Meta AI Lifts Instagram Scroll Time as Lawsuits Mount

Instagram global time spent grew by double digits year over year in the second quarter, Meta CFO Susan Li told investors, largely from AI upgrades to Feed and Reels recommendations. The biggest single Reels ranking release delivered a 15 basis point increase in sessions, with clear strength in reshares and total time spent.

Those gains arrive as Meta books $2.4 billion in legal charges tied to youth-safety proceedings and faces multi-state trials that treat the same engagement engines as evidence of design for addiction.

Double-Digit Time Spent Lands on Instagram

Li laid out the engagement picture on the July 29 earnings call. Instagram time spent rose double digits. Facebook video time spent rose 9% globally and more than 10% in the US and Canada. Instagram itself hit 2 billion daily active users.

Family daily active people reached 3.60 billion. Ad impressions across the family of apps climbed 14%. Average price per ad rose 12%. Total revenue hit Q2 revenue of $60.8 billion, up 28% year over year.

• Instagram time spent: double digits year over year
• Reels ranking impact: 15 basis points more sessions
• Instagram daily actives: 2 billion
• Legal charges this quarter: $2.4 billion

Free cash flow sank to $784 million after heavy capital spending and the legal hit. Operating margin compressed from 43% to 31%.

The revenue print and the margin print pulled in opposite directions. Top-line growth reflected both higher impressions and higher prices. Cash generation absorbed the legal charge and the compute bill at once, leaving free cash flow far thinner than the revenue pace alone would suggest.

Every Public Reel Now Runs Through an LLM

Meta no longer ranks mainly on clicks and watch history. Earlier this year the company hit a milestone: every public Reels and Feed post on Instagram is automatically processed through a large language model. The models pull topics, tone and other dimensions.

Those signals feed ranking, recommendations and content-policy enforcement. This quarter Meta also began using its Muse family of models for video topic classification and summarization. Early results look positive.

Mark Zuckerberg described the shift as giving systems a first-principles grasp of what content is about and why it is compelling, plus a deeper read on user goals. The result is more relevant matching.

• LLM analysis of every public Reel and Feed post for topics and tone
• Muse models handling video classification and summarization
• Deeper user history drawn into a new ranking architecture
• Faster inference so predictions update in real time
• Over half of recommended Instagram Feed content now less than one day old, more than double a year ago

The same stack expands the content universe. Zuckerberg pointed to Muse Image and Muse Video models that will generate nearly infinite personalized material beyond friends and followed creators.

Topic and tone extraction turns each post into a structured signal before any ranking decision runs. That step lets the system compare a new Reel against a longer slice of user history without waiting for fresh click data. Policy enforcement rides the same pass, so the classification layer serves distribution and safety review from one read of the content.

Fifteen Basis Points and the Reshare Spike

The flagship Reels update combined faster inference with architecture that pulls longer user history. Sessions rose 15 basis points. Reshares and time spent moved most, which Meta reads as better content-to-user fit.

The company is rolling the same approach into main Feed, where early results look comparable. Fresh content moves faster too: new videos get scored at creation by the largest ranking models.

On X, users immediately framed the numbers as attention engineering. One widely seen take called the LLM layer “evil” and told people to touch grass if they are not actively curating. Another observed that every public post now has an AI reader before any human does. That crowd read lines up with the legal theory of intentional design for prolonged use.

Reshares matter because they signal voluntary distribution. A user who forwards a Reel has already decided the match is worth another person’s time. Meta treats that spike, alongside the session lift, as evidence the ranking change improved fit rather than merely increasing noise.

Creators Gain Reach While Ad Inventory Grows

Longer sessions give creators more distribution. Reshares rose with the ranking change, a direct signal that matching improved. More time also creates more ad slots without raising load to the point of user revolt.

Meta’s ads systems are absorbing the same LLM upgrades. A new generative recommender reasons over ad content and user preferences together instead of scoring every ad one by one. Early pilots showed a 1% lift in app event conversions on Instagram. Broader user-understanding models produced an 8.3% rise in ad clicks and 15.7% uplift in conversions on Facebook.

Nine million small businesses now use at least one AI ad creative tool. Advantage+ tools reached a $75 billion annual revenue run-rate. The engagement flywheel and the monetization flywheel turn together. Readers watching ad risk on the platform can also look at how Meta ads put money at risk in other ways.

When sessions lengthen, inventory expands without a matching rise in daily actives. Price per ad can climb while impression counts also climb, which is the pattern the quarter showed. Creators capture more organic reach from the same ranking pass that fills those extra slots.

States Queue Trials Over the Same Engagement Engine

Li told investors that youth-related scrutiny remains a major priority in several markets. The company recorded $2.4 billion in charges related to legal proceedings this quarter. Multiple youth-focused lawsuits are set for trial later this year.

Earlier in 2026 a New Mexico jury ordered Meta to pay $375 million for consumer-protection violations tied to child safety and mental-health claims. A California jury found Meta and Google negligent in a separate case and awarded $6 million total, with Meta carrying most of the share. In July Meta disclosed that four states were seeking $1.4 trillion in penalties ahead of an August trial in Oakland.

1. Earlier in 2026: New Mexico jury awards $375 million on child-safety and mental-health claims
2. 2026 California verdict: $6 million total negligence award shared with Google
3. July disclosure: four states seek $1.4 trillion ahead of Oakland trial
4. Q2 2026 close: company books $2.4 billion in legal charges
5. Later this year: additional multi-state youth trials on the calendar

More than two dozen states and a large federal multidistrict litigation have advanced similar theories: that features were designed to maximize time on platform for young users. The new LLM depth and deeper history models give plaintiffs fresher technical exhibits. Meta continues to contest the claims and notes it is giving users more controls.

What we know

• $2.4 billion legal charges booked in Q2 2026
• Prior 2026 verdicts of $375 million (New Mexico) and $6 million shared (California)
• Four-state trial seeking $1.4 trillion set for August
• Company outlook still flags possible material loss from youth matters

What’s unconfirmed

• Final damages or settlement size in pending trials
• Whether the newest ranking models will be treated as aggravating or mitigating evidence
• Exact contribution of LLM ranking to any measured youth time-spent figures

Each verdict and each disclosure lands on the same core allegation. Plaintiffs argue that ranking systems built to raise sessions and minutes are, by design, systems built to hold young users longer. The technical detail Meta now cites for investors becomes discovery material in that framing.

Your Algo Page Lets Users Push Back

Meta is not only optimizing for time. Instagram users can open a “Your Algo” page and write natural-language prompts to tune recommendations. Facebook launched “Shape Your Feed.” Early retention among people who engage the controls sits above 80%.

Zuckerberg framed the broader LLM work as a step toward greater personalization that better matches goals, not just past clicks. The company is also building foundation models meant to power organic and ads ranking together, plus fully LLM-native recommenders. A first research milestone this half showed healthy scaling when a large model was continuously pre-trained on recommendations data.

Still, the controls sit inside a system whose primary measured success this quarter was more sessions and more minutes. Competitors are testing similar levers; YouTube tests custom feed AI that lets viewers rewrite their own ranking logic.

LLMs add a first-principles understanding of what the content is about and why it is compelling, as well as a deeper understanding of what people are interested in and what their goals are when they’re using our apps.

Mark Zuckerberg said that on the earnings call, tying the tech directly to more relevant and engaging content.

Natural-language controls give users a direct channel into the same models that rank their feeds. Retention above 80% among people who try the tools suggests the interface is usable enough to keep. Whether those prompts can offset the default pull of deeper history and faster inference remains an open product question the quarter did not settle.

Investors and Courts Read One Stack Differently

The quarter put a single technical stack in front of two audiences with opposite scorecards. Investors heard double-digit Instagram time spent, a 15 basis point session lift, 14% more ad impressions, and 28% revenue growth. Courts and state attorneys general hear the same lifts as proof that engagement design is working as alleged.

The $2.4 billion charge already prices part of that collision into the income statement. Operating margin moved from 43% to 31% with the legal hit and capital spend both landing in the same period. Free cash flow at $784 million shows how quickly those items can shrink cash conversion even when revenue accelerates.

• Investor lens: time spent, sessions, impressions, price per ad, revenue growth
• Plaintiff lens: time spent, sessions, deeper history, LLM ranking depth
• Shared exhibits: Reels release notes, Muse rollout, fresh-content ratios
• Shared unknowns: youth-specific time-spent splits, final trial damages

Meta’s defense points to user controls and contests the design-for-addiction theory. The “Your Algo” page and “Shape Your Feed” tools are part of that answer. Plaintiffs can still treat the primary success metrics of the quarter as consistent with the pattern they describe. Both readings draw from the same public numbers.

Capital Spend Underwrites the Next Ranking Wave

Heavy capital spending helped pull free cash flow down even before the legal charge. The compute bill supports faster inference, larger ranking models at creation time, and the Muse family used for classification and future generative supply. A new 1 GW data-center venture with BlackRock in El Paso was announced around the earnings window, extending that capacity path.

Margin compression from 43% to 31% therefore mixes one-time legal cost with ongoing infrastructure cost. The legal line may not repeat at the same size each quarter. The compute line is more likely to stay elevated while foundation models for organic and ads ranking, and fully LLM-native recommenders, remain on the roadmap.

Subscription surface area is growing in parallel. Meta One packages extra AI features for a fee, and business agents plus Muse-powered creative tools continue to scale. Those products recycle the same model investment into revenue lines that sit beside the core ads engine rather than inside it.

Facebook Feed Gets the Same Treatment Next

Meta is already extending the Instagram LLM pipeline to more Facebook surfaces. Early Feed results on Instagram look comparable to the Reels lift. Zuckerberg said the same framework will roll across Facebook’s main feeds.

The company also shipped Meta One, a subscription with extra AI features, and continues to scale business agents and Muse-powered creative tools. Capital spending remains elevated to support the compute. A new 1 GW data-center venture with BlackRock in El Paso was announced around the earnings window.

For users the next months will show whether natural-language controls and fresher content offset the pull of deeper personalization. For creators and advertisers the inventory and matching gains look durable. For the courts the same technical stack that produced the 15-basis-point session jump is now part of the record heading into trials that already forced a multi-billion-dollar charge this quarter.

The irony is structural. Better understanding of content and users produces both the double-digit time-spent print investors cheered and the prolonged-engagement pattern regulators and plaintiffs are trying to price.