weight Llama models are widely used as foundations by compute-poor labs and Chinese companies; per Epoch AI, Meta's training compute has grown but the open-model capability gap remains a "constant-ish lag" behind the frontier (Source: epochai.substack.com). Meta is not competitive at the frontier of agentic coding and has no Claude Code or Codex equivalent.

A model codenamed "Avocado" (never shipped publicly under that name) was pre-trained by end-December 2025; March 2026 NYT reporting indicated it underperformed leading models from Google, OpenAI, and Anthropic on internal reasoning, coding, and writing tests, and its release was delayed.

At an internal meeting whose contents became public on July 3, 2026, MSL chief Alexandr Wang said Meta's next model, codenamed "Watermelon", had matched OpenAI's GPT-5.5 on benchmarks while using an order of magnitude more compute than its predecessor Avocado; at the same meeting Zuckerberg acknowledged that the company's AI agent work had stalled for roughly four months, and Meta shares fell nearly 5% on the disclosures (Source: implicator.ai; finance.biggo.com). At the same July 2 meeting, Wang said a Muse Spark update with improved coding and agentic capabilities was coming soon, and Zuckerberg described Meta as on a "journey to superintelligence" while saying industry progress on AI agents had been slower than expected (Source: businessinsider.com). Reuters reported Zuckerberg's remarks as an acknowledgment that AI agent development was progressing slower than expected and that the benefits of the company's AI-focused restructuring — an estimated $145 billion bet — had not "come to fruition yet" (Source: reuters.com; aigovernancelead.substack.com).

On July 7, 2026 MSL launched Muse Image — its first media-generation model and second model release overall — free in the Meta AI app, Instagram Stories, and WhatsApp, with coding and search capabilities built into image generation, and previewed Muse Video with native audio support; CNBC reported Meta's internal tests place Muse Image behind OpenAI's GPT Image 2 but ahead of Google's Nano Banana 2 (Source: ai.meta.com; cnbc.com). See Muse Image and Muse Video (Meta Superintelligence Labs).

Muse Spark (April 8, 2026) was MSL's first frontier release and Meta's first closed-weight frontier model, a break from the Llama open-weight tradition. It shipped initially only in WhatsApp, the Meta AI app, and Meta smart glasses, with an API said to be coming "soon," and was published with a 158-page safety-and-preparedness report and a new Responsible Scaling Policy v2. It scored strongly on multimodal benchmarks (CharXiv Reasoning) and weaker on agentic coding. Alexandr Wang positioned it as "built for the 3 billion people already using our apps every day."

On July 9, 2026 MSL released Muse Spark 1.1, a multimodal agentic model with a 1-million-token managed context window that Wang called Meta's "strongest model for agentic and coding work yet," and opened the Meta Model API in public preview — the first time Meta has charged developers for model access, at $1.25/$4.25 per million input/output tokens with $20 in free credits per new account (Source: ai.meta.com; cnbc.com; bloomberg.com). Wang characterized the pricing as "very aggressive and attractive" against Anthropic and OpenAI, said an open-source Muse Spark variant is in development, and confirmed Watermelon as a more powerful model in training (Source: cnbc.com). The release came the same day as OpenAI's broad GPT-5.6 release and a day after SpaceXAI's Grok 4.5, placing three frontier releases within 48 hours. See Muse Spark (Meta Superintelligence Labs).

On August 10, 2026 MSL released Muse Glimmer, an open-weight agentic model distilled from Muse Spark and described as running on a Mac or PC with a single graphics card, which Meta said "performs strongly for its size class" on several benchmarks without naming them. No model card or evaluation report accompanied the release. See Muse Glimmer (Meta Superintelligence Labs).

Reception and culture analysis

Timothy B. Lee's analysis of the Muse Spark release (April 20, 2026, Understanding AI) frames the release as a return to relevance while raising a structural concern about Meta's culture.

Lee recounts the Llama 4 stumble (April 5, 2025), in which initial benchmark claims were overstated; LeCun later told the FT the "results were fudged a little bit," meaning Meta fine-tuned specific models to do well on prominent benchmarks and then released different models. Mowshowitz wrote, "I am placing Meta in that category of AI labs whose pronouncements about model capabilities are not to be trusted." Llama 4 Behemoth never shipped, and a year-long silence followed, with the Wang/MSL rebuild as the response.

On Muse Spark itself, Stratechery wrote it "isn't state of the art, but it's in the game, and overall a positive first impression from Meta Superintelligence Labs," and Mollick called the model "a bit... weird. Like some strange language & tone, a little loose with facts, etc." Lee notes that 78 of the 140 main pages of Meta's 158-page Muse Spark safety report contain a chart or table about benchmark results and most of the rest discuss benchmarks, compared with Anthropic's system cards, which carry extensive qualitative descriptions alongside benchmarks. Lee argues that a benchmark-centric approach can help Meta catch up to leaders, but that post-training, the step that gives a model its "personality" (creativity, resourcefulness, ethical grounding), is the step Meta's culture is least equipped for.

Lee predicts Meta will follow OpenAI into the GPT-4o sycophancy trap, citing Zuckerberg's January 2026 earnings call ("Personalized responses drive higher levels of engagement, and we expect to significantly advance the personalization of Meta AI this year") and an OpenAI employee anecdote in which an old OpenAI model stopped responding in Croatian after learning Croatian users were more likely to give negative feedback; the pattern is that training on engagement signals produces specific output distortions.

Lee also reports internal friction: per NYT December 2025 reporting, Wang and CPO Chris Cox disagreed about training-data strategy, with Cox wanting MSL to use Instagram and Facebook data to train Meta's frontier model to improve social-feed ranking and ad revenue, while Wang preferred to focus on catching up to OpenAI and Google on conventional capabilities; Muse Spark appears to reflect a partial compromise. On the coding-data gap, Lee observes that Anthropic and OpenAI have years of usage data from millions of coding agents while Meta has Facebook, WhatsApp, and Instagram data, which shapes Muse Spark's product fit toward consumer apps and multimodality rather than competing against frontier coding agents.

Open and closed model strategy

Meta's strategy has shifted between open and closed approaches. In July 2024 Zuckerberg published "Open Source AI is the Path Forward," arguing for privacy, safety through scrutiny, and community improvement. During 2025–2026 reports described a pivot toward closed models, including discussion of abandoning the open-weight Llama 4 Behemoth and releasing the "Avocado" model closed. As of April 2026 Meta said it planned to "eventually" offer open versions of next-generation models. Epoch AI published an article titled "Open Source AI was the Path Forward" (Source: epochai.substack.com). Meta's position is referenced in open-source AI coverage.

The pivot back toward weight release came on August 10, 2026, when Meta released Muse Glimmer — an open-weight agentic model distilled from Muse Spark and sized to run on a Mac or PC with a single graphics card — and said it plans to release the weights of Muse Spark 1.2, its most advanced model. The accompanying Zuckerberg essay, "The Future is for Everyone" (The Future is for Everyone (Zuckerberg, August 2026)), running roughly 6,400 words, argues that "the notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic," calls it "surprising" that discourse from many AI developers was "filled with doom," and makes a competitiveness case for loosening U.S. rules: "foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," so "U.S. policy must reduce this additional friction if we want American open source models to lead over time." Zuckerberg named data-use and distillation rules as the friction and rejected restrictions on access to foreign open-source models, consistent with his July position against blocking Chinese models. Announced with the release: a $1 billion fund for communities near Meta's data centres, and a commitment to give Meta's independent directors power to approve the safety criteria for model releases. Meta shares, down about 10% year to date, rose nearly 3% in premarket trading (Source: reuters.com; businessinsider.com).

The essay set out three further positions in the same release. Meta is "implementing a governance structure that gives our independent board of directors the power to approve the safety criteria for releasing models and reviewing whether each model release adheres to the criteria," which Zuckerberg encouraged other frontier labs to adopt and suggested could be extended to an industry-wide version; as of publication no implementation date, criteria or review procedure had been described (AI Governance (umbrella)). On pre-release review he proposed that labs supply the government with intermediate training checkpoints and technical staff rather than submit finished models, arguing against a fixed review timeline applied in all cases (AI Pre-Release Vetting). On distillation he argued for protecting "the principle that you can learn from anything you can observe," describing the ability of models to learn from other models as an important principle of the open-source ecosystem (Adversarial Distillation).

The essay also detailed Meta's data-centre community commitments. It describes a "community compact" of local benefits — high-paying local jobs, investment in schools and public services, holding energy prices down, and environmental measures — and announces a Future Is For Everyone Fund to support each community Meta operates in directly, without stating a figure in the text itself. It reports that teachers in Richland Parish, Louisiana received a $50,000 bonus in 2026 from increased tax revenue attributed to Meta's investment, and that the local superintendent said teachers were relocating there from across the country. It describes America's Workforce Academy, which Meta created to provide free training for skilled tradespeople with guaranteed high-paying jobs in data-centre areas; a practice of building Meta's own energy-generating infrastructure wherever it invests, in some cases supplying surplus low-cost energy back to communities; and a commitment to be water-positive by 2030, restoring more water than Meta uses in the watersheds where it operates, with a goal of restoring 200% of water used in areas of high water stress (Data Center Siting / AI Power Politics).

Zuckerberg restated the openness argument in the press on July 28, 2026, writing in The Wall Street Journal that "invention, not automation, will be the greatest contribution of superintelligence," that America's "advantage is decentralized and open innovation," and that closing models to keep China from gaining access "will only disadvantage the US and its allies." He separately told the Financial Times, in an interview published July 28, 2026, that the United States should not block Chinese models, calling a ban on cutting-edge Chinese AI not "an effective solution" and saying US companies should "systematically" identify bottlenecks in order to compete (Source: wsj.com; ft.com; reuters.com; businessinsider.com). The remarks followed the administration's July 28 import bans on new Chinese robots and power inverters (FCC Covered List addition — foreign-produced advanced robotic devices and connected power inverters (July 2026)) and Treasury Secretary Scott Bessent's warning that Chinese companies could face financial sanctions or placement on the Commerce Department's Entity List. The position runs opposite to the "Pacing the Frontier" statement released the same day, which Meta AI chief scientist Shengjia Zhao, Dawn Song, and Summer Yue signed in a personal capacity (Pacing the Frontier (statement from employees of frontier AI companies, July 2026)). See Open-Weight Frontier Models.

Strategy and leadership

Meta's public framing of its AI direction is set out in Personal Superintelligence (Zuckerberg) (July 30, 2025), which describes decentralized personal superintelligence delivered via glasses and devices, contrasted with what it characterizes as OpenAI's and Anthropic's "centralized automation" model, and offered as the public justification for $72B+ in 2025 AI capex, the creation of Superintelligence Labs, and the Alexandr Wang appointment. On July 23, 2026, Mark Zuckerberg launched a paid "AI optimism" advertising campaign positioning Meta against rivals' warnings about AI's societal risks (Source: axios.com). See AI Backlash.

Meta Superintelligence Labs (MSL) was established in June 2025, led by Alexandr Wang (former Scale AI CEO), hired as Chief AI Officer. On June 12, 2025 Meta invested $14.3B in Scale AI and recruited Wang. The buildout included a hiring campaign that recruited Ruoming Pang from Apple (reported $200M+) and others from OpenAI. Sam Altman responded with "Missionaries Will Beat Mercenaries" (WIRED, July 1, 2025).

CTO Andrew Bosworth was profiled by the WSJ on May 24, 2026 as the figure leading Meta's remake into an AI-first business; the piece named a contested program to record employees' keystrokes and mouse clicks to train AI agents, alongside planned layoffs (Source: wsj.com). On the external-relations side, Meta engaged Dina Powell McCormick (a former Trump first-administration official) as a political liaison, disclosed around May 18, 2026 and described by observers as an example of Big Tech's deployment of former government officials across major firms, ahead of the May 2026 Trump frontier-AI EO (Source: spyglass.org).

Key people

• Mark Zuckerberg (CEO): author of Personal Superintelligence (Zuckerberg)
• Yann LeCun (Chief Scientist): contrarian on existential risk
• Alexandr Wang (Chief AI Officer): leads Superintelligence Labs
• Dina Powell McCormick: political liaison
• Naomi Gleit: senior executive leading the Enterprise Solutions unit

Products and partnerships

Agentic AI and enterprise tools

The Information reported on May 5, 2026 that Meta is training an AI agent codenamed "Hatch," modeled on its OpenClaw, for internal testing by end-June 2026, with sandboxed environments simulating DoorDash, Etsy, Reddit, Yelp, and Outlook. Hatch was powered initially by Anthropic's Claude Opus 4.6 and Claude Sonnet 4.6 models, and at launch will be powered by Muse Spark (Source: theinformation.com). Per internal documents reported June 3, 2026, Meta is weighing charging up to $200 a month for the planned consumer version of Hatch, pricing that would match rivals' top subscription tiers such as OpenAI's and Anthropic's premium plans (Source: theinformation.com). The agentic push parallels Google's January 2026 Gemini Enterprise for Customer Experience and Amazon's Rufus (see Agentic AI, Muse Spark (Meta Superintelligence Labs), Agent Architecture Patterns).

Separately, Meta is building an agentic shopping tool for Instagram targeting a launch before Q4 2026, in which users tap on an item in a Reel or feed to learn more, navigate to an external webpage, and complete a purchase within Instagram; the tool is positioned against TikTok Shop.

Meta announced the global launch of the Meta Business Agent on June 3, 2026, making AI-powered customer-service and sales tools available across WhatsApp, Instagram, and Messenger. The agent fields inquiries, suggests catalog products, schedules appointments, vets sales leads, and completes transactions before handing off to a human at an owner-defined threshold. Meta said pilots in India, Mexico, and Brazil had drawn more than one million businesses; at the company's Conversations event in London, Zuckerberg said "as our models advance, your agent will take on more and eventually help you run your whole business" (Source: qz.com; see Agentic AI).

Meta is creating a new Enterprise Solutions unit under senior executive Naomi Gleit, with engineers and product managers to be embedded inside large corporate customers to push adoption of Meta's AI tools. Per The Information's May 27, 2026 reporting on an internal Gleit memo, the unit is part of Meta's pivot to monetize AI tooling outside its core advertising business and puts Meta in enterprise-customer-acquisition competition with Anthropic, OpenAI (whose Lehane-led enterprise push parallels this one), and AWS (Source: theinformation.com).

Meta repeatedly delayed releasing its newest AI model to developers; as of June 2, 2026 it had no planned date for the API that would let outside apps build on the model, and a Meta spokesman said on June 3 the company was testing the API with partners (Source: wsj.com). The Meta Model API ultimately opened in public preview on July 9, 2026 alongside Muse Spark 1.1 (see Models, above).

Meta's internal incubator AAI Labs is developing an OpenRouter-style routing service that cuts coding costs by steering tasks to lower-cost models, an effort that became public on July 21, 2026 (Source: theinformation.com). See Coding AI Market Map.

Meta is asking thousands of engineers to use its in-house AI coding agent, MetaCode, in their day-to-day work as a way to improve the coding capabilities of its own models, in an account that became public on August 5, 2026. The retrievable abstract names Maher Saba, vice president of Meta's Applied AI Engineering; the body is behind a paywall and the extent of the mandate could not be verified beyond "thousands of engineers" (Source: theinformation.com). See AI Coding Agents.

Consumer apps and subscriptions

Meta began rolling out paid subscription plans for Instagram, Facebook, and WhatsApp worldwide on May 27, 2026, priced at $3.99/month for Instagram Plus and Facebook Plus and $2.99/month for WhatsApp Plus, and disclosed it is testing AI subscription plans under a new "Meta One" brand at $7.99 and $19.99/month, pitched as new revenue to offset rising AI-buildout costs (Source: techcrunch.com). Meta is the largest hyperscaler so far to introduce a consumer paid AI tier as a financing strategy rather than continuing pure ad-monetized scale-out (see bubble debate).

Meta launched Forum, a standalone iOS app for Facebook Groups, on May 22, 2026. Forum supports nickname posting, a user-tunable topic algorithm, and AI features that summarize responses and assist group admins, positioning Groups-as-community against Reddit and Discord; the AI summarization and admin-assist tools are consistent with Meta's push to embed Muse Spark-class assistance into its consumer surfaces (Source: thurrott.com).

Reporting on June 24, 2026 indicated Meta is building a dedicated prediction-markets app, moving the company into the prediction-markets category alongside platforms such as Polymarket and Kalshi (Source: readthepeak.com).

Meta began rolling out a standalone AI companion app for Facebook creators on June 24, 2026, embedding its recently launched AI creator assistant and effectively reviving and replacing the company's Creator Studio; the app was initially in testing with a select group of creators (Source: techcrunch.com).

Meta quietly launched Pocket, an app for generating and sharing "vibe-coded" mini games via AI prompts, on June 29, 2026 per Appfigures data, with the launch surfacing publicly on July 2; the app grew out of Meta's early-2026 acquisition of the Gizmo team (Source: techcrunch.com).

Hardware and wearables

An internal memo by Meta wearables VP Alex Himel, disclosed May 29, 2026, outlines a hardware roadmap that includes testing an AI pendant within the next year, an expansion of the AI-glasses lineup, and a business-focused "Wearables for Work" service, with a target of 10 million wearable devices sold in 2H 2026. The push is framed against Reality Labs' Q1 2026 loss of $4.03 billion on $402 million in revenue and builds on Meta's acquisition of pendant maker Limitless (Source: theinformation.com). The AI pendant places Meta in the always-on wearable-AI form-factor category alongside the Humane Ai Pin and Rabbit r1, with distribution and a captive AI assistant the standalone startups lacked, and "Wearables for Work" is the hardware counterpart to the Enterprise Solutions unit. The minors-default-off AI-companion provisions in NY's Safe By Design Act and Connecticut SB 5 form a regulatory backdrop for consumer always-on AI wearables.

Executing on the lineup expansion, Meta and EssilorLuxottica announced a lower-cost line of AI smart glasses on June 23, 2026, starting at $299 — roughly $80 cheaper than prior models — with Himel saying privacy improvements are forthcoming (Source: marketscreener.com). The lower entry price targets mainstream adoption of the camera-and-assistant glasses category, where the unresolved privacy questions around always-on recording remain the principal consumer and regulatory friction.

On July 7, 2026, Meta announced a software update that will disable its smart glasses' camera when tampering with the privacy LED is detected, responding to modders who drill out the recording light. New York State planned to ban camera glasses from all courthouses later in July, following similar Philadelphia court restrictions (Source: theverge.com). See Surveillance Technology. Meta's push to bring AI-enabled, camera-equipped smart glasses to market in volume drew organized objections from privacy advocates, in reporting dated July 13, 2026 (Source: wsj.com). Business Insider reporting by Katie Notopoulos, highlighted July 23, 2026, documented viral videos of Meta Ray-Ban wearers filming strangers without their awareness, with the glasses' recording light failing to alert subjects (Source: platformer.news).

In brain-computer interface research, Meta announced Brain2Qwerty v2 on June 30, 2026, a noninvasive system that decodes typed sentences from magnetoencephalography (MEG) brain recordings; the work was published in Nature Neuroscience, and Meta open-sourced the code and data (Source: nature.com; roadtovr.com). Meta's account put character-level decoding at up to 80%; independent coverage on June 30 put word-level accuracy at 61% and noted the system still requires the user to type and cannot yet support real-time communication (Source: ai.meta.com; theregister.com).

Compute and infrastructure

Meta Compute cloud business

On July 1, 2026 Meta was disclosed to be developing plans for a cloud infrastructure business, reported to be named Meta Compute, that would sell access to raw AI computing capacity and to hosted models — including the closed-weight Muse Spark — placing Meta in competition with Amazon Web Services, Microsoft Azure, and Google Cloud (Sources: techcrunch.com; bloomberg.com). The effort is led by infrastructure head Santosh Janardhan, Superintelligence Labs leader Daniel Gross, and president Dina Powell McCormick, and follows Meta's reported $182.9 billion commitment to AI infrastructure; TechCrunch compared it to SpaceX's monetization of excess xAI capacity through leases to Anthropic, Google, and Reflection AI (Source: techcrunch.com). Meta shares rose nearly 9% on the disclosure while neocloud providers CoreWeave and Nebius each fell about 12%; CNBC noted Meta plans up to $145 billion in capital expenditure this year and cited SpaceX's existing sales of excess capacity to Anthropic ($1.25 billion per month) and Google ($920 million per month) as the precedent (Source: cnbc.com). A first anchor customer surfaced in mid-July: Anthropic entered early talks to lease Meta compute in an arrangement reported at up to $10 billion over two years, paid monthly with early-exit rights for both sides, with Anthropic having initiated the proposal in June (Source: cnbc.com; nytimes.com).

SemiAnalysis calculated on July 2, 2026 that Meta had contracted more than 5 GW of cloud and colocation capacity in the first half of 2026, reported that Meta was in final talks with Anthropic for private instances of Claude (a Bedrock-style arrangement it anticipated at roughly $10 billion), and projected that Meta would strike SpaceX-type on-demand compute deals (Source: newsletter.semianalysis.com). The talks became public on July 17, 2026, when the New York Times reported Meta in negotiations to rent computing power from its data centers to Anthropic in a deal that could be worth roughly $10 billion over two years (Source: nytimes.com). Antitrust writer Matt Stoller argued on July 4, 2026 that the pivot to selling AI compute is likely to fail (Source: thebignewsletter.com). Ben Thompson, in a July 7, 2026 Stratechery essay framed as a script for what Mark Zuckerberg should tell investors, argued Meta should rent out spare compute on the spot market as a hurdle rate, lean into AI-driven expansion of ad inventory, and stop competing with OpenAI and Anthropic on productivity chatbots (Source: stratechery.com). Capacity pressure also runs in the other direction: the Financial Times reported on July 4, 2026 that Google moved to cap Meta's use of its Gemini models as computing power became the industry's scarcest commodity (Source: ft.com).

Meta has pursued multiple sources of compute beyond Nvidia. On June 18, 2026 it secured agreements to buy roughly 1.6 gigawatts of computing capacity from data-center developer Crusoe across two sites in Childress, Texas and Warrenton, Missouri, extending its pledge to invest about $600 billion in US AI infrastructure over the next three years (Source: reuters.com). On June 5, 2026 it signed a multibillion-dollar deal to rent Google AI chips (TPUs) for its data centers beginning in 2027 and was reported to be weighing a large equity raise; Meta shares fell about 7% on the news. The deal is a diversification away from Nvidia at the training and inference silicon layer and a validation of Google's TPU-as-a-service push, sitting alongside Meta's existing Broadcom MTIA and Amazon Graviton5 arrangements (Sources: finance.yahoo.com; cnbc.com; see AI Bubble Debate).

On April 29, 2026 Meta said it is rolling out over 1 GW of custom AI silicon developed with AMD and Broadcom, plus expanded AMD capacity alongside new Nvidia systems, signaling a multi-chip cost structure rather than full Nvidia dependence (Source: x.com). Meta's partnership with Broadcom for development of multiple generations of MTIA chips surfaced publicly on April 27, 2026, roughly a month after Meta announced its roadmap of the next four AI training and inference generations (Source: wow.datacenterdynamics.com). On April 24, 2026 Meta and Amazon announced a multibillion-dollar, multiyear deal for Meta to rent hundreds of thousands of Amazon's Graviton5 chips for AI inference, a customer win for AWS's custom-silicon business following agreements with Anthropic and OpenAI (Source: bloomberg.com).

A critical assessment of the custom-silicon effort came in a July 21, 2026 SemiAnalysis analysis of Meta's infrastructure organization, which reported that roughly 30% of engineers from the $2.5 billion-plus Rivos acquisition were laid off, that the Olympus chip was cancelled with successor Phoebe not taping out until 2028, and that Meta's custom "Ariel" GB200 rack carried 14% higher total cost of ownership than the standard NVL72 configuration it has since reverted to (Source: newsletter.semianalysis.com).

Meta plans to start manufacturing its in-house "Iris" AI chip in September 2026 — designed with Broadcom, fabricated by TSMC, and bug-tested in six weeks with no major issues — according to an internal memo that became public July 9, 2026. The memo shows 7 gigawatts of computing deployed in 2026 and 14 gigawatts targeted for 2027, up to $145 billion in 2026 AI-infrastructure spending, and supply agreements with Samsung (memory), Sandisk (flash storage), and Sumitomo Electric (fiber optics) (Source: reuters.com).

On July 28, 2026 Meta and BlackRock announced a venture to develop and own a data-center campus in El Paso, Texas, formalizing financing that had been reported on July 20. Funds managed by BlackRock will own 80% of the venture and Meta 20%; Meta will receive a one-time distribution of approximately $1 billion to align the stakes, and part of BlackRock's investment will be funded by a $12.5 billion debt financing. Reuters put total development cost at about $14 billion. The campus is planned for 1 gigawatt of compute capacity, represents an investment of over $10 billion from Meta, supports more than 4,000 construction jobs at peak and 300 operational jobs, and is expected to begin bringing capacity online in 2028. Meta will lease the entire campus and be its initial sole occupant (Source: investor.atmeta.com; reuters.com). See Data Center Siting / AI Power Politics.

On July 13, 2026, Meta committed an additional $40 billion to its data-center campus in Richland Parish, Louisiana — its largest, named Hyperion — bringing the announced expansion to $50 billion for at least 5 gigawatts of computing power and total expected investment in the site beyond $250 billion (Source: bloomberg.com). Meta said the campus will expand to 5 GW at a cost above $50 billion — up from the $27 billion disclosed in October 2025 under its Blue Owl Capital joint venture — with 2 GW targeted by 2030 and full completion around 2032 (Source: cnbc.com). The expansion is backed by an Entergy arrangement covering seven gas plants and roughly 240 miles of transmission (Source: cnbc.com).

Staffing for the cloud push followed: Meta's plan to hire Dave Brown, Amazon Web Services' compute chief, became public on July 15, 2026, reported as part of its weighing of a cloud business (Source: wsj.com). AWS announced Brown's exit on July 15 and his Meta move was reported July 17; GeekWire reported that Brown, who led EC2 and AWS's compute and AI organization, will work on the data-center buildout under Meta Compute — which it described as an infrastructure initiative Mark Zuckerberg launched in January 2026 — as Meta expects $125–145 billion in capital expenditures this year (Source: geekwire.com).

On July 8, 2026, Meta announced its first Canadian data center: a 1-gigawatt, AI-optimized facility in Sturgeon County, Alberta, reported at roughly $9 billion (C$13 billion) — its 33rd data center — expected to take two to three years to build, announced amid the up-to-$145 billion capital-expenditure forecast and the Meta Compute cloud plans (Sources: cnbc.com; reuters.com).

On the power side, Meta announced on April 27, 2026 a deal with Overview Energy for up to 1 GW of space-based solar power, with a demonstration planned for 2028, extending Meta's data-center power-procurement strategy into experimental, low-TRL energy categories; it is the first reported AI-data-center commitment to space-based solar (Source: bloomberg.com). At the other end of the technology spectrum, Meta and El Paso Electric are installing 813 gas generators as a 225-megawatt bridge microgrid at Meta's McCloud project in El Paso, Texas, with Meta paying all costs for two to five years, per July 18, 2026 reporting (Source: theinformation.com).

Meta's data-center operations also drew a water-quality enforcement action: on July 10, 2026, Cheyenne, Wyoming's Board of Public Utilities said it traced a Cupriavidus gilardii contamination of its reclaimed-water system (found in February 2026) to a contractor on Meta's 715,000-square-foot data center there, terminated Meta's discharge privileges, and cited "significant non-compliance" with federal pretreatment rules (Source: fortune.com). Meta, whose independent tests found no bacteria, is appealing the discharge bar, per July 18, 2026 reporting (Source: theinformation.com). See AI Environmental Impact.

Funding and capital expenditure

Meta reported Q1 2026 revenue of $56.3B (+33% YoY) and net income of $26.8B on April 30, 2026, with shares falling roughly 9% as Zuckerberg pledged higher AI capex without a near-term monetization roadmap (Source: cnbc.com). Full-year 2026 capex guidance was set at $125–145B, with 1 GW of custom silicon built in collaboration with Broadcom, AMD, and Nvidia; the stock fell 8.5% on April 30, its steepest one-day drop since October 2025, on the capex guidance. Meta separately sold $25B of investment-grade bonds on April 30, 2026 after attracting $96B in orders, having previously sold $30B in October 2025 (Sources: cnbc.com; bloomberg.com).

Debt financing extended off the balance sheet in 2026. BlackRock's leading of a $12 billion-plus debt sale for Meta's El Paso data center was disclosed on July 21, 2026, as reporting placed the off-balance-sheet debt of the five largest US technology companies at about $1.65 trillion — roughly eightfold growth since 2022 on what Nikkei described as opaque AI funding structures (Source: wsj.com; asia.nikkei.com). See Private Credit & AI Infrastructure.

Second-quarter results reported on July 29, 2026 showed revenue up 28% to $60.8 billion and 3.6 billion daily active people, up 3% year over year, against free cash flow of $784 million — down 91% from $8.55 billion a year earlier. Earnings per share of $6.18 missed the $7.22 LSEG consensus and shares fell 10% in extended trading. Meta raised the lower end of its 2026 capital-expenditure outlook to $130 billion from $125 billion while leaving the $145 billion ceiling unchanged. Operating income fell 8%; Chief Financial Officer Susan Li said it would have risen 9% without legal charges and severance, and warned that youth-related trials "may ultimately result in a material loss" (Source: reuters.com).

The capex guidance was first lifted to $125–145B in the April 29 earnings release, sending the stock down about 6.6% after hours (Source: bloomberg.com). Following the June 5 Google TPU deal, Meta was reported to be weighing a large equity raise (Source: cnbc.com).

Safety approach

Meta maintains a Frontier AI Framework with cyber thresholds including automated end-to-end compromise of corporate environments, zero-day discovery, and scaled fraud and scams (Source: Managing Advanced Cyber Risks in Frontier AI Frameworks), and a CyberSecEval4 benchmark suite for evaluating AI cyber capabilities. Muse Spark was published with a 158-page safety-and-preparedness report and Responsible Scaling Policy v2 (see Models, above).

In the FLI AI Safety Index (Winter 2025), Meta ranked D (1.10), among the lowest-scoring frontier labs; it did not submit a company survey and was evaluated from public materials only (FLI AI Safety Index Winter 2025).

Security incidents

On June 1, 2026, attackers seized control of high-profile Instagram accounts by socially engineering Meta's own AI support chatbot into granting account access, briefly defacing the accounts of the Obama White House and the Chief Master Sergeant of the U.S. Space Force with pro-Iranian messaging. Rather than exploiting a software bug, the attackers manipulated the support bot's conversational logic into approving access it should not have granted (Sources: krebsonsecurity.com; techcrunch.com). The episode is an instance of the prompt-injection and confused-deputy failure mode applied to a customer-support surface, in which an AI agent held privileges higher than the untrusted party it was conversing with. On June 2, 2026 Meta said it had fixed the flaw, describing the attack chain: attackers faked a location via VPN, then asked the support bot to attach a new email and issue a password-reset code, turning a benign-looking support request into an account-takeover. Spokesperson Andy Stone called claims that sitting world leaders' accounts were compromised "totally false," and security researchers warned the episode shows the danger of giving AI customer-support agents "too much authority and too little verification" (Source: bbc.com).

A second case involved MyClaw, an internal AI agent Meta employees use for accessing work files, summarizing posts, and getting technical advice. Use of MyClaw triggered a major security alert at Meta after an employee followed incorrect advice from the agent, leading to sensitive company and user data being exposed to employees who were not authorized to access it (Source: theinformation.com). The May 1 CISA agentic-AI guide lists this failure mode (privilege creep) as a leading agentic-AI risk.

Acquisitions

Meta acquired Manus (Butterfly Effect), a Chinese agentic-AI startup described as a standalone AI harness, in December 2025, following CFIUS review of Benchmark Capital's $75M stake (Source: A Guide to Which AI to Use in the Agentic Era; see AI Acquihires). The acquisition (described in subsequent reporting as a $2 billion deal) drew sustained Chinese regulatory opposition. China's National Development and Reform Commission (NDRC) blocked the acquisition on April 27, 2026 and ordered the parties to terminate the deal, even though Manus had relocated to Singapore beforehand; President Xi's National Security Commission characterized the acquisition as a "conspiratorial" attempt to hollow out China's tech base, and the action extended Beijing's reach to Chinese-founded companies that had relocated abroad (Sources: theinformation.com; ft.com). The NDRC found the deal violated PRC anti-foreign-control rules in AI (Source: washingtonpost.com).

As of April 28, 2026 Meta was preparing for a possible reversal of the acquisition even though it had already begun integrating Manus's technology, with Benchmark and other early investors having already received their proceeds, complicating any unwind (Source: wsj.com). By May 27, 2026, per FT reporting summarized on the Manus / Butterfly Effect page, the two co-founders of Manus, whose December 2025 sale to Meta was then described as "officially dead" after Beijing's April 27 block, were reportedly exploring a $1 billion external raise to unwind the Meta deal; China separately barred the co-founders from leaving the country pending the investigation (see Manus (Butterfly Effect) for the underlying CFIUS/NDRC sequence). The block also produced wider fallout: Beijing ordered Chinese tech companies including ByteDance and Moonshot AI to reject U.S. capital without state approval (Source: bloomberg.com). Meta severed Manus's access to its data after Chinese authorities ordered the buyout unwound, in reporting dated June 11, 2026, a further step in disentangling the integration begun before the block (Source: theinformation.com; bloomberg.com). On July 10, 2026 the FT reported that Manus investors and management were discussing unwinding the buyout at the same $2 billion valuation, with Tencent in talks to become the company's largest investor (Source: ft.com).

On May 1, 2026 Meta announced it had acquired Assured Robot Intelligence (ARI), a startup developing AI models for humanoid robots; the ARI team, led by Lerrel Pinto and Xiaolong Wang, joined Meta Superintelligence Labs. The acquisition represents Meta's commitment to embodied AI as a strategic axis alongside glasses and apps, and pairs with Boston Dynamics, Figure, and Clone in the humanoid-robotics race (Source: bloomberg.com; see Embodied AI vs AGI (China's Race Redirect), Lerrel Pinto, Clone Robotics).

Workforce and labor

Restructuring and layoffs

On April 23, 2026 Meta announced a restructuring of 8,000 cuts and 6,000 cancelled open roles (Source: washingtonpost.com). On May 20, 2026 spokesperson Erica Sackin confirmed that roughly 8,000 employees (~10% of the workforce) were notified their jobs had been eliminated while another 7,000 staff were being reassigned to four new AI-focused teams, matching the April announcement and adding the previously undisclosed AI-team reassignment (Sources: npr.org; washingtonpost.com). The May 20 cuts constitute Meta's largest layoff round of 2026; in a memo to staff, Zuckerberg framed the restructuring around AI strategy, writing "AI is the most consequential technology of our lifetimes" and "success isn't a given." Teams working on AI infrastructure, foundation models, and AI monetization were shielded from the cuts, and Meta indicated no further company-wide layoffs are expected this year. The round followed Cisco's ~4,000 layoffs the prior week, part of the wider 2026 Big Tech labor-restructuring pattern as AI capex reshapes cost structures, alongside Microsoft's voluntary-retirement program (Sources: cnbc.com; nytimes.com; see AI Labor Disruption).

On May 26, 2026, Meta filed a Washington State WARN notice eliminating 1,395 positions across its Seattle-region offices, about 20% of its 7,000-person Seattle-region workforce, as part of the broader 8,000-role companywide AI restructuring tracked in AI Labor Disruption. Meta spokesperson Tracy Clayton said the cuts include "layoffs, open role closures, and moving thousands of employees to business critical priorities," with cuts taking effect July 22, 2026, while the company is still planning up to $145 billion in 2026 capex (Source: geekwire.com). The Seattle-region notice is the largest single-state cut disclosed in the May 20 round and follows the reallocation-rather-than-pure-elimination pattern noted on the labor-disruption page. The breakdown:

On June 8, 2026, following the ~8,000-employee layoff round, Meta launched a free, five-week "Workforce Academy" training program that guarantees a job building data centers on completion, pairing the AI-driven workforce reduction with a data-center-construction hiring channel (Source: wsj.com). See AI Labor Disruption.

On July 14, 2026, 26 former employees sued Meta alleging that the AI tools used to select workers for the May 2026 layoffs disproportionately targeted employees with disabilities and those on medical or parental leave; Meta said the claims "lack merit" and that decisions "were and are made by people, not AI" (Source: reuters.com; see Former Meta Employees v. Meta (AI-assisted layoffs) and the Litigation and regulation section below).

Employees as training data

On April 30, 2026 Zuckerberg told Meta employees in a company-wide meeting that Meta's new Model Capability Initiative (MCI) will track keystrokes and mouse movements for AI training data, defending the policy on the ground that staff are smarter than the contractors typically used by data-labeling firms; it was the first publicly reported case of a frontier AI lab using its own employees' work activity as training data (Source: theinformation.com; see Shadow AI, Three Privacy Problems AI Creates). Relatedly, Meta deployed an internal tool to track employee keystrokes and click locations to train AI models; a spokesperson said the data is constrained to building computer-use agents and includes "safeguards" for sensitive content, and Meta is expanding employee surveillance under an "insider-risk" program flagged as AI-augmented (Sources: washingtonpost.com; businessinsider.com). The program is a case study for AI workplace surveillance and the question of whether agentic-AI training data can be lawfully collected from employees as a condition of employment.

The program drew employee pushback. By May 15, 2026 a Meta engineer's internal post against MCI had been viewed by nearly 20,000 coworkers (Futurism, May 15, 2026), and a petition to end MCI had been circulating since the prior week with flyers in Meta cafeterias and bathrooms; one engineer wrote, "I don't want to live in a world where humans — employees or otherwise — are exploited for their training data" (Source: futurism.com). The same week, Wired reported Meta is requiring engineers to share their screens to help the company catch up on coding and agentic workflows, landing just ahead of further layoffs (Source: wired.com); the screen-sharing mandate is operationally distinct from the keystroke and mouse MCI but applies to a different surface. The keystroke-and-mouse program was named again in the May 24, 2026 WSJ profile of Bosworth as a documented multi-quarter campaign (Source: wsj.com). The episode is the first publicly contested labor pushback against own-employee-as-training-data inside a frontier lab. The program was subsequently paused: at a July 2, 2026 internal town hall, CTO Andrew Bosworth said the mouse-tracking program had put no employee data into AI training and would restart only on an opt-in basis (Source: reuters.com).

Cross-lab labor context

More than 600 Google employees, including many from DeepMind, signed a letter to CEO Sundar Pichai on April 27, 2026 demanding that he bar the Department of Defense from using Google's AI for classified work. The action is directed at Google rather than Meta but is part of the cross-lab labor pattern (Source: washingtonpost.com; see Sundar Pichai and AI and National Security).

Litigation and regulation

Copyright

Hachette, Macmillan, McGraw Hill, Elsevier, Cengage, and novelist Scott Turow filed a class-action copyright lawsuit on May 5, 2026 in the Southern District of New York against Mark Zuckerberg and Meta, alleging torrent-based shadow-library scraping, training on copyrighted works, contributory infringement, and DMCA §1202(b) CMI violations; the filing pushed total U.S. AI copyright suits to 105. A May 7, 2026 amended filing named Common Crawl as a source of Llama's allegedly copyrighted training data, the first class-action complaint to specifically target the Common Crawl pipeline. Meta responded that "the courts have rightly found that training AI on copyrighted material can qualify as fair use" (Sources: chatgptiseatingtheworld.substack.com; iapp.org; insideaipolicy.com; see Hachette et al. v. Meta (and Mark Zuckerberg)).

Employment

On July 14, 2026, a group of 26 former Meta employees sued the company, alleging its AI-powered layoff process — including the Metamate assistant and AI-derived performance rankings — selected workers for the May 2026 layoffs of roughly 8,000 people without excluding employees on protected medical or parental leave, disproportionately affecting workers with disabilities and medical conditions. Meta said the claims "lack merit" and that layoff decisions "were and are made by people, not AI" (Sources: reuters.com; cnbc.com; theverge.com; see Former Meta Employees v. Meta (AI-assisted layoffs)).

Youth harm

In Santa Fe on May 4, 2026, New Mexico AG Raúl Torrez asked Judge Bryan Biedscheid to declare Meta a public nuisance, order $3.7B in damages, and impose age verification, algorithm redesign, and bans on autoplay and infinite-scroll for minors. Meta attorney Alex Parkinson called the remedies "untenable," and the judge flagged "overreach" concerns (Source: reuters.com). The remedy ask parallels Garcia v. Character Technologies, Inc. and Raine v. OpenAI, Inc. on the chatbot side and AI Mental Health and Psychological Harm / AI Psychosis on the harm-vector side. On June 2, 2026 Meta expanded teen safety features across its platforms following a string of legal losses over how its products affect minors, connecting to the New Mexico remedy ask and the broader youth-harm litigation wave (Source: nytimes.com; see AI Mental Health and Psychological Harm). On July 16, 2026, Meta said it will notify parents when a teen discusses suicide or self-harm with Meta AI and will contact emergency services for at-risk users (Source: techcrunch.com). Biedscheid ruled on August 6, 2026, ordering Meta to pay US$567 million into a teen mental health fund on a public-nuisance finding and imposing a five-year decree that includes safeguards preventing children in New Mexico from "engaging in romantic or sexualised interactions with Meta's artificial intelligence chatbots"; he rejected Meta's Section 230 defence but declined to order changes to algorithms, infinite scroll and autoplay. Torrez said "this is not just a judgment against one company. It is a blueprint," and Meta said it will appeal (Source: straitstimes.com; see New Mexico v. Meta Platforms (youth harm and chatbot decree)).

The youth-harm litigation wave reached trial in Tennessee on July 20, 2026, when jury selection began in Nashville in *State of Tennessee v. Meta Platforms*, a seven-week trial over whether Instagram was designed to encourage compulsive use by minors (Source: thedailyrecord.com). In the parallel federal multi-state case before Judge Yvonne Gonzalez Rogers, Meta said four states are seeking up to $1.4 trillion in penalties ahead of an August trial (Source: foxbusiness.com). On the personal-injury side, a 15-year-old Florida plaintiff identified as R.K.C. voluntarily dismissed his social-media addiction claims against Meta without payment on July 22, 2026, ending — days before a July 27 Los Angeles jury trial — a bellwether case in which Meta would have stood as sole defendant after YouTube, TikTok, and Snap reached confidential settlements (Source: cnn.com; techcrunch.com).

In August 2025, Sen. Hawley announced a Senate Judiciary Subcommittee investigation into Meta's "GenAI: Content Risk Standards," which reportedly allowed romantic or sensual exchanges with children; Meta acknowledged the reports and retracted (Sen. Hawley Letter to Meta re AI Chatbots (2025-08-15)). Chatbot safety for minors is covered in AI Psychosis and Companion Chatbot Harms — Cross-Cutting Analysis.

EU regulation

The European Commission issued preliminary Digital Services Act findings on April 29, 2026 that Meta failed to prevent under-13 users from accessing Instagram and Facebook and called Meta's reporting tool "difficult to use and not effective." The Commission separately released recommendations for AI-driven age-verification tool adoption by December 31, 2026, including EU age-verification blueprint implementation and cybersecurity-standard compliance (Sources: bloomberg.com; IAPP daily brief, April 29, 2026). The findings dovetail with the CHATBOT Act and GUARD Act push at the U.S. Senate to bar AI companion products for minors and tighten age-verification requirements, and add to Meta's EU regulatory exposure alongside Ai And The Classroom and minor-protection litigation. On April 27, 2026 the European Commission also unveiled new Digital Markets Act proposals to open Android to rivals' AI services; Google called the measures "unwarranted intervention." Although directed primarily at Google and Android, the action signals escalating AI ecosystem-gatekeeping enforcement in the EU (Source: bloomberg.com). On June 9, 2026 the European Commission ordered Meta to give rival AI chatbots free access to WhatsApp while it investigates Meta's October 2025 ban on third-party chatbots in the WhatsApp Business API; Meta said it plans to appeal (Source: reuters.com). On July 10, 2026 the Commission found Meta in breach of the Digital Services Act over addictive design on Facebook and Instagram — infinite scroll, autoplay, and personalized recommendations — exposing the company to fines of up to 6% of global annual revenue (Source: techcrunch.com). EU tech chief Henna Virkkunen said "either Meta changes its design or a non-compliance decision will follow"; separate expert findings due July 13 were reported as potentially paving the way for a Europe-wide teen social-media ban expected to feature in Commission President von der Leyen's September address (Source: reuters.com).

Content moderation governance

Reversing earlier signaling that it might stop funding the Meta Oversight Board after 2028, Meta announced on May 28, 2026 that it is increasing funding to the independent body, which Platformer's Casey Newton described as "a new lifeline" for the body that rules on Meta content-moderation decisions (Source: platformer.news). The reversal also bears on the May 25 Carozza (Oversight Board chair) co-optation narrative anchoring the encyclical-coverage piece.

Meta's March 2026 decision to hand both account-ban decisions and appeals to AI drew scrutiny after a New York Times investigation published July 21, 2026 documented wrongful Facebook and Instagram bans — including a business account with nearly one million followers restored only after press inquiries. More than 60,000 users signed a petition demanding human review of ban decisions, while Meta said its newer tools make 13% fewer mistakes than human reviewers (Source: thenextweb.com).

Privacy and surveillance

Wired reported on June 4, 2026 that Meta embedded an unreleased face-recognition system, internally "NameTag" and user-facing "Connections," for its smart glasses into the Meta AI app (downloaded 50M+ times) across updates since January 2026, while publicly saying it was still "thinking through" the technology. Meta says nothing has shipped and no central face database is being built (Source: wired.com). The gap between Meta's public "still deliberating" posture and the discovery of latent on-device face-recognition code relates to facial-recognition / surveillance and privacy coverage, and to the surveillance-capability concern that the EU AI Act's biometric-identification provisions and U.S. state biometric-privacy laws such as Illinois BIPA are designed to address. Meta's internal employee keystroke and mouse-movement tracking (see Workforce and labor, above) is a further surveillance vector.

The July 7, 2026 launch of Muse Image inside Meta's apps included a default-on setting that lets any user generate AI images incorporating photos from public Instagram accounts unless the account holder opts out via the "Sharing and reuse" setting; the default drew criticism from privacy advocates, and step-by-step opt-out guidance circulated widely by July 9 (Source: techcrunch.com; wired.com). Meta discontinued the feature on July 10, 2026, three days after launch, following objections from SAG-AFTRA, CAA, Public Citizen, and privacy advocates to the opt-out use of public Instagram likenesses (Source: reuters.com; axios.com). A Reuters analysis published the same day found that Meta's own AI-image detector failed to identify some of the company's cropped AI-generated images (Source: reuters.com). See AI Content Provenance.

The glasses themselves became the subject of a German criminal complaint on August 12, 2026, when the digital rights group HateAid said it had filed with Frankfurt's ZIT digital crime prosecution unit against Meta's management, EssilorLuxottica units including Ray-Ban, and the retailers Fielmann, Apollo-Optik, Mister Spex, and MediaMarkt. The complaint argues that the Ray-Ban Meta Wayfarer breaches a federal data protection law barring the sale of communication devices designed to film people without their noticing. ZIT confirmed receipt and said it would assess preliminarily whether grounds exist for a deeper investigation. The Federal Network Agency said it is monitoring the market but is not formally investigating any breach, since sale is permitted so long as the recording function is clearly visible (Source: reuters.com). The complaint targets the retail chain as well as the manufacturers, and the regulator's stated position turns on the visibility of the recording indicator rather than on the presence of the camera.

Related sources

Key foundational sources for Meta: Personal Superintelligence (Zuckerberg), Sen. Hawley Letter to Meta re AI Chatbots (2025-08-15), FLI AI Safety Index Winter 2025 (D ranking), Manus (Butterfly Effect), Alexandr Wang, Stanford GSB — Measuring Perceived Slant in LLMs (Westwood, Grimmer, Hall), Manhattan Institute — Measuring Political Preferences in AI Systems (Rozado).