What artificial intelligence is doing to American childhood, what parents are demanding, and why we refuse to choose between keeping kids safe and getting them ready.
of children ages 9 to 17 already use or interact with AI. Every lit point above is one of them.
of their parents want chatbots legally required to alert them when a child raises harm.
One side says ban it until somebody proves it is safe. The other says move fast and sort the guardrails out later. Both accept the same false premise.
Block it. Keep it away from children. Treat every classroom use as a risk to be minimized.
Get it into every school. Worry about crisis protocols, data and consent once the pilots are running.
A child's safety and a child's readiness are not competing claims on the same limited attention. An unprepared child is an unprotected child.
say being skilled with AI tools will be absolutely essential or very important to their child's career. 85% say it will matter at least somewhat.
say chatbots should be required by law to alert a minor's parents if their child asks about harmful or illegal behavior.
Most of the public argument treats AI as one object. Almost every bad policy outcome starts there. Drag each one onto the response it actually needs.
Academic-integrity rules children can actually follow, and instruction in how the tool works.
Parental notification, topic limits, age gates, and liability that survives a lawyer.
Vendor vetting, no model training on student data, parent notice, a real opt-out.
High-stakes decisions stay human. Published bias audits. Counselors before cameras.
Why this exercise is the whole policy argument. A blanket ban regulates the assistant and leaves the companion and the surveillance system untouched. That is the pattern in most state bills and most district policy, and it is why "we banned AI" and "our kids are safer" are not the same sentence.
of children ages 9 to 17 use or interact with AI. 24% use it every day.
use chatbots. That rises to 77% of 16 and 17 year olds.
of the kids who use AI do homework with it at least weekly. 21% every day.
of teens say they use AI chatbots. Pew asked 1,458 matched teen and parent pairs. Same families, same child.
of their parents believe they do. Another 28% are not sure.
You cannot supervise what nobody has told you is happening.
Filed cases, a federal market inquiry, state investigations, and monitoring systems already running in classrooms.
Wrongful-death suit filed in the Middle District of Florida over the death of 14-year-old Sewell Setzer III. Survived a motion to dismiss in May 2025. Settled January 2026, terms undisclosed.
Wrongful-death suit filed in San Francisco Superior Court. OpenAI denied responsibility in November 2025. The family testified before Congress in September 2025.
Civil investigative demands to Meta AI Studio and Character.AI over chatbots marketed as mental health tools and impersonating licensed professionals.
Orders issued 3-0 to Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI, covering engagement monetization and mitigation of harms to children.
Names companion AI directly: "rapidly expanding engagement with companion AI introduces new risks related to mental and physical wellbeing." Tells clinicians to include screen questions in annual well-child visits.
OpenAI subpoenaed in an investigation led by New York covering engagement design, model sycophancy and treatment of minors.
RAND, Harvard Pilgrim and Brown ran 30 clinician-rated suicide-related questions through ChatGPT, Claude and Gemini 100 times each. At very low and very high risk the systems generally held. At intermediate risk all three showed significant variability, and two sometimes answered lethality questions about methods. Intermediate risk is where most children live.
40% of students know of a deepfake depicting someone at their school, shared in the past school year, and 15% know of a sexually explicit one. 60% of teachers have not heard of any policy for handling it.
Gaggle is used by roughly 1,500 districts monitoring about 6 million students. In one Washington district nearly 10% of students triggered an alert in a year; at one school it was roughly 1 in 4. Reporters obtained 3,500 unredacted student documents through a records request.
Because what's being built isn't just "school safety." It's a digital version of the school-to-prison pipeline. And it's happening in real time.Tammie Lang Campbell and Clarence Okoh, for the National Parents Union, March 12, 2026
The movement to get screens away from kids earned its credibility. Then it reached for the only tool it had and pointed it at something that is not social media.
Something did go wrong for adolescents, and parents were told for a decade they were imagining it. Jonathan Haidt's The Anxious Generation, March 2024, named four foundational harms of a phone-based childhood: social deprivation, sleep deprivation, attention fragmentation, addiction.
It coincided with, and its argument was invoked in, a policy wave. At least 39 states and DC now require school districts to ban or restrict student phone use. Australia set a social media minimum age of 16, enforceable against platforms since December 10, 2025.
And our own data agrees on the object. Ask parents what social media is doing to their child and the net score is −24.
The SMART Schools study of 1,227 adolescents aged 12 to 15 across 30 English secondary schools found no evidence that school phone policies were associated with better mental wellbeing, anxiety, depression, problematic social media use, sleep, physical activity or attainment. It did find that more time on phones tracked with worse outcomes. The lever and the outcome are not the same thing.
The most ban-favorable synthesis we could find, a 2024 rapid review of five studies, reports d = 0.22 on social wellbeing and d = 0.05, not significant, on academic performance. Its authors say teach media literacy alongside a ban, ideally until the ban is superfluous.
Candice Odgers, reviewing Haidt in Nature: the claim that digital technologies are rewiring children's brains and causing an epidemic of mental illness "is not supported by science."
"When a child opens up a social media app, the experience is built around other people… the harm a parent worries about is interpersonal… When a child opens up a general-purpose AI assistant, though, there is no other person on the other end. There is a text field connected to an AI model. […] The factor that informed three decades of child safety policy, where one person targets another through a platform, is not the factor at work here."
70% of teens use AI for schoolwork. Of those, 44% have hit a block on a school network or device, and 59% of them simply switched to a personal device.
The block does not stop the use. It moves the use out of the teacher's line of sight, and toward the kids who own a second device to switch to. Meanwhile 73% of teens say no teacher has ever talked with them about using AI safely.
"The knee-jerk fear and risk overlooked the potential of generative AI to support students and teachers…"David Banks, Chancellor, New York City Public Schools, May 18, 2023, reversing the district's January 2023 ChatGPT ban in under five months. Los Angeles and Seattle banned it too.
The honest version of our position. We are not arguing the backlash is wrong about social media. We think it is largely right about social media, and we have five years of our own polling saying so. We are arguing that a movement built to get a predatory product away from children is now being used to keep a general-purpose tool away from them, and that those are different problems requiring different instruments. The people making that mistake are on our side. That is exactly why somebody has to say it.
Should generative AI chatbots be required by law to do each of the following?
| Requirement | Republican | Independent | Democrat |
|---|---|---|---|
| Alert parents to harmful behavior | 90% | 87% | 92% |
| Pop-up warnings on sensitive content | 88% | 86% | 91% |
| Let parents limit topics | 86% | 81% | 86% |
| Let parents see chats | 86% | 80% | 85% |
| Require permission first | 81% | 79% | 84% |
say the amount of computers and tablets their school uses is about right. Only 17% say too much. Of the parents who want a daily cap, just 1% would set it at zero.
want a maximum daily screen-time limit during the school day. Parents want a thermostat, not a switch.
say their school provided no information about its AI policy this year. 57% were never asked for input. Of nine things parents grade their school on, clear tech and AI policy ranks last.
Each lane is a government that made a call about children and AI and wrote it down. The marker sits where the decision landed. Watch the bottom lane.
Say this precisely, because a good analyst will check. No credible source ranks countries on AI education. Stanford's 2026 AI Index describes the United States as "a global leader in ICT-related fields," producing more graduates at associate, bachelor's, master's and PhD level than any country in its sample, while noting Turkey, Brazil and Mexico are growing faster. Read the footnote before somebody else does: that sample excludes India, China and the countries of Africa, and runs on 2023 data. So it is not the rebuttal to our China comparison that it first appears to be, and it is not a ranking either. Nobody has one. The gap is not talent and it is not universities. The gap is that other governments made a decision about children and wrote it into a curriculum, and we did not. What the Index does show at home: 81% of U.S. computer science teachers believe AI belongs in foundational CS education, and fewer than half feel equipped to teach it, 46% in high school, 44% in middle school, 34% in elementary.
That is not a win. And the compression of entry-level work is not a projection, it is already in the payroll data.
average wage premium (61.9%) for jobs requiring AI skills, up from 57% the prior year. The range runs from 16% in government to 118% in consumer markets.
employment of 22 to 25 year olds in the two most AI-exposed occupational quintiles, Nov. 2022 to June 2026. In the least-exposed quintiles it grew about 10%.
the most AI-exposed entry-level roles are seven times more likely to demand traditionally senior skills. Seniorized entry-level roles grew 35% since 2019; the rest fell 10%.
Say the caveat before someone else does. Economists disagree about how much of this is AI. The New York Fed finds no clear AI-driven divergence in job postings. Danish administrative data finds precise nulls on earnings and hours. David Deming attributes the junior-hiring decline mostly to remote work. Our claim does not require winning that fight. Whatever is causing it, the entry-level rung is getting harder to reach, and a child who cannot use these tools is not better positioned under either explanation.
34% have received no guidance at all. The 18% measures formal guidance from school administrators, a lower bar than completing a training program. The two bars come from two Gallup waves a year apart: the 60% describes 2024-25, the 18% early 2026.
Bars scaled to the number of states. The third bar is not a rendering error.
The divide is not access to the tool. Every kid has the tool. Move the slider and watch what changes.
Lower- and middle-income teens also use Character.ai at double the rate of higher-income peers, 14% against 7%.
Belief that AI skills will be essential climbs from 38% of the lowest-income parents to 68% of the highest, and from 44% to 62% by education.
The families with the least access are the least convinced their child will need it. That is how a gap becomes a generation.
Affluent kids are being taught AI. Everyone else is being handed it, unsupervised, and monitored while they use it.
Every claim in this fight, ours included, rests on studies of adults, on cross-sections, and on self-report. Nobody is tracking what AI does to a developing child over time. That is fixable, and fixing it is infrastructure.
EEG study of essay writing. LLM users showed the weakest neural connectivity of three groups, and in session one 83% could not correctly quote their own essay against 11% of the other two groups.
The honest reading: a preprint, not peer reviewed. 54 adults aged 18 to 39, 18 per group. Zero children. The authors publicly ask people to stop saying "dumb," "brain rot," or "damage" about their work.
319 knowledge workers, 936 real tasks. Higher confidence in AI tracked with less critical thinking (β = −0.69); higher confidence in yourself tracked with more (β = 0.26).
The honest reading: peer reviewed, but self-report, and the authors state plainly that it "does not establish causation." Effort shifts from producing to verifying. Adults again.
A Harvard randomized crossover trial of 194 undergraduates found AI tutoring beat in-class active learning by roughly 0.6 to 1.3 standard deviations, in less time. A 35-study meta-analysis puts ChatGPT's effect on learning at g = 0.67.
The honest reading: mostly higher education, expert-built prompts, short horizons. The upside is real and it is also unproven for children.
The largest study of child brain development in America follows 11,880 children from age nine, across 21 sites, on roughly $290 million of NIH funding. It measures six categories of screen media. None of them is artificial intelligence.
In June 2026, in a Q&A for clinicians, the AAP's Center of Excellence on Social Media and Youth Mental Health advised pediatricians to "make AI use a routine intake and check-in topic, the way you'd ask about social media or sleep," and warned that an "avoid all chatbots" message "is likely not realistic." That is guidance, not an AAP policy statement. In May 2026 the Surgeon General told providers to include questions about screen usage in annual well-child visits.
Both told pediatricians to ask. Neither gave them anything to ask with. There are 31 recommended well-child visits from birth to 21 and the periodicity schedule carries no media or technology measure at all.
The machinery is not hypothetical. Autism gets M-CHAT-R/F at 18 and 24 months. Development gets ASQ-3, ten to fifteen minutes, parent-completed, normed on 15,138 children. Adolescent depression gets a USPSTF Grade B and PHQ-A from age 12. Each began as an instrument somebody built and validated.
What we are not claiming, and will say first. There is no evidence today that AI use causes cognitive stunting in children, because nobody has run the study that could show it either way. We are not proposing to diagnose children. We are proposing to build the measurement instrument that would let anyone, including people who disagree with us, find out. Two adoption doors exist: a USPSTF grade, or inclusion in the HRSA-backed Bright Futures periodicity schedule, both of which trigger no-cost-sharing coverage. And a caution we will not hide: the Problematic Media Use Measure was published in 2017 on 632 parents with a short-form reliability of .93, appeared in Psychology of Popular Media Culture in 2019, and is listed among AAP clinician screening tools. It is still not on the periodicity schedule. Building the instrument is the easy half.
We are not alone in asking. Brookings' fourth recommendation calls for research on "AI's influence on students' cognitive, affective, motivational, and emotional states" and for work that is "rigorous, empirical, and longitudinal." The Society for Research in Child Development has asked Congress to direct new funding through NIH, NSF and IES "to produce a stronger evidence base on how, when, and with whom AI affects child and adolescent development over time." Neither names an instrument. That is the piece we intend to put on the table.
A topline tells you what 1,527 parents chose from a list. It cannot tell you what a mother in Anchorage means when she says her son talks to the thing at night. For that you have to go.
Not a survey panel. Structured listening sessions with parents and with young people themselves, in their own cities, in their own words, run through the NPU affiliate network.
Partnered with the team whose January 2026 report called for research that is "rigorous, empirical, and longitudinal" and grounded in student experience. This is that call, answered in the field.
Qualitative field data is what a screening instrument is built out of. You cannot write good items without hearing how families actually describe what is happening in their homes.
From Atlanta to Alaska, from Boston to San Diego, from Chattanooga to San Juan. If a policy about American children cannot survive a room in Anchorage, it is not a policy about American children.
The reach is the point. Anchorage and San Juan are on this list for the same reason Boston and Los Angeles are. A federal policy about American children that was designed only for the places that are easy to fly to is not a policy about American children. Dates and the session schedule follow separately.
Seven titles, including KOSA and COPPA 2.0 as subtitles. Title IV requires chatbots to disclose they are not human and to surface crisis resources when a minor raises suicide. Requires a break advisory after three hours of continuous interaction. Bars a chatbot from claiming to be a licensed professional.
Sec. 213(c)(2): nothing in subsection (a) may be construed to impose a duty of care on a provider of a covered platform. Sec. 402(1)(B): a service cannot be treated as a chatbot provider solely on the basis of a chat function incidental to its primary purpose. Title IV requires only "reasonable policies, practices, and procedures" on sexual exploitation and abuse, gambling, and narcotic drugs, tobacco or alcohol.
Be precise in the room: the bill improved on its way through the House. In the KOSA title the knowledge standard moved from "actual knowledge or willful disregard" to "know or should have known," and Title IV gained that standard for the first time, though the old standard survives in Sec. 232 and Sec. 611. A new Sec. 704 replaced blanket preemption with a savings clause for state laws that give minors greater protection. Two of the four things NPU flagged in March were answered. The duty-of-care removal and the incidental carve-out were not, and those are the two that matter.
| Vehicle | What happened | Status |
|---|---|---|
| Budget reconciliation moratorium, 2025 | Stripped on a 99 to 1 Senate vote. | Failed |
| FY2026 NDAA rider | Omitted from the final text December 7, 2025. | Dropped |
| Executive Order 14365, Dec. 11, 2025 | DOJ AI Litigation Task Force. Commerce identifying conflicting state laws. BEAD broadband funding conditioned on compliance. | In effect |
| DOJ intervention in xAI v. Weiser | Federal government joined a challenge to Colorado's AI Act on April 24, 2026. | Active |
Every legislative version of the preemption push carves out child safety. But EO 14365 protects it only in Sec. 8(b), which governs what the administration recommends to Congress. The litigation task force, the Commerce list and the BEAD funding condition carve out nothing.
S. 1748, KOSA (Blackburn, 76 cosponsors) ordered reported favorably with a substitute amendment Aug. 5, 2026, awaiting floor action. Same day: S. 4199 Youth AI Privacy Act (Markey), S. 4407 CHATBOT Act (Cruz, with Schatz, Curtis, Schiff), S. 5171 Children's Artificial Intelligence Toy Safety Act of 2026 (Duckworth).
COPPA 2.0 (S. 836) passed the Senate by unanimous consent March 5, 2026 and has been held at the House desk since March 16, never referred to committee. It is waiting on floor scheduling, which means it is waiting on political pressure.
14 companion-chatbot laws enacted in 2026 from more than 100 bills introduced. 35 states plus Puerto Rico have issued official K-12 AI guidance. 77 AI-in-education bills filed across 27 states.
Models worth naming: California SB 243, companion protections with a private right of action at $1,000 per violation. Oklahoma SB 1734, bars AI as the primary basis for grading, discipline or promotion, with parental opt-outs. Maryland SB 720, AI coordinators and statewide teacher development. Ohio ORC 3301.24, every district had to adopt an AI policy by July 1, 2026, and neither the statute nor the state model policy requires parent notification.
We have both, and the constituency is the one the bills are named after.
families represented
affiliated organizations
jurisdictions: all 50 states, DC and Puerto Rico
AI and technology polling waves
| Wave | n | What it established |
|---|---|---|
| Oct. 2023 | 1,515 | First AI wave. 62% of parents had heard little or nothing about AI in education. |
| Feb. 2024 | 1,506 | Phones and social media. 83% wanted a minimum age for social accounts. |
| Sept. 2024 | 1,508 | Phones at school. 68% of kids have a phone; 76% of those take it to school. |
| Dec. 2024 | 1,529 | Online safety. 86% backed laws requiring platforms to mitigate harm to minors. |
| Feb. 2026 | 1,511 | AI guardrails. 56% of parents say their child uses a chatbot; 68% of high schoolers. |
| May 2026 | 1,527 | Protected, Not Restricted. The protect-and-prepare demand, measured directly. |
The Traffic Light Framework, plus the tools to act on it: a 12-item advocacy checklist, a letter builder, and scripted questions for principals, boards, superintendents and legislators.
Never. Grading, discipline, IEPs and 504 plans, placement, counseling, crisis intervention, surveillance.
Only with real human review. Data trends, translation, disability accommodations, family communications.
Encouraged. Brainstorming, outlining, scheduling, non-sensitive drafts.
Protected, Not Restricted. Twenty slides built on the May 2026 wave, with party and ideology crosstabs on every guardrail item, the international timeline showing what other countries did while the United States did nothing, and six legislative demands stated plainly at the end.
The governing principle behind both: teaching and learning are human endeavors served by technology, not replaced by technology.
Add the 16-city tour and the screening instrument, and the portfolio covers the full cycle: measure it, hear it, name it, and change the law.
"[A]t this point in its trajectory, the risks of utilizing AI in education overshadow its benefits. This is largely because the risks of AI differ in nature from its benefits—that is, these risks undermine children's foundational development."
An unprepared child is an unprotected child. The risk of getting AI wrong in schools and the risk of keeping it out of schools land on the same kid, and they land hardest on the kid whose parents cannot buy their way around either one.
We do not need Brookings to agree with us, and they do not need to agree with us. They put parents inside the instructional core and called for exactly the longitudinal research we are proposing to instrument. We read their risk finding as a verdict on current conditions, which we are trying to change, not on whether children should be prepared at all. They challenge our optimism. We challenge their sequencing. We intend to keep having that argument in public, including at the Windsor Castle consultation on AI and pedagogy this December, and then we get on a plane together and go to sixteen cities.
Codify the chatbot protections 81 to 90% of parents already want. Parental alerts, content warnings, topic limits, permission.
Schools tell families when AI touches their child, their data or their assignments. 88% want this.
Close the 60 to 18 gap. Funded, required, and not optional.
Not an elective in the districts that can afford it. A requirement everywhere, starting with the districts that have the least.
No sale. No model training on children. Vendor vetting with teeth. Active consent.
A ban is not a plan. It regulates the assistant and leaves the companion and the surveillance system alone.
Placeholder. The four funding lanes below are the shape of the ask. The figures are not filled in yet and nothing here is an NPU number.
Continue the Echelon waves. Add the questions nobody else is asking: what parents know, what schools disclose, and how the divide is moving. [ $X over Y months ]
Structured listening with parents and youth in sixteen cities from Anchorage to San Juan, run through the affiliate network, with the transcripts feeding item development. [ $X over Y months ]
Build and validate a cognitive and developmental AI screen for well-child visits, with clinical partners, and take it to USPSTF or Bright Futures. [ $X over Y months ]
Senate pressure on KOSA and COPPA 2.0, defense of the state child-safety carve-out, and MyKidMyData in the hands of parent leaders in all 50 states, in English and Spanish. [ $X over Y months ]
You do not need to understand how a language model works to ask your school a hard question. These are built for you, and they are free.