Executive Summary

On a single day, October 14, 2026, attorneys general in several US states sued OpenAI. This is not one case. Arizona filed in Maricopa County Superior Court, Arkansas in Faulkner County Circuit Court, Idaho in the Fourth Judicial District in Ada County, Iowa in Polk County District Court, Montana in the First Judicial District in Lewis and Clark County, and West Virginia in Mason County Circuit Court — each on its own. Kansas was reported in Shawnee County and Nebraska in Buffalo County; for those two we could not obtain the complaints, so we name the courts on the strength of the reporting and go no further. The defendants are the same in every complaint we read: eight OpenAI entities, OpenAI Global, LLC among them, named as one bundle.

One correction belongs here first. Even the count of states differs from announcement to announcement. The Arkansas attorney general's press release lists the states that filed similar suits against OpenAI that day and counts nine, Utah included. The Montana, Iowa and Nebraska announcements and the wire stories list eight, and the Utah attorney general's newsroom carries no announcement of an OpenAI filing. What is confirmed is that the lists disagree, which is why the rest of this piece calls the states by name instead of counting them.

Set the six complaints side by side and the paragraphs overlap. Some sentences are identical down to the punctuation. The reason is in the signature blocks. All six we obtained list the same private firm, with offices in New York and Los Angeles, as co-counsel, and the individual attorneys' names match too. State attorneys general retaining outside firms for large cases is common practice, so the overlapping paragraphs are best read as traces of a common draft. What matters for §3 is that the common draft folds into a different shape under each state's law.

Nor do the counts collapse into consumer protection law alone. The Montana complaint builds six branches — two consumer protection counts, a privacy act count, and negligence, design defect and failure to warn. The Arkansas complaint stacks four deceptive trade practices counts, three counts under the Arkansas Children and Teens' Online Privacy Protection Act, and five common-law counts on top. West Virginia pairs its consumer credit and protection act with public nuisance, pulling into its prayer the cost of restoring schools' digital safety education and youth crisis response capacity. We do not dig into that doctrine itself. Only this much is worth noting: calling this a consumer protection case already narrows the frame once.

OpenAI's answer belongs in the same place. Company spokesperson Drew Fusatteri gave several outlets the same statement.

Everything set out below is what the plaintiffs wrote in their complaints, not what any court has decided. The Montana attorney general's notice says as much itself. That is why this article attaches, to each allegation, which complaint and which paragraph it came from. None of the cases has reached an answer.

The first thing to look at in these suits is the dates. The clauses aimed at minors and at design are not old law. They are statutes that switched on one by one between last autumn and this summer. The minor provisions of Montana's privacy act took effect on October 1, 2025. Nebraska's Age-Appropriate Online Design Code Act took effect on January 1, 2026, and the attorney general's authority to seek civil penalties under it began on July 1. Arkansas's Children and Teens' Online Privacy Protection Act switched on that same July 1. And Nebraska's enumerated list of design features was amended once more on July 17 — the latest-arriving clause in this entire case.

Montana hides one more date. The 2025 amendment that added the minor provisions also removed the cure period. A cure period is the step that requires the attorney general to give notice and an opportunity to fix the problem before suing. Twelve and a half months after that step disappeared, the complaint arrived. We did not, however, find any place where the complaints rest on this. Nothing in the six we read contains language referring to a cure period.

Line the dates up and each clause is a different age. Montana's is the oldest of them. As of the filing date, Nebraska's design code act had been in force nine and a half months, and Arkansas's act and the Nebraska attorney general's penalty authority three and a half. The last clause to switch on was not yet three months old.

▲ The order in which the clauses switched on. What to watch is that the period each count covers differs by clause. The consumer protection counts reach back to the November 2022 release of ChatGPT, but the privacy and design counts attach only from the day those laws took effect. Nebraska's statutory dates and language were confirmed through law firm commentary.

That difference carries through the whole piece. Inside one complaint, some counts cover nearly four years and others cover three months. Which means summarizing this as a case that fines four years of data gets it wrong. The sharper the clause, the shorter the window it covers. Put the other way around: what is new in this case is not the dollar figure. It is the clause.

The word design splits three ways in this section. How the model was trained and tuned; what features were bolted onto the product; and what gets collected by default. What the complaints plead is the middle and the last. There are passages describing a model tuned to be agreeable, but what is pleaded under a clause is features and defaults. When this article writes design below, it means those two.

3.1The Montana statute talks about time spent, directly

"Keeping young users engaged longer," the phrase the coverage carried, reads like a reporter's metaphor. Open Montana's privacy act and the phrase is sitting there in the statute.

Montana's complaint at ¶258(c) pleads precisely that clause. But reading that one paragraph and summarizing it as "Montana went after design" narrows the frame. Paragraphs ¶257 through ¶259 lay out nine branches of privacy act violation: processing without consent, collecting precise geolocation, failing to use reasonable care to avoid heightened risk to minors, the design feature, failing to maintain data security practices, failing to provide a privacy notice, retaining more than necessary, collecting for undisclosed purposes, and failing to conduct a data protection impact assessment. Not doing the assessment is itself one of the boxes. And one of those nine, ¶258(d), is a box about collecting the sensitive data of adults before obtaining consent. The summary "a children's privacy case" narrows once more here.

What makes the design clause interesting is its conditional. What is forbidden is not the design feature itself but using it without consent. So obtain consent and the problem goes away? The same statute closes that road in advance. Montana's complaint at ¶252 says the privacy act does not count as consent any agreement obtained through dark patterns. Consent extracted by design is void because of the design. The complaint does not, however, plead dark patterns as a separate count. ¶252 is a paragraph introducing the law, and we read it no further.

3.2Nebraska rewrote its list three months earlier

Nebraska's Age-Appropriate Online Design Code Act goes one step further. It names the covered design features and enumerates them. When first enacted, that list held five: infinite scroll, rewards for frequency or duration of visits, notifications and push, in-game purchases, and appearance-altering filters. The vocabulary of social media and games. A chatbot has no infinite scroll, so it seemed natural to expect a fight over whether chatbots fell inside the list at all.

It went the other way. An amendment passed in April 2026 and effective July 17 grew the list to eleven and added this subsection.

Sitting next to infinite scroll and push notifications, that subsection is hard to read as pointing at anything but a chatbot. And 89 days separate the day it switched on from the day of the filings. We will not build a bridge between those two facts here. The Nebraska attorney general's announcement says the design code act was pleaded alongside the consumer protection act and the deceptive trade practices act. It is equally a fact that the design code act had carried a chatbot-shaped subsection for three months before the filing. But we never obtained the Nebraska complaint itself. Whether the complaint picks up that subsection we could not confirm, and what we could not confirm is written down as unconfirmed. Setting the two facts side by side is as far as this article can go.

It is also hard to treat as one state's outlier clause. A bill with nearly identical language is pending in New Jersey. Nebraska's is law and New Jersey's is a pending bill, though, so the same sentence does not yet have force in two states. Closer to the facts: a model bill is in the middle of being copied.

Where the coverage noted this, it stopped at two states — Arkansas pleading its children and teens online privacy protection act, Montana pleading its privacy act, each alongside consumer protection. Nebraska's design code act adds one more seat at that table.

3.3One fact, translated three ways

What about the remaining states? Arizona has no comprehensive state privacy act. So the Arizona complaint translates the same fact into non-disclosure. ¶240 says the company concealed that it collects personal information from minors, and ¶243 says ChatGPT encouraged minors to disclose personal information to it alone and at times discouraged them from telling anyone else, parents included. The same conduct is pleaded in Montana as a privacy act violation, in Nebraska as a design act violation, and in Arizona, Idaho, Iowa and West Virginia as the deception of not telling. One fact, translated three ways.

▲ One fact translates into three different legal theories depending on the state. Montana pleads a design-feature clause, Nebraska a named feature on its design code list, and the four states without a comprehensive privacy act plead non-disclosure as deception.

Iowa shows where the translation runs out. In its body the Iowa complaint cites its own state's consumer data protection act and says that the absence of a gatekeeping mechanism in ChatGPT is not merely a problem but unlawful. Yet the counts plead only the consumer fraud act. The law it invoked in the body was never raised to a count. The complaint does not say why, and this article does not speculate.

3.4The same design carries a different price tag in each state

When the translation splits, so does the price. The civil penalties the complaints request run from $5,000 to $50,000 per violation. Same company, same product, same feature — cross a state line and the price tag multiplies by ten. The bars below give the per-violation maximum under the statute each complaint invokes.

▲ Each bar counts a different "one." Because each state defines a single violation differently, as §5 shows, these bars cannot be added together or multiplied against each other. Nebraska is marked in orange not because the figure is the largest but because its route is different: the design code act carries no penalty provision of its own, and a violation is deemed a separate violation of the deceptive trade practices act, borrowing that statute's provision. Kansas is reported as $10,000 per violation plus an additional $10,000 where the victim is a protected consumer, but we could not obtain that complaint and left it off the chart. Montana appears on two rows because it seeks penalties under two statutes.

More important than the length of each bar is what each bar counts. And that question is §5 of this article.

In suits like these, what plaintiffs usually struggle with most is proving knowledge. If the company says it did not know the user was a minor, knowledge has to be proven from outside. These complaints do not look outside the company for it. They pick up the very machinery the company built to protect children, and the figures the company published of its own accord.

4.1A three-column age label becomes the evidence

The Arkansas complaint at ¶194 says OpenAI holds data classifying users by age. That classification has three columns.

Those three columns are not a paper taxonomy. On August 18, 2026, two months before the filings, OpenAI shipped a teen version of ChatGPT that carves out 13- to 17-year-olds, and the route into it runs through exactly this classification. Users are not made to verify age up front; minors are picked out from the age they entered, account history and usage patterns and moved over automatically, with a path to confirm age by selfie opening only when an adult is caught by mistake — which is where the third column sits.

What happens here is simple. The label built to screen children out becomes the evidence that the children were recognized. Building an age-prediction model satisfies the actual knowledge element on its own. The diagram below is that route.

▲ A diagram of the structure of ¶194 of the Arkansas complaint. The three boxes on the left are machinery built to protect children; the box on the right is where that machinery is put to work in the complaint. The statutory language puts actual knowledge and willful avoidance of knowledge in the same box.

4.2When an evaluation score becomes a disclosure item

The second scene cuts deeper. On October 27, 2025, OpenAI measured and published how its models respond in sensitive conversations. The post reported, as numbers, the share of desirable responses across three categories: self-harm and suicide, emotional reliance, and signs of psychosis or mania. The West Virginia complaint at ¶239 says that not having disclosed those numbers and the misclassification rate of age prediction until then was a half-truth, and a half-truth is deception. This is the moment an engineering document crosses over into a consumer disclosure document.

Which makes carrying the numbers across accurately matter. The three figures the complaint cites are 77 percent for self-harm, 50 percent for emotional reliance and 27 percent for psychosis and mania, and all three are values for the model that has since been replaced. The complaint does not hide that. It attaches the qualifier that this was the model deployed at the time. In the same announcement the new model's values rise to 91, 97 and 92 percent. So drop the qualifier and write "ChatGPT's self-harm compliance rate is 77 percent" and the sentence becomes false in that instant. In fairness, both sets belong side by side.

▲ OpenAI's own evaluation, published October 27, 2025. The grey bars are the values cited at ¶239 of the West Virginia complaint; the orange bars are the post-replacement values from the same announcement, which the complaint does not write down at that spot. The denominator is conversations in the evaluation set, not users. For the record, the number 77 is used once more in that same announcement with a different meaning — a statement that agreement between expert graders ran in the 71 to 77 percent range, a figure that distorts the source the moment it is quoted as a single value.

The complaint's account does not depart from the source. But underneath it there is a paradox. Numbers published voluntarily are cited as the basis for an allegation about why they were not disclosed sooner. Had they never been measured, there would have been no figures to cite. Nor is not measuring the safer side. One of the nine branches the Montana complaint pleads as privacy act violations is the failure to conduct a data protection impact assessment. Measure, and the value becomes a disclosure duty; do not measure, and the not-measuring becomes a failure of care.

4.3The 12 percent is not a figure the company published

One number in the complaints needs carrying across with particular care: the roughly 12 percent rate at which the age-prediction system misclassified minors as adults. The source of that figure is not an OpenAI announcement but a Wall Street Journal report, and the West Virginia complaint at ¶111 cites it with a "according to reporting" qualifier attached. Yet when the figure reappears in the same complaint at ¶245(h), in the list establishing the knowledge element, the qualifier is gone. Inside a legal document that is a back-reference to an earlier paragraph, but to a reader who only reaches the later passage it looks like a figure the company disclosed. Which is why this article says it at the start and says it again. The 12 percent rests on newspaper reporting.

Other figures from the same announcement only mean something with their denominators attached. The share of weekly active users having conversations with explicit indicators of suicidal planning or intent is 0.15 percent; measure the same category by messages and it is 0.05 percent. A number that counts users and a number that counts messages are different, and swapping one for the other tells a different story. The complaints, for their part, do not switch the denominators when they cite these figures.

If the sections so far asked what is being challenged, this section's question is what is being counted. The Montana complaint lays out that arithmetic across two paragraphs. The order matters.

¶260 builds the bridge first. It is the sentence saying a privacy act violation is itself an unfair or deceptive act prohibited by the consumer protection act, with the enabling provision given in parentheses. Having joined the two statutes, ¶261 sets the multiplication on top of that bridge.

One act of collection becomes one count under each of two statutes. And the penalties are requested per violation. So in this case the multiplier is the log count. What sets the amount is not what the user typed but what the company gathered, and how many times.

5.1What counts as one violation splits by state

What a single violation is, though, is written differently in each state. Lay all six complaints against each other and the definition splits four ways. Montana, as we saw, counts one act of collection as one violation. Idaho and West Virginia count each deceptive representation or omission multiplied by each transaction it produced. The relevant sentences in those two complaints are nearly identical once you swap the state names. Arizona counts one unfair, false or deceptive act as one, and Iowa counts one misrepresentation as one.

None of the four definitions counts by messages or by users. And more importantly, not one of the complaints states how many of that unit there actually were. Leaving the amount out is not an omission; it is the part a trier of fact decides at trial. The upshot is that this case has no total. Which means any sentence estimating the size of the penalty by multiplying a complaint's unit price by some count is a guess. Nor will we make one here.

Before a violation can be counted, jurisdiction has to stand up. Free ChatGPT users pay nothing — how is that a consumer transaction? The West Virginia complaint fills that gap at ¶233. Free use is a transaction, it says, because the company receives something of value in exchange, and that something includes the user's personal information, the conversations used to train models, and ad exposure. To put consumer law onto a free service you have to characterize the data as consideration. The complaint does exactly that.

Since advertising has come up, one boundary needs drawing. ChatGPT has carried ads since February 2026, and the Arizona complaint says the personal information collected feeds advertising revenue. OpenAI, for its part, states that it does not serve ads to accounts identified as under 18. Not showing ads to minors and not collecting minors' data are two different statements. What this case contests is the latter. And no complaint says OpenAI sold minors' personal information. The word "sell" does appear in Iowa's prayer for relief, but it is the language of an injunction asking that collecting or selling without age verification and consent stop going forward.

5.2Delete and do not delete point at the same store

Move to the prayers for relief and two requests to delete data appear. Their scopes differ. Montana asks for deletion of personal information obtained in violation of the privacy act. That is a narrow set, pickable only by following lineage. Iowa asks for deletion and destruction of all data of Iowa residents obtained before judgment. No qualifier. We do not decide which is the heavier request. Iowa did not raise a privacy act count, so it may simply have had no legal hook on which to hang an "obtained in violation of" qualifier. We confirm only that the scopes differ.

And an order running in the opposite direction already sits on the same store at the same company. That is the log preservation order out of the copyright litigation. The duty to keep preserving logs going forward ended in October 2025, but logs already segregated and preserved before that cutoff remain under preservation. The history is in the piece Pebblous wrote in July, "New York Times Seeks Sanctions Over OpenAI's Deleted Chatbot Logs", so we will not repeat it here. All this section needs is the direction.

▲ What is to be preserved is output logs that would otherwise have been deleted; what is to be deleted is minors' personal information obtained in violation, or all of one state's residents' data. How far the two sets actually overlap is written neither in the complaints nor in the order. So the overlap is left as a question mark. There is also no basis yet for asserting that they collide head-on.

Say an order to delete comes down. What does having deleted look like? On the research side this question is in worse shape. The recent survey literature holds that the techniques for stripping particular data out of a trained model are further along than the means of verifying that it was stripped. Verification has a circularity in it. Confirmation requires comparing against a model retrained from scratch without that data — and had retraining been cheap, the removal techniques would not have been needed in the first place. On top of that, the party whose data it was cannot look inside the deletion process. This is not a claim that it is impossible. What this article can write down is that a way to confirm deletion from the outside has not been settled. Why the route of obtaining consent after the fact is closed was covered in our report on Canada's privacy law.

One day before the attorneys general filed, a class action was docketed in the Northern District of California. It is a case against the OpenAI Foundation, and it contests different ground. This one challenges the training of models on minors' data as such, and asks for the share of the model's value that data contributed. The state suits write training use in only as factual background and never raise it to a count. It passes by as the West Virginia complaint explains what free use buys, and as the Montana complaint notes that conversations are used in training.

So the same data gets priced two ways. The states ask for a civil penalty per violation, and the class action asks for a share of the value that data created. The first prices data by how many times it was counted; the second by what it left behind in the model. The two methods demand different records. The first asks when you gathered what, and how often; the second asks which of it went into which model. Not many organizations can answer both.

This article does not mix the figures or the remedies of the two cases. They are separate cases in separate courts. For the same reason, three more things need keeping apart. The suit Florida filed separately in June this year is not among the October 14 filings. The wrongful-death actions brought by bereaved families are a different kind of case. And when Utah sued Snap in June this year it included counts about chatbot features, which makes it the immediate precedent in the line of state attorneys general reaching chatbots with consumer law.

On the legal fate of chatbot conversation records, Pebblous covered the Character.AI bankruptcy two days ago. That one is about conversation records treated as an asset when a company collapses; this one is about those records becoming the multiplier on a penalty while the company is running perfectly well. Call it a third way of pricing the same data.

Why Pebblous cares

Records kept for ourselves are going outside

What Pebblous does is judge whether data is in a condition fit for AI, and leave that judgment as a record someone can read later. What these suits change is who that record is for. Until now a data quality record was an inward-facing document. Are the labels right, are there duplicates, is the distribution skewed. The record these complaints demand is an outward-facing document. What was collected from minor users and when, what form the consent took, how long it was held, and which of it was obtained in violation.

Montana's prayer for relief goes straight at that point. If an order comes down to delete personal information obtained in violation, organizations that can pick out what to delete part from those that cannot. That difference is lineage. The same list of requested relief holds one more item that is neither money nor an injunction: a demand that if a model or agent escapes its intended controls, the attorney general be notified immediately. That is also where a recordkeeping duty moves from after-the-fact penalties to standing reporting.

Liability appears where the three layers come apart

This article has drawn out three distinctions. The first is what was collected. Which fields, under which defaults, from whom. The second is what is written down as known. The age the user stated, the age the model predicted, the age that was verified. The Arkansas complaint offers the mere existence of those three columns as evidence of knowing conduct. The third is what was disclosed. Evaluation scores, misclassification rates, retention periods. The West Virginia complaint says what went undisclosed was a half-truth, and that this is deception.

Every pipeline that accumulates AI-generated labels, predictions and scores has those same three layers. Most organizations record only the first, leave the second inside the model, and never ship the third at all. The moment the three layers come apart is the place where liability appears. A value you measured turns into an item you have to disclose; a value you never measured lands in the box marked failure of due care. Neither side is the safe one. So the choice is not whether to measure but in what form to keep what you measured.

What you can count in your own organization today

The first question is this. Can you count the records of minor users separately? If you cannot, the answer is not "there are none" but "we do not know," and not knowing sits right up against the statutory language these complaints invoke. Montana's clause puts actual knowledge and willful avoidance of knowledge in the same box. It is also worth seeing that the answer changes with how you ask. Ask by survey and 58 percent of children aged 9 to 12 say they have used a chatbot; measure through a monitoring app installed on the child's device and 21 percent of 10- to 12-year-olds leave an access record. Neither is wrong. Asking whether someone has tried it and reading the device log are different questions. Teen usage itself was covered separately in the Pebblous story from July.

Second, what will you define as one instance? A message, a session, a user, an act of collection? The complaints split it four ways, and none of them counted the actual number. That it gets multiplied without being defined is the structure of this case. Third, when an order to delete arrives, can you select the targets? Down to the share that has worked its way into aggregate metrics, caches, evaluation sets and trained weights. And how will you show that it was deleted? Fourth, what do you do when it overlaps an order not to delete? A preservation order from a different case can land on the same logs. Whether your record structure can obey both orders at once is something you can only check before the orders arrive.

Korea has no clause aimed at that spot yet

If you run a service in Korea you are already handling guardian consent for children under 14. The Personal Information Protection Commission's guideline on children's and adolescents' personal information also recommends setting defaults at a high level. That guideline, however, is advisory material for self-assessment, while Nebraska's design act compels the same ground by statute and attaches a per-violation penalty. And since February 2025 the e-commerce act has prohibited six types of dark pattern: hidden renewal, drip pricing, pre-checked options, false hierarchy, obstructing cancellation and withdrawal, and repeated interference. All six aim at design that makes you spend more money. There is as yet no clause aimed at design that makes you stay longer. The shutdown law, which regulated time spent directly, was repealed in August 2021. What those two US states have pleaded here is precisely that empty spot. The while it stays empty is the time to prepare.

The same measure, turned on ourselves first

Where Pebblous stands is on the side of counting what happened to data one instance at a time, and leaving it in a form readable both inside and outside. We are not in the race to build bigger models. And turning the same measure on ourselves produces an empty field. Pebblous's DataClinic outputs still have no formal field stating where this record came from and when, and whether it can be deleted now. We have never separated and counted the records of minor users either. This being an article that counts someone else's empty fields, writing down our own comes first.

References

Primary — the complaints (downloaded and read directly)

• 1.State of Montana v. OpenAI Global, LLC et al., First Judicial District Court, Lewis and Clark County, 2026-10-14 — complaint PDF. ¶252, ¶¶257–259, ¶260, ¶261 and the prayer for relief cited in the body come from here.
• 2.State of Arkansas v. OpenAI Global, LLC et al., Faulkner County Circuit Court, No. 23cv-26-2984, 2026-10-14 — complaint PDF. ⚠️ It is a scan, so the causes of action were read through optical character recognition. The ¶194 verbatim in §4.1 comes from here.
• 3.State of West Virginia v. OpenAI Global, LLC et al., Mason County Circuit Court, 2026-10-14 — complaint PDF. ¶111, ¶233, ¶239, ¶245, ¶249.
• 4.State of Idaho v. OpenAI Global, LLC et al., Fourth Judicial District, Ada County, 2026-10-14 — complaint PDF. ¶237 and prayer for relief B.
• 5.State of Iowa v. OpenAI Global, LLC et al., Polk County District Court, 2026-10-14 — petition PDF. ¶241 and prayers for relief C, D and H.
• 6.State of Arizona v. OpenAI Global, LLC et al., Maricopa County Superior Court, 2026-10-14 — downloaded via this collection of the complaints. ¶57, ¶232, ¶240, ¶243.
• 7.⛔ The Kansas complaint (Shawnee County District Court) and the Nebraska complaint (Buffalo County District Court, No. D09CI260000488) could not be obtained. The Nebraska attorney general's server was unreachable from this working environment. That is why the account of those two states stops at the court and the statutes pleaded.

Primary — statutes

• 8.Mont. Code Ann. § 30-14-2811(2)(b) — the design feature clause. In the same act, § 30-14-2802(6) excludes from consent any agreement obtained through dark patterns, and § 30-14-2820(1) joins a privacy act violation to a consumer protection act violation. The 2025 amendment (SB 297) brought in the minor provisions and removed the cure period.
• 9.Neb. Rev. Stat. § 87-1302 — the definition of covered design features in the Age-Appropriate Online Design Code Act. The 2026 amendment (LB 838, effective 2026-07-17) grew the list to eleven and added subsection (k). Penalties run through § 87-1309. ⚠️ Both the section numbers and the language were confirmed through law firm commentary and legal databases; the Nebraska legislature's own server could not be reached.
• 10.Ark. Code — the Children and Teens' Online Privacy Protection Act (Act 952, effective 2026-07-01) and the Deceptive Trade Practices Act · Iowa Code § 714.16(7), § 714.16A, § 715D.4(2) · Idaho Code § 48-606(1)(e) · A.R.S. § 44-1531(A) · W. Va. Code § 46A-6. These are the basis for the unit prices in the §3.4 chart.

Primary — official and company announcements

• 11.Arkansas Attorney General, press release (2026-10-14) — the only announcement that counts nine states, Utah included, as filing that day.
• 12.Montana Attorney General, press release (2026-10-14) · Nebraska Attorney General, press release (2026-10-14, via a reprint) — the Nebraska one enumerates all three statutes: consumer protection, deceptive trade practices and the design code act.
• 13.OpenAI, Strengthening ChatGPT's responses in sensitive conversations (2025-10-27) — the source for the §4.2 chart. ⚠️ The page itself was blocked, so the figures were checked against the Wayback Machine capture.
• 14.OpenAI Help Center, Ads in ChatGPT — the policy of not showing ads to accounts identified as under 18. · OpenAI, A milestone in expanding access to AI (2026-08-31) — primary source for the advertising revenue figure. · OpenAI, launch of ChatGPT for teens (2026-08-18) — basis for the automatic routing described in §4.1. ⚠️ The company's own announcement pages were unreachable, so these were confirmed by cross-checking multiple independent secondary reports.

Secondary — scholarship, surveys and reporting

• 15.Sam Schechner & Georgia Wells, Wall Street Journal (2026-03-15) — the sole source for the roughly 12 percent age-prediction misclassification figure. It is not a number OpenAI published.
• 16.Kristen Soltis Anderson et al., The Common Sense Media Census: AI Use by Tweens and Teens, 2026 — a survey of 1,204 respondents aged 9 to 17 (2026-03). · Anne J. Maheux et al., Generative Artificial Intelligence Applications Use Among US Youth, JAMA Netw Open 9(2):e2556631 (2026-02-02), doi:10.1001/jamanetworkopen.2025.56631 — passive sensing of device logs across 6,488 households. The gap between the two is a difference of method, not of age band.
• 17.Towards Reliable Forgetting: A Survey on Machine Unlearning Verification, ACM Computing Surveys 58(12), 2026-05-15 · Verification of Machine Unlearning is Fragile, arXiv:2408.00929 — the verification gap in §5.2.
• 18.Personal Information Protection Commission (Korea), Guideline on the Protection of Children's and Adolescents' Personal Information (2022) · Korea Fair Trade Commission, the six dark pattern types under the amended E-Commerce Act (in force 2026-02-14). ⚠️ The statutory text itself was not read verbatim, so the law and the guideline are referred to by name only.

Adjacent cases and adjacent Pebblous articles

• 19.Doe v. OpenAI Foundation, N.D. Cal. (2026-10-13) — the separate class action · New York Times v. OpenAI, S.D.N.Y. — the log preservation order · Florida's filing (2026-06) · Utah v. Snap (2026-06).
• 20.New York Times Seeks Sanctions Over OpenAI's Deleted Chatbot Logs (2026-07-13) · Consent Cannot Be Obtained After the Fact — Canada's New Line (2026-05-29) · Character.AI Is in Bankruptcy. Where Do Your Chat Logs Go? (2026-10-14) · US Teens Hid It From People and Confided Only in the AI (2026-07-03).