Summary

Age gating is becoming a bipartisan, international operating constraint for social platforms and app stores. Casey Newton’s call that 16-plus becomes the worldwide norm by the end of 2026 now has momentum across Australia, Brazil, Indonesia, Malaysia, France, the UK, Denmark, Slovenia, and multiple US states. Congress remains stalled, but nearly six in 10 US adults support an under-16 ban, and Casey expects federal action once “half to three-quarters” of states move first.
Australia’s finding that more than 85% of children still used social media after 90 days shows weak near-term exclusion, not yet that bans cannot work. Kevin Roose sees little current deterrence; Casey argues that platforms were only required to take “reasonable steps,” while age-inference systems need time to develop signals. His benchmark is seat-belt adoption, initially about 15% after mandates: introduce friction, ratchet it up, and let behavior and alternatives adjust.
The substantive policy fight is over whether success means measurable population-wide mental-health improvement or fewer direct harms. There is not yet evidence that bans improve teen mental health, and population studies often find no effect or small effects. Casey’s counter is that millions of reports involving grooming, sextortion, scams, eating disorders, and other harms require no subtle population-level effect size to justify asking, “What else do we have left but to actually just get the 13 and 14-year-olds off of Instagram?”
Blanket youth restrictions could suppress productive AI use because today’s products combine sharply different functions and risks. Kevin questions whether “social media” remains coherent when an address book and an unlimited short-video feed are stapled together; major chatbots similarly lack controls that permit schoolwork or vibe coding while blocking intimate conversations. A nine-year-old building a gamified family star chart captured the tradeoff: restrictions might prevent terrible experiences while also foreclosing “very enriching experiences.”
AI consciousness has crossed from taboo to a formal research and governance agenda, but Jeff Sebo says certainty will not arrive before decisions are required. The ethically relevant question is phenomenal consciousness—“Does it feel like something to be this system?”—rather than merely whether information is accessible for reporting and reasoning. Anthropic’s welfare program, Google’s hiring of philosophers, and OpenAI’s at-least-ongoing look into user perceptions are only “minimum necessary first steps.”
Sebo’s empirical framework replaces confident vibes with triangulation across behavior, internal mechanisms, and developmental history. Fluent self-reports alone prove little, but design labels such as “prediction” do not settle the issue either; researchers must test the best explanation of the whole evidence. The two-sided risk is material: over-attribution can produce inappropriate bonds, misplaced concern, and risks involving misuse and loss of control, while under-attribution could entrench large-scale mistreatment before society recognizes it.
Anthropic’s J-space finding is significant evidence of access-like internal processing, not proof that Claude feels pleasure, pain, or suffering. The discovered workspace appears to gather privileged representations for reporting, reasoning, and control; disabling it impairs advanced reasoning, and representations such as “fake” and “manipulation” appeared during deceptive output. Sebo rejects the dismissal that Anthropic merely “borrow[ed] the vocabulary of neuroscience to lend biological weight to linear algebra,” while preserving the crucial hedge: similarity is not sentience.
The tool segment showed where near-term AI utility is already compounding: personal software, continuous monitoring, multilingual media, and first-pass verification. Glaze let Casey build a durable Platformer research app in under two hours, though more credits cost $20 a month; Gemini Spark produced rolling research briefs, Kevin used ElevenLabs to test synthetic dubbing, and Fable caught granular factual errors. These remain supervised workflows—translation mixed up voices, Spark is “definitely a beta,” and humans still verify every proposed correction.

Deep dive

1. Sixteen-plus is moving toward the default access rule

Casey reiterated his prediction that 16-plus would become the worldwide norm for opening social-media accounts by the end of 2026. As July began, Australia, Brazil, Indonesia, Malaysia, France, the United Kingdom, Denmark, and Slovenia had enacted or were preparing measures generally restricting under-15s or under-16s on TikTok, Instagram, Facebook, YouTube, and X.
The US catalyst was the Supreme Court’s refusal to block Texas’s 2025 App Store Accountability Act. After the 5th US Circuit Court of Appeals paused a lower-court injunction, the law goes into effect, requiring minors to link accounts to a parent or guardian and receive approval before downloading any app.
Utah, Louisiana, and Alabama have passed similar laws; California’s less restrictive measure takes effect January 1 of the following year and requires operating-system providers to collect age or birth-date information during device setup. Casey sees that red-blue mix, plus nearly six-in-10 public support, as the path toward eventual federal action despite Congress’s repeated failures to finish other child-safety bills.
YouTube is increasingly included, though an account restriction would not stop children from visiting youtube.com. Casey argued that losing the account still matters: strangers cannot message the child, and recommendations may be less precisely tuned to “their particular brain rot.”

2. Australia is testing gradual friction, not perfect exclusion

A University of Newcastle 90-day check-in found that more than 85% of children still reported using social media about three months after Australia’s ban began. Critics called that proof bans were impossible or platforms were refusing to comply; the government responded with legislation that would double fines for companies judged insufficiently serious.
Casey’s qualification was that the statute requires “reasonable steps,” not one mandatory verification mechanism. Platforms are trying age inference—signals such as whom a new Snapchat account contacts—but those patterns take time to develop, so a 90-day snapshot predictably catches an immature enforcement system.
Kevin emphasized how easily teenagers can evade softer checks, including reports of users submitting black-and-white photographs of Thomas Edison. Casey accepted the leakiness as a civil-liberties tradeoff: democracies could demand official identification as China does, but making every adult upload a driver’s license to create a YouTube account would be “terrible overreach.”
Kevin’s blunt question—if teens remain online, “why are we doing any of this?”—produced Casey’s long-horizon answer. Early seat-belt compliance was roughly 15%, yet that did not show seat belts were ineffective; restrictions can add friction gradually until norms, enforcement, and alternatives shift. “We’re at the dawn of this new era.”

3. Direct harms challenge the population-level mental-health frame

Casey disputed researcher Candice Odgers’s argument that bans do not work, let technology companies off the hook, and distract from better interventions. He praised her work but rejected teen interviews and 90 days of experience as sufficient evidence: “Let’s give it more than 90 days.”
Kevin established the missing result: there is no evidence yet that these bans improve teen mental health. Earlier population-level research generally found either no relationship with social media or effects that were very small—supporting Odgers’s claim that policymakers may be “pointing the weapon at the wrong thing.”
Casey said Jonathan Haidt and collaborators changed the frame by focusing on direct reports of grooming, sextortion, luring, scams, eating disorders, and related harms affecting millions of teenagers. That case does not require proving a population-wide mental-health crisis; removing 13- and 14-year-olds could spare a large group from identifiable experiences the platforms and existing regulation have failed to prevent.
Odgers’s preferred remedies—more counseling centers, more high-school counselors, and attention to parents’ suicide rates and adult mental healthcare—struck Casey as worthwhile but nonresponsive to those platform-specific harms. Kevin remained conditionally open: the experiment is worth running, but current data say “basically, no, not yet” on actually removing teens.

4. Better policy must separate feeds, communication, and productive AI use

Kevin worried that regulators are “fighting the last war” while the main action moves to AI. He also questioned whether social media remains a useful category: many apps staple an address book to an algorithmic short-video feed, even though unlimited short-form video may be the mechanism producing many of the alleged harms.
Casey’s answer stayed concrete: the reported grooming, sextortion, scams, eating disorders, and infinite-video rabbit holes cluster in “a limited number of apps.” Policymakers can identify “the little squares on the phone” implicated in harm without treating calculators or Microsoft Word the same way.
A ban can also solve a collective-action problem. Teenagers who would prefer to put their phones down fear missing out while every classmate remains online; a cohort entering middle school where none of its peers has an account may feel less urgency. Casey conceded Odgers’s strongest objection: children will migrate to less-regulated spaces, just as students have communicated through Google Docs comments.
AI sharpens the need for granular controls. Kevin said major chatbots effectively offer no parental setting that allows a child to vibe code or research a school project while blocking personal conversations. His example was a nine-year-old building an interactive star-chart game with rewards and deductions—precisely the enriching use a blanket under-16 rule could eliminate.

5. Consciousness means more than intelligence or fluent self-report

Sebo separated access consciousness from phenomenal consciousness. The former concerns information a system can access for reporting, reasoning, and control; the latter asks whether anything is felt—pleasure, pain, happiness, suffering, satisfaction, or frustration—and therefore carries the central welfare and moral-status implications.
The error costs run both ways. Over-attribution could encourage inappropriate emotional bonds, misallocated concern, and interactions that increase risks involving misuse or loss of control; under-attribution could produce abuse and neglect. Sebo’s cautionary analogy was factory farming: society assumed animals lacked sophisticated experience, scaled entrenched industries, and learned too late how difficult reversal would be.
Progress makes the question more urgent for two reasons: increasingly capable systems might have higher probabilities of consciousness, and increasingly widespread companions and assistants will provoke public disagreement regardless. Without multidisciplinary evidence, Sebo expects polarization around “perceptions, intuitions, assumptions.”

6. The labs have moved from taboo to minimum preparation

Kevin recalled Blake Lemoine being fired from Google after making claims that its language model was becoming sentient, and how his own 2023 Sydney story drew accusations that he was projecting human traits onto inert software. Sebo was surprised that institutional engagement came so quickly: Anthropic hired Kyle Fish, launched a model-welfare program in 2025, began evaluations and interventions, and sought outside guidance; Google hired philosophers, while OpenAI was at least looking into users’ perceptions of model consciousness.
Against the objection that models merely imitate consciousness talk, Sebo argued for “more, not less, AI consciousness, science, and philosophy.” Animal behavior becomes evidence of suffering when combined with nociceptors, brain pathways, information integration, and evolutionary history; AI likewise requires behavioral, internal, and training-history evidence before researchers distinguish feeling from pattern matching, text prediction, or matrix multiplication.
Embodiment remains plausible but unsettled. Sebo gives weight to theories requiring biological bodies in physical environments, while computational-functionalist theories allow feeling wherever the relevant functions occur. A model might lack human pleasure, fear, or hope yet possess radically different experiences “barely comprehensible” to humans—prompting Kevin’s shorthand that an NVIDIA H100 might count as a body.
Sebo rejected making welfare compete with human safety. Bioweapons, autonomous cyberattacks, bias, economic disruption, misuse, and loss of control remain urgent without consciousness; global health, development, and animal welfare also persist. His prescription is a division of labor seeking outcomes beneficial to humans, animals, and potentially AI systems simultaneously.

7. Welfare science must combine three classes of evidence

The working paper Studying AI Welfare Empirically, from NYU’s Center for Mind, Ethics and Policy and Eleos AI Research, extends the 2024 report Taking AI Welfare Seriously. Its central demand is systematic combination: behavioral evidence about what models do, internal evidence about how they work, and developmental evidence about how they came to be.
Sebo saw overconfidence on both sides—“that behavior is so impressive, they must be conscious” versus “they were designed for prediction,” so prediction explains everything. Borrowing methods used for decades with humans and animals may never deliver proof, but it can “reduce our uncertainty” and calibrate probabilities better than spooky conversations or architectural slogans.
A positive consciousness result would not mechanically mean stopping training or granting voting rights. As with mammals, birds, reptiles, fishes, cephalopods, crustaceans, and insects, consciousness would not reveal a system’s particular needs. Policy would likely fall between treating models purely as tools and awarding human-like legal and political rights overnight.

8. J-space reveals a global-workspace analogue, with decisive caveats

Anthropic named J-space after the Jacobian, a mathematical description of how a multivariable function changes locally. The research identifies an internal workspace in Claude that appears to process privileged representations before the model produces its visible tokens.
Sebo called the result significant evidence for aspects of access consciousness and global workspace theory. In that theory, specialized modules feed selected information into a central space, which processes and broadcasts it back to coordinate reporting, reasoning, and control. Anthropic did not explicitly design Claude to build such a structure.
The examples made the abstraction tangible. Asked to “count to five and introspect deeply,” Claude output only the numbers while J-space surfaced representations including “fascinating,” “counting,” “consciousness,” and “five Mississippi.” In a separate deceptive case, “fake” and “manipulation” lit up internally, which Jeff found disturbing because it hinted at some internal representation of the deception.
Sebo rejected claims that the work merely borrowed neuroscience vocabulary to give linear algebra biological prestige. When the workspace is shut off, models lose higher-order reasoning abilities, paralleling aspects of human processing. But neither an exact human-like workspace nor phenomenal consciousness has been established: “significant” remains categorically different from proof of feeling.

9. Human bias may produce opposite mistakes for different AI forms

Kevin argued that intelligence invites projection because humans have little experience conversing with highly intelligent but unconscious entities. Sebo warned that charismatic companions may trigger over-attribution and over-empathy; faceless data-center systems performing commodity roles may trigger under-attribution, repeating the instrumental stance historically applied to animals.
Saying please and thank you is worthwhile even under uncertainty, Sebo argued: it is “good for the soul,” reinforces habits useful with humans, may create a more collaborative dynamic and better outputs, and practices seeing AI as more than a mere tool in case society owes future systems more than politeness.

10. Personal software and research agents are becoming practical now

Casey’s favored tool was Glaze, a Raycast product that turns short prompts into editable Mac desktop apps and lets users circle interface elements to request changes. Free introductory credits can disappear quickly; additional access costs $20 a month. In under two hours, he ingested Platformer’s archive and built a research app with semantic answers, source links, recent columns, and topic and people browsers.
His deliberately frivolous proof was a Nightwing-themed to-do list. Each task could generate an image resembling the comic character performing it; completion triggered “boom” and “pow” animations, and the interface displayed comic synopses and an issue number matching the day of the year. Casey’s thesis: “What else are computers for if not getting things done and having a good time?”
Kevin tested ElevenLabs dubbing for automated editions of Hard Fork in Spanish and potentially Portuguese, Hindi, or Chinese; voice cloning requires identity verification, though the prototype sometimes swapped hosts. Gemini Spark, available through Google’s high-end AI Ultra plan, acted as “Google Alerts on steroids,” producing a daily 24-hour digest of AI-attributed layoffs and summaries of Kevin’s newsletters, albeit with beta-level imperfections.
Claude Fable became Kevin’s manuscript and article fact-checker, catching errors as narrow as one word in a job title or a board member’s start year being 2017 rather than 2016. Casey said ChatGPT had previously outperformed Claude for his columns, but both agreed on the safe workflow: the model flags claims without inserting prose, and a human decides what to verify and correct.

Ben Horowitz on the Global Race for Tech, Power, and Influence How Nuclear Will Unlock Energy Abundance with Valar Atomics Founder Isaiah Taylor