Australia Kicks Kids Off Social Media + Is the A.I. Water Issue Fake? + Hard Fork Wrapped

Summary

• Australia’s under-16 social-media ban is the clearest regulatory proof yet that major platforms have “lost this argument.” Ten services—including Instagram, TikTok, YouTube, Snapchat, X and Reddit—must use age checks rather than birthday menus, threatening younger-user engagement. Kevin Roose expects other democracies to follow within one or two years if Australia’s teenagers do not visibly suffer, even though meaningful mental-health results may take five to 10 years.
• Platform exclusions and product architecture could reroute engagement without eliminating teenagers’ online activity. Roblox and Discord escaped the initial list, while YouTube can retain under-16 viewers through a personalized logged-out feed; Casey expects Meta may answer with logged-out versions of Instagram, Threads and Facebook. Teens are already using false-age backup accounts, migrating to CoverStar, Yope and Lemonade, and even exchanging phone numbers.
• Both hosts expect modest net benefits, while preserving the costs for young creators and vulnerable teenagers who depend on online communities. Casey’s public-health analogy is that many people use alcohol or tobacco without ruin, yet societies still restrict them for minors; the missing counterclaim is, “Looking at Instagram is really good for your kid.” Kevin, influenced by apparently successful school phone bans, thinks critics underestimate how often “sometimes this stuff just works.”
• The viral claim that one chatbot prompt consumes a bottle of water is, in Andy Masley’s account, a category and denominator failure. Google’s data suggests roughly 0.26 milliliters on-site and perhaps two milliliters including electricity generation, versus an average American daily footprint of about 1,600 liters; raising that footprint 1% would require approximately 8,000 median prompts. Skipping one pair of jeans, he estimates, corresponds to roughly two million prompts.
• AI’s industrial water demand looks manageable to Masley, but its electricity demand remains a material unresolved infrastructure risk. Data-center water use could reach the equivalent of about 250 square miles of irrigated corn—roughly 1% of US corn—by 2030, while electricity demand may climb from 0.5% to 5% of the national grid versus water reaching about 0.5%. Coal plants staying open is already a negative, and Masley is “really agnostic” about whether renewable procurement and grid investment will outweigh the damage.
• The year-end policy picture combines attempted federal deregulation, looser chip exports and renewed pressure for executive accountability. A draft Trump order would challenge state AI laws and potentially withhold funding despite a 99–1 Senate rejection of a state-law moratorium; separately, Washington would let NVIDIA sell H200 chips to China while taking 25% of sales. The hosts call that a lobbying victory, while the strong response to their Roblox CEO interview showed audience demand for accountability about child safety.

Deep dive

1. Australia imposed a real age gate, not another birthday box

• The law bars Australians under 16 from holding accounts on Facebook, Instagram, Threads, TikTok, Twitch, Snapchat, YouTube, X, Reddit and Kik. Platforms cannot rely on a self-declared birth date; new adult users may also need to prove their age, commonly through a video-selfie estimation system that works often but has sometimes been fooled by video-game characters.

• Casey traced the policy’s unusually fast path to the wife of the premier of Australia’s second-smallest state reading Jonathan Haidt’s The Anxious Generation and deciding something should be done. Australia passed the legislation less than a year later, last November, then gave platforms roughly a year to prepare before the restrictions took effect.

• Roblox and Discord are the omissions generating the strongest platform “eye rolls”; Minecraft and Steam are also outside the initial list. Industry critics see the government “picking winners and losers,” while eSafety Commissioner Julie Inman Grant said Roblox had introduced features including age estimation and was acceptable for now—without ruling out adding it later.

• YouTube retains a structural advantage because minors can watch without accounts and still receive a personalized feed based on viewing, browser and location signals. Instagram has no equivalent logged-out experience, making the policy feel to Meta like “a giveaway to YouTube”; Casey’s practical prediction is that Meta will build logged-out versions of Instagram, Threads and Facebook.

2. Teen workarounds will test whether the ban changes behavior or merely traffic

• Australian teenagers had a year to prepare. Some created backup accounts claiming to be 19 or 26, while others downloaded uncovered services such as CoverStar, Yope and Lemonade—a predictable migration that could send young users into smaller, less regulated corners rather than remove them from the internet.

• The most revealing adaptation is deliberately low-tech: teenagers are exchanging phone numbers. Casey joked about this “ancient technology,” but the shift matters because direct messaging and video calls preserve social coordination outside the platforms covered by the ban, even as it does not reproduce the full social-media experience.

• A platform representative cautioned Casey that children will not suddenly put down their phones and “go play in the park”; they have grown up online, and one law will not restore the 1980s. Casey accepts that premise but argues fewer appearance anxieties, eating-disorder rabbit holes and all-hours notifications could still leave teenagers materially better off.

• Australia will convene academic researchers to track the affected cohort and publish findings, creating the closest thing yet to a national natural experiment on youth social media. Kevin expects limited short-term movement but potentially “pretty steep differences” after five or 10 years, once a micro-generation has passed through childhood under the rule.

3. The hosts accept painful edge cases and still expect a positive net result

• Casey kept the strongest objection intact: a 13-, 14- or 15-year-old creator could lose a real business, while an LGBT teenager in an unsupportive household could lose resources that made them feel less alone. Some teenagers also use Instagram or TikTok safely for entertainment and still become “self-actualized humans.”

• His answer is public-health logic rather than a claim that every use is harmful. Plenty of people can smoke or drink without ruining their lives, yet societies restrict tobacco and alcohol until children are older; those restrictions ultimately reduced consumption, and Casey thinks most countries are better off because of them.

• The asymmetry that decides his view is that nobody argues, “Looking at Instagram is really good for your kid,” or warns that a 14-year-old must join TikTok to avoid missing out. Social media is, at best, “a tolerated distraction” and, at worst, immensely harmful—making exclusion a “no-brainer” below 13 and a harder but still defensible choice from 13 through 15.

• Kevin said school phone bans had “radicalized” him somewhat: based on everything he has seen and the teachers and parents he has spoken with, they have been universally successful, despite earlier warnings about isolation and emergencies. That experience makes him skeptical of predictions that removing social media must generate worse knock-on effects.

4. Australia’s experiment could become a template before its results mature

• Casey defines success as some improvement in teenage mental health, perhaps accompanied by higher test scores. His prediction remains carefully bounded: the law probably will not get teenagers off their phones, and some will find harmful unregulated spaces, but “on balance” he expects at least a positive marginal effect.

• Kevin rejects a 12-month verdict because the meaningful horizon is closer to five or 10 years. Yet he doubts policymakers will wait: after one or two years, if Australia’s teenage cohort shows no sign of “falling apart,” many democracies may adopt similar restrictions without waiting for definitive longitudinal evidence.

• That diffusion has arguably begun. Denmark and Norway are pursuing bans for under-15s, while Malaysia has said it intends essentially to copy Australia’s law—turning what began as one country’s youth-safety intervention into a potentially broad regulatory model.

• Casey’s larger verdict is that platforms remained profitable after the post-2016 backlash but finally lost political trust: “We ran your experiment,” governments are now saying, “We didn’t like the result.” Preventing an entire age group from using their products is “what it looks like to lose an argument.”

5. The “bottle per prompt” story fails on assumptions and scale

• Andy Masley entered the debate after hearing people treat chatbot use as incompatible with environmentalism. The old claim that a chatbot used 10 times the energy of a Google search struck him as unimpressive—1,000 searches would not ordinarily make anyone question a person’s values—and repeated misconceptions sent him through enough rabbit holes to become a prominent commentator “basically just…doing this as a hobby.”

• His emblematic mismatch was a pizza-place conversation in which someone criticized a ChatGPT user’s environmental credentials and then ordered a meat-lovers pizza. Masley’s point was not that individual choices never matter, but that one activity can be “so much worse than the other,” and moral attention detached from magnitude obscures “where the actual bad guys are.”

• The bottle-of-water claim, he said, rests on a strained reading of a Washington Post article: it effectively assumes 10 to 20 prompts to produce a 100-word email, counts evaporation from lakes behind hydroelectric dams, and holds model efficiency around its 2020 level. He called the resulting talking point “a horrible misreading of a kind of inconclusive article.”

• Google’s current data puts a median prompt near 0.26 milliliters of on-site data-center water, perhaps rising to about two milliliters after electricity generation. Against an average American’s roughly 1,600-liter daily footprint, that is an 800,000-to-one ratio: about 8,000 prompts to add 1%, while one forgone pair of jeans represents roughly two million prompts.

6. Industrial water use is large in gallons but ordinary with a denominator

• Masley welcomed Kevin’s shift from personal prompts to system-level infrastructure, because climate debates should emphasize “the big systematic things.” Millions of gallons per day sound shocking beside household use, but the same framing applies to berries, steel and almost any national industry; raw gallon totals reveal little without comparison.

• The strongest comprehensive baseline he cited is Lawrence Berkeley National Laboratory’s report on data-center energy and water use. Its latest good data is from 2023, both because AI is moving quickly and because companies are not always transparent, so any present-day estimate must acknowledge substantial growth while making explicit assumptions about its pace.

• In 2023, Masley estimates US data centers consumed water equivalent to roughly eight copies of his 15,000-person hometown. That is significant from an individual perspective—effectively adding eight towns across the country—but still a load he expects the broader water system generally to absorb and manage.

• By 2030, data-center water use could equal around 250 square miles of irrigated corn, or approximately 1% of America’s corn. Household leaks already use substantially more water; Masley therefore expects localized problems resembling those created by other large industries, not a nationally disruptive water disaster.

7. Electricity, not water, is the environmental question Masley leaves open

• Masley draws a full order-of-magnitude distinction: AI-related electricity demand may rise from about 0.5% to 5% of the national grid by 2030, while its water share may reach only about 0.5%. The water case looks “open and shut” to him; the power case is “a huge question mark.”

• The harms are not hypothetical. Data-center demand has already coincided with coal plants staying open, and Masley allows that the right answer may sometimes be not to build a facility. He rejects any blanket entitlement for data centers and asks instead that each project’s emissions, grid constraints and competing needs be weighed.

• The upside case is also conditional: AI companies’ renewable-energy purchasing could create economies of scale in renewable energy, but Masley explicitly says that may not cancel the negative effects. With “so many unknown unknowns,” his position is agnosticism—not an assumption that any new energy demand is bad, or that AI’s demand will be fine.

• Kevin noted that accelerationists embraced Masley’s work—White House AI czar David Sacks reposted him—even though some may use it as a cudgel against environmentalists. Masley called those alliances unfortunate but maintained that bad anti-AI arguments are “drowning out the good ones,” leaving critics less prepared to contest labor harms, surveillance or even X-risk.

8. Proportion does not erase democratic legitimacy or local trade-offs

• Kevin’s steel-man was political, not numerical: society never collectively chose a sudden build-out of data centers, grid strain and capital devoted to a technology many people neither requested nor find useful. Masley answered that disliking AI is a valid reason to oppose its resource use—but not a reason to misstate how much shutting every AI data center would improve America’s water position.

• His comparison is animal agriculture, where roughly a quarter of a person’s dietary water footprint comes from crops eventually fed to animals. If water is the concern, prioritization should follow potential impact; abolishing AI infrastructure tomorrow, in his account, “is not really gonna make much of a dent” beside larger uses.

• Local economics complicate a simple harm narrative. Data centers can bring substantial tax and utility revenue to poor communities, funding upgrades where the binding water problem is aging pipes, lead contamination or delivery cost rather than scarce freshwater. A large new customer might improve access in specific places, though Masley refuses to generalize that into “data centers good all the time.”

• Masley praised Hank Green for showing bad data on both sides, including Kevin’s critique that companies spotlight per-prompt inference while training and longer reasoning chains receive less attention. His one disagreement was Green’s heavily hedged tone: complexities and error bars are real, but even the high end of a reasonable range does not make chatbot use a serious personal water concern.

9. Wrapped paired audience scale with an escalating AI federalism fight

• Hard Fork produced 3,524 minutes of content in 2025—about 59 hours or two and a half days—across 52 episodes, with 56 expected by year-end. Spotify recorded 1.4 million listening hours, equivalent to 160 years continuously; the audience’s estimated listening age was 35 to 45, matching demographic data.

• The year’s “AI regulation regulation” was President Trump’s stated plan to curb state rules such as California’s requirement that labs disclose new model risks and catastrophe plans. A circulated draft directed the US attorney to sue states where laws might burden interstate commerce and contemplated withholding broadband or other federal funds.

• Kevin supports the theory of one federal framework but objects that proponents of preemption have no concrete federal plan or route through Congress: asked what replaces state law, they “start…hemming and hawing.” Casey warned that barring state action without a substitute would radically loosen what labs may do.

• The political and legal foundation looks uncertain. The Senate rejected an earlier state-law moratorium 99–1, another attempt tied to defense authorization failed, and Casey’s reading was that executive preemption did not seem “particularly legal”—while acknowledging that a possible Supreme Court outcome remains unknowable.

10. China policy and the Roblox interview exposed two accountability gaps

• China earned “other country of the year” after DeepSeek’s R1 triggered a market shock over whether a Chinese lab had matched frontier US models for far less money. Kevin never fully believed the spending narrative, but the episode helped people realize China was not far behind; newer DeepSeek systems appear a couple of steps off the frontier, while Chinese open-source models gained adoption among US companies.

• Trump said NVIDIA could now sell H200 chips to China while the US government takes 25% of sales. It is not clear whether China will let its companies use them, since it has been developing domestic chips through companies such as Huawei; Casey also cited reporting that DeepSeek had obtained several thousand banned Blackwell chips through third countries—evidence, he argued, that H200s would find eager users.

• Kevin called the reversal “a complete victory for the NVIDIA lobbying machine.” He rejected the administration’s argument for keeping the global AI stack American: giving the leading geopolitical rival a key ingredient for frontier systems, while US labs race to retain their lead, made no strategic sense to him beyond successful lobbying.

• The year’s biggest listener response followed the hosts’ interview with Roblox CEO David Baszucki. Parents reconsidered allowing Roblox, families described serious harms, and Casey was stunned by Baszucki entertaining prediction markets in children’s games—“essentially onboarding them” into gambling—while Kevin found him defensive despite his team knowing the hosts planned to ask about child safety.

• Casey connected that encounter to CEOs favoring creator podcasts, including a pay-for-play ecosystem where some hosts reportedly accept $25,000 for an interview. Friendly formats may dull executives’ instincts, but his standard remained simpler: literal billionaires should answer basic safety questions without treating scrutiny as an attack. Kevin’s takeaway was that “doing journalism has consequences.”