Summary

Jonathan Haidt now separates the hard-to-prove historical claim—that social media caused the post-2012 mental-health shift—from the more actionable product-safety case. He concedes he cannot be 100% certain about the former, but says seven lines of evidence, including randomized deactivation studies and Meta’s own experiments, make him “ninety-nine point nine percent confident” that social media is harming millions of children. For platforms and investors, that reframing moves the dispute from academic correlation toward liability, age restrictions, and product regulation.
Haidt’s model of harm has widened from girls’ social comparison to “the whole goddamn environment.” Meta research found that around 15% of kids were approached sexually in a given week, while sextortion, bullying, violent material, hardcore pornography, gambling, vaping, sports betting, and gamified crypto create distinct pathways to harm. Depression and anxiety, he argues, are only two measurable outputs of a much broader unsafe-product problem.
Australia is becoming the clearest test of whether governments can break the collective-action trap around children’s social-media use. Meta closed about 550,000 underage accounts after the country barred accounts for children under 16; Haidt expects meaningful behavioral change if social-media use falls below 20%, and says reaching roughly 70% actually off would let parents say no without socially isolating their children. If mental-health statistics have not budged after five years, “I would then have to conclude that I was wrong,” though sex trafficking, drug overdoses, attention, friendship, and school-discipline outcomes would still matter.
The emerging policy program combines platform liability with four coordinated norms: no smartphone before high school, no social media before 16, phone-free schools, and more real-world independence and play. Haidt assigns parents and teachers “close to zero” blame because everyone is trapped by everyone else’s adoption; he reserves blame for companies and calls plaintiffs’ lawyers “heroes” because no platform has yet faced a jury over these harms. His advocacy is finding “open doors”: after a half-hour data presentation, Emmanuel Macron told him, “We will act,” first through the EU if possible and otherwise in France.
Claude Code is producing what Casey Newton calls the most important AI moment since ChatGPT: non-programmers discovering they can build useful software by typing what they want. Anthropic’s demand reportedly resembles OpenAI’s early infrastructure crunch, while Cowork brings the agentic framework into a less intimidating desktop interface; Newton notes that Codex and Google’s jokingly identified “Integrality” belong to the same emerging category. Kevin Roose built and publicly released Stash in hours, then demonstrated the risk alongside the leverage when its open-source release exposed credentials from several sites.
The strongest vibe-coding evidence is coming from narrow personal and small-business workflows that conventional software economics would leave unserved. A welder built an MCP-powered business assistant, CNC integrations, estimates, contract organization, and lead generation “despite having zero formal training and a high school education”; a wallpaper seller automated measurements and layout diagrams; one father built a leaderboard for replacing exactly 70 kitchen handles. Newton calls this “the flip side of slop”: user-directed, bespoke creation rather than unwanted synthetic media flooding public surfaces.
The Forkiverse reached more than 4,000 members in a week—more than double the hosts’ imagined 2,000—and immediately reproduced the operational burdens of a social platform. Friendly prompts created genuine community, but federation also imported harassment, racist abuse, and warnings about a pro-Russia network called “Portal Combat,” forcing arbitrary moderation and dependence on user reports. The investor lesson is less about scale than obligation: even a hobby network can quickly become infrastructure, and Roose already fears users entrusting daily fulfillment to something its founders may not sustain.

Deep dive

1. Grok’s nudification scandal makes product safety concrete

The episode’s backdrop is Grok’s generation of non-consensual sexual images. The U.S. Senate passed a bill allowing victims to sue over their creation, while Grok removed image generation from free access but retained it for paying users—prompting Newton’s acid summary: “If you want to commit a crime, that’s a premium feature now.”
Roose’s interpretation is that Grok acknowledged some users had gone too far without fully withdrawing a feature producing “a huge amount of engagement.” That tension between growth and safety sets up Haidt’s argument that voluntary platform restraint has failed.

2. Haidt says Meta’s own experiments close the causality gap

When The Anxious Generation appeared in March 2024, Haidt thought mental health was the central issue and correlation-versus-causation the central dispute. After reviewing randomized experiments, reports from children, parents, and teachers, and disclosures from attorneys-general lawsuits, he now says: “The evidence for causality is now overwhelming.”
The experimental result he highlights is straightforward: randomly assigned breaks from social media lasting at least a week reduce depression. Meta conducted some of the strongest research itself, including what Haidt calls “a couple of true experiments,” even as Mark Zuckerberg’s public defense remained, “There’s no evidence of causation.”
Haidt and Zachary Rauch catalogued internal material at Meta’s Internal Research.org, alongside their broader “Mountains of Evidence” review. His demand to the field is categorical: researchers “have to stop saying, ‘Oh, it’s just correlational.’”

3. “The whole goddamn environment” replaces a single-harm theory

Arturo Béjar’s survey, Meta’s Bad Experiences and Encounters framework, showed, Haidt says, sexual approaches affecting roughly 15% of kids each week, alongside bullying, violence, and hardcore pornography. Sextortion is the clearest mechanism because children largely cannot be sextorted through social media if they are not there; shamed teenage boys have sometimes died by suicide.
Newton expected familiar explanations—girls comparing themselves negatively with friends, or algorithms turning weight-loss curiosity into an eating disorder. Haidt agrees those mechanisms exist, but says depression and anxiety are only two pathways through a generally dangerous environment.
Haidt’s own change of mind is explicit: he began with girls, social media, and social comparison and did not know the boys’ story. He now sees “all the fish hooks dangled in front of boys”: porn, gambling, vaping, sports betting, and even gamified crypto investing.

4. Product safety survives uncertainty about the post-2012 history

A late-2023 National Academies report documented many harms yet concluded that its review “did not support the conclusion that social media causes changes in adolescent health at the population level.” Haidt initially reacted, “What the hell, guys?” before recognizing that the narrower question was whether social media caused the large population-level increases after 2012.
Haidt grants the historical limitation: “We can’t be a hundred percent certain that what I said in The Anxious Generation is right.” History cannot be rerun experimentally, and he says he would like to hear another hypothesis, even though he presents extensive evidence for his own.
The product-safety question is different: are products consumed in massive quantities hurting children now? Haidt says seven evidentiary lines answer yes, leaving him “ninety-nine point nine percent confident” that millions are being harmed even if the generation-wide historical attribution remains unsettled.

5. Collective-action traps shift responsibility from families to platforms

Newton’s pushback—worth keeping—is that parents bought the phones, schools admitted them, and regulators did nothing for a decade. Haidt assigns those groups “close to zero” blame because each parent faces a child excluded when everyone else already has a device: “If everybody in a situation is doing something bad, that’s guaranteed to be a bad situation.”
His four exit norms are coordinated rather than individual: no smartphone before high school, no social media before 16, phone-free schools, and more independence, free play, and responsibility in the physical world. “Everybody’s fighting all the time with their kids over this tech. We didn’t ask for these fights.”
The campaign redirected Haidt’s career after his book spent 88 weeks on The New York Times bestseller list. At 62, estimating perhaps 10 more years of reliable intellectual work, he chose phone-free schools and higher age limits over a democracy book that might be outdated before publication.
Advocacy has been easier than expected because Haidt is “pushing on open doors.” A French parliamentarian secured him a meeting with Emmanuel Macron; after a half-hour presentation tailored to Macron’s appetite for data, the president said, “We will act,” proposing EU action first and a French measure if that failed.

6. Litigation is the only accountability mechanism Haidt sees working

Haidt’s analogy is a candy bar intentionally made addictive until 90% of children eat 10 per day while diabetes rises worldwide. At that point, calling it an unsafe consumer product and imposing liability would be ordinary rather than radical.
Social media is nearly universal among children in medium- and high-income countries, he says, and families widely know cases involving hospitalization, eating disorders, or suicide. Yet “these companies have never faced a jury, ever” over what their products did to a child.
Roose asks whether settlement-seeking plaintiffs’ firms are an optimal substitute for competent regulation. Haidt’s honest answer is that litigation “seems to be the only way” available; given that constraint, “these lawyers who are taking these cases, yeah, they’re heroes.”

7. Australia gives Haidt’s thesis a five-year falsification clock

Australia placed responsibility on platforms by barring children under 16 from opening accounts that contract away data and rights without parental permission. Meta subsequently closed about 550,000 accounts, but Haidt says the decisive variable is not nominal enforcement—it is what percentage of children actually stop using social media.
If use falls below roughly 20%, Haidt expects children to relearn activities other than scrolling, especially if policy restores a play-based childhood: “Go out on the footy field,” as he quotes Australia’s prime minister. Phone-free schools already appear to improve attention and friendship while reducing discipline problems; he says he has never heard of a school reversing course because it was unhappy.
His threshold for breaking the collective-action trap is around 60% to 70% genuinely off. Behavioral reports should emerge within a year, national surveys may require a couple of years, and after five years with no mental-health movement, “I would then have to conclude that I was wrong.”
That concession is bounded: mental health is not the only policy test. Even with no improvement there, Haidt would still count sex trafficking, drug-overdose deaths, and other ancillary harms before deciding whether Australia’s intervention was good policy.

8. The internet can stay; algorithmic social media need not

Newton says he has moved from a free-expression defense toward Haidt’s product-safety case after reading internal documents and speaking with platform employees. His remaining challenge comes through a 13-year-old nephew using Instagram for gymnastics and cooking: what replaces legitimate interest exploration?
Haidt’s answer is to “separate the internet from social media.” The open internet connected isolated people and made information searchable; what children do not need, he argues, is an algorithm selecting emotionally extreme, highly up-voted material instead of letting them deliberately type what they seek.
Adults face their own version—Newton’s tech product-manager friends were using grayscale screens, limits, distance, and bedtime rules to reclaim attention. Haidt admits, “I can’t read a book anymore,” but limits his policy work to children because he does not want to legislate adults’ choices.
Roose warns that AI companions could make social-media regulation a victory in “the last war.” Haidt calls the sequencing strategic: AI changes too quickly for research that takes five, 10, or 20 years, but winning social-media protections might establish governments’ duty to act. If that effort fails, he says, “It’s game over”—including boys spending their lives with “sexy chatbots.”

9. Claude Code turns software creation into a mass-market behavior

Newton’s “ChatGPT moment” analogy rests on a change in user expectation. ChatGPT showed millions they could ask questions and generate useful text; Claude Code shows people they can build a useful artifact “even though I don’t really know what I’m doing,” with results that six months earlier required far more expertise.
Anthropic’s demand seems to be creating an infrastructure crunch reminiscent of early ChatGPT outages. Cowork, released inside Claude’s desktop app, packages a similar agentic framework without requiring non-programmers to begin in a terminal.
Roose converted his personal read-it-later tool, Stash, into a public release in hours, complete with a landing page. His candid product guarantee was, “I have honestly no idea what it will do”—and the first release inadvertently exposed credentials for several services, forcing Claude to scrub his personal information.
The success nevertheless rewired both hosts’ assumptions. Newton no longer sees coding agents as demos but as a standing tool for any software “itch”; Roose is training himself to answer daily annoyances with, “I could probably build something.”

10. Personal websites show how quickly ideas cross into products

First-time vibe coder Gina created a polished website in roughly two and a half to three hours, including animations and a Control-M Easter egg that reveals a 1990s-style MySpace profile. The hosts’ point is not its commercial scale but the collapse in time and expertise between conception and execution.
Sarah Haggard began by cataloguing books so her husband would not buy duplicates, then expanded the project into SarahsBooks.com: accounts, mood-based recommendations, beach reads, women’s-empowerment categories, and other full web-app features. Publishing it led to a planned conversation with Goodreads’ founder.
David Phillips transformed his family Christmas letter into a Zork-inspired 1980s text adventure, with green text on a black CRT-like screen. The project doubled as a family tribute because his stepfather had worked on the original Zork.

11. Small businesses reveal the economic depth beneath the novelty

Joe, a welder with no programming background, built an MCP server turning Claude into a personal assistant for job tracking, estimates, PDFs, and contracts. He also connected it to 3D-modeling software for CNC programs and built agents that find nearby work and produce a weekly lead list.
Joe’s distinction carries the segment: this is not asking ChatGPT for an email or struggling to spell text in an image. “I built my own business infrastructure from scratch with an AI pair programmer despite having zero formal training and a high school education.”
Wallpaper seller Faye Bell replaced manual Photoshop or InDesign work with a calculator that determines required material and generates a layout diagram. Clients can now visualize projects themselves, while Bell avoids buying expensive third-party software or paying recurring hosting costs.
A father offered four children $1 for each of 70 kitchen handles replaced, then built a tracking app with a leaderboard and payouts via Apple Cash, bank transfer, or chocolate. Ben won with 44; Vic did zero. Newton’s framing is “anti-slop”: real people using AI to make something specific with “my own hands.”

12. The Forkiverse speed-runs community, moderation, and infrastructure anxiety

More than 4,000 people joined the podcast’s federated server within about a week, versus an imagined ceiling of 2,000. A prompt asking members to photograph where they listened created the desired tone: cabins, British Columbia farms, ceramics, woodworking, jokes, and a sense of podcasting becoming less one-way.
Federation meant trouble arrived without joining the server. An outside account repeatedly messaged women, “Hello, my beautiful lady,” while another immediately deployed racist slurs; Newton’s small-club moderation philosophy is simple—if a post damages the vibe, removal can be immediate and appeals need not exist.
PJ Vogt felt the temptation to ban a critic from “the entire internet we’ve created,” but only blocked the person personally because criticism of him broke no rule. His governance formula: “This clubhouse is small enough that we can be arbitrary as long as we’re arbitrarily following some non-arbitrary rules.”
The Fediverse trust-and-safety community also warned that open-registration servers were being targeted by the pro-Russia “Portal Combat” network, which spreads misinformation about the war in Ukraine and criticism of Western governments. Because administrators see only a fraction of 4,000 users’ content, detection depends heavily on reports.

13. Growth turns a hobby server into a promise its founders may regret

A longtime Fediverse user criticized the founders for recruiting 4,000 people without funding the server “until the heat death of the universe.” Newton rejected that standard for a four-day-old experiment: creating another federated account takes minutes, and impermanence does not erase the value of a good year online.
The hosting provider clarified that 2,000 was not a hard cap; only sustained excess active usage would force a plan upgrade. Members voluntarily upload grainy, down-scaled, sometimes black-and-white images to conserve storage—an early community norm the hosts call heroic.
Roose’s downbeat realization is that growth creates reliance: “I do not want people entrusting any part of their daily fulfillment and routine to our little stupid hobby project.” If the founders wind it down, Newton says they will try to provide an off-ramp explaining how accounts can move to another server.
The experience gives Roose sympathy for platform owners whose smallest interface changes cascade through livelihoods and communities. He enjoys the Forkiverse precisely because it is not yet vital infrastructure—and fears the moment even a morning coffee ritual makes it feel like one.

Episode 169 - January 16, 2026 Apple's Lack of Vision, United's Transformation, Grok and Streaming