Still In Charge · Insights

What the Agreement Actually Says

By Thomas Tornatore, Founder · September 13, 2026

Fifteen minutes of your attention, spent well. Three places to look in any AI product's paperwork, shown on documents anyone can open.

I wrote last week that if a product is good for your child, it survives fifteen minutes of your attention. While that is true, I left the next logical question unanswered. Fifteen minutes doing what?

Before I dive in, a quick word: Still In Charge provides public-document education, not legal advice. I am not a lawyer. This article identifies and explains language in public documents; it does not determine whether any provision is valid, enforceable, or applicable to a particular person or situation.

Among American kids age 9 to 17 who use AI at all, 85% use it for help with schoolwork or homework, and 38% have used an AI-powered learning tool.

Here is the "what."

This week I read the public agreements behind three AI study tools. Everything below is quoted from a document anyone can open. I am not telling you what these companies are trying to do. I am telling you what their own paperwork says, and showing you the three places to look, so you can do this yourself with whatever your kid is actually using.

One: who operates it

Gauth is a homework app widely used by American teenagers. Its US privacy policy, last updated January 22 of this year, names the operator in the last sentence of its opening paragraph: "The Platform is operated by TikTok USDS Joint Venture LLC ("we" or "us")."

That is not a gotcha. It is the document's own opening paragraph. But it is the kind of thing that never appears on the app store listing or the home page, and it changes how the rest of the document reads, because later sections describe sharing with a named business partner and cite Executive Order 14352 by number. That order is a real executive order, it is public, and it is short enough to read. It was signed September 25, 2025, titled "Saving TikTok While Protecting National Security," and it is the order that cleared the deal creating that United States joint venture in the first place.

One more thing I noticed, which cost me an additional ninety seconds. Legal documents give a long company name a short nickname in parentheses, then use the nickname from there on. Watch what happens across these two.

The privacy policy: "We also share information with TT Commerce & Global Services LLC and its affiliates ("TT Commerce & Global Services")."

The terms of service, last updated the same day: "TikTok USDS Joint Venture works with service providers and business partners, such as BD TikTok USA LLC and its affiliates (hereinafter referred to as "TT Commerce & Global Services")."

Same nickname. Two different companies. One date. From that point on, every time either document uses that nickname, you cannot be certain which company it means. I do not know which is correct. Neither, presumably, does a parent.

Two: where the promises stop

Same document, further down, with the first four words in bold: "No Sale or Sharing. The Platform does not sell your personal information or share your personal information with third parties for purposes of cross-context behavioral advertising where restricted by applicable law."

Read to the end of the sentence. Most people read the bold lead-in and the first eight words and stop, which is what a bold lead-in is for.

The sentence makes two promises. The first is broad: the Platform does not sell your personal information. The second is much narrower: it does not share that information with other companies for "cross-context behavioral advertising," which is the industry's term for tracking a person across unrelated apps and websites in order to aim ads at them.

Then it closes on five words positioned after both statements: "where restricted by applicable law." Nothing in the sentence tells you how far back they reach.

Which promise do those five words qualify? If they attach only to the sharing half, what stands in front of them is a flat, unqualified promise not to sell. If they attach to both halves, the promise not to sell reaches exactly as far as the places where selling is already illegal, which is a description of compliance rather than a commitment.

I am not going to tell you which reading is correct, because I do not know, and neither does anyone who has not argued it in front of a judge. That is the whole point. A promise you cannot pin down is not one you can lean on. And you did not need a lawyer to notice the qualifier was sitting there. You needed to finish the sentence.

Three: what your child's work becomes

This is the one that stopped me, and it is the hardest of the three. I am going to work through it more slowly than your fifteen minutes will require, so you can see what it looks like done properly. Once you have watched it happen once, finding it in the next document takes about a minute.

In Gauth's list of what it uses information for, one bullet reads: "To train and improve technology, such as our machine learning models and algorithms. We do not use any data obtained through Google Workspace APIs to develop, improve, or train generalized AI or machine learning models."

One bullet. Two sentences.

The first says your child's work trains the models. To know what "your child's work" covers you have to read across two separate documents. The privacy policy's list of what it collects includes "your written prompts, uploaded files, and voice recordings." The terms of service spell out that user content includes "any text, math problems, answers, photographs, data and information." For a homework app, that means the photo of the worksheet.

The second sentence says your child's Google Workspace data does not train the models, because Google's API terms require that restriction. The protection exists precisely where a third-party company had the contractual power to demand it, and not where only a parent was watching.

The terms of service go further than the privacy policy does. Submit user content and you grant a "non-exclusive, worldwide, unlimited, irrevocable, royalty-free, fully sublicensable (through multiple tiers) and fully transferable" right to use your child's input, including to "modify, sell, offer for sale, create derivative works based upon and otherwise use" it.

Three of those terms carry most of the weight. Irrevocable means you cannot take it back. Fully transferable means the company can hand the right to someone else. Fully sublicensable through multiple tiers means that someone else can hand it on again, and so can the next one after that. A separate sentence in the following paragraph reads: "To the fullest extent permitted by law, you waive any moral rights you may have in any User Content you submit." Apply the rule to that one too. It has a qualifier at each end, the same as the license. Moral rights in the United States are narrow and cover mostly fine art, so it is doubtful a phone photo of a worksheet carries any at all. The sentence may be asking you to give up nothing. It is written to work either way.

The thing being described is a photograph of your kid's math worksheet.

I told you to read to the end of the sentence. That applies to what I quote, too. The license I just showed you has a qualifier at each end, and I am going to read both of them rather than hope you do not notice them. It opens with "As permitted by applicable laws." It closes with "in order to provide the Services to users and develop the Services and similar products." Read them carefully and they do less than they appear to. The opener is not a promise; it is a statement that the company intends to obey the law, which was already true. The closer sounds like a real limit until you notice that "similar products" appears exactly once in the entire terms of service, in that sentence, and is never defined. It reads like a fence. Nothing is on the other side of it. But it is there, in the sentence, and if I had quoted the license without it, I would have been doing the exact thing this issue exists to warn you about.

What a better document looks like

Khan Academy's privacy policy, updated January 30, eight days after Gauth's, lists five privacy principles at the top. The second one reads: "We do not sell your personal information to third parties. We established Khan Academy as a nonprofit organization so that our mission of education and your trust will not be in conflict with a for-profit motive." Read it again and look for the qualifier. There is nothing after "for-profit motive."

It also states that, for the AI features it provides to schools, which include Khanmigo, its AI teacher tools, and Writing Coach, its service agreements with the companies supplying those models "do not permit the providers to use Input data to train their models." That is a genuine commitment, but it is a bounded one, and the boundary in this case is the word schools. It sits inside a section called Schools and Student Use. It is not a blanket statement about everyone who opens Khanmigo at the kitchen table.

Another program, MagicSchool, effective date March 19, runs a plain list under the words "We do not:" and one line of it is: "Use Student Data to train, fine-tune, or improve artificial intelligence or machine learning models, including large language models, or permit any third-party AI provider to do so."

There is a qualifier above that list too, and you should not skip it. The paragraph introducing the list says the commitments "reflect MagicSchool's approach to Student Data where applicable." Where applicable is doing the same job that where restricted by applicable law was doing in Gauth's policy. It is a smaller hedge in a much plainer document, but it is the same shape, and you start to see the same pattern repeating.

MagicSchool also tells you where the promise stops. A paragraph directly underneath the list says those commitments "apply only to Student Data processed by MagicSchool as a service provider or processor on behalf of Schools or other Customers and do not alter this Privacy Policy's terms governing personal information processed by MagicSchool as a controller."

In plain terms: MagicSchool wears two hats. When a school hires it, MagicSchool is working on the school's behalf, and the promises above apply. When MagicSchool is acting for itself, a different part of the policy governs, and those promises do not apply. The company told you which hat it is wearing and when.

That is a real limit, and it is one you can read for yourself. A good agreement is not one that promises you everything. It is one where you can find the promises and see where they stop.

The fifteen minutes

Three questions, in order, answerable from documents that are already public.

Who is named as the operator. Not the brand on the app icon. The entity in the policy.

Where the promises stop. Read past the bold lead-in to the end of the sentence, and watch for every "where," "except," and "to the extent."

What happens to the work your child puts in. Search the terms for "train" and for "license," then read what those paragraphs actually cover, start to finish.

None of this takes a phone call, a lawyer, or a fight. These companies wrote the documents, posted them, and can be held to them. The one thing they cannot do is read them for you.

What the checking turned up

A few more things: about these documents, about my sources, and about what I got wrong along the way. You should be able to check all of it.

Khan Academy's last-updated date is invisible on a phone. On a computer it sits right under the headline. On any screen narrower than 767 pixels it disappears, because of an ordinary responsive design rule. Open that policy on the device most parents actually use and you cannot see when it was last changed.

Gauth's terms contain no arbitration clause and no class action waiver, which is better than a great many consumer agreements and worth pointing out. Plenty of them take away your right to go to court, and your right to join with other people in a single case. Gauth's do not. They do require an informal resolution process first, require that any lawsuit be brought in California courts rather than where you live, cap total liability at "the greater of $100 or the amount you have paid us in the past 12 months," and bar any claim not brought within one year of the event that caused it. That one-year window is the clause a parent would actually get caught by.

That last clause binds the user to "TIKTOK US JOINT VENTURE," a name that appears nowhere else in either document. Everywhere else the company is TikTok USDS Joint Venture. Two documents, one company, and the paperwork is not consistent about what it is called.

A class action was filed on August 31 in the Central District of California, which is the venue Gauth's own terms specify, alleging that the platform transmits the full text of users' questions to Google through embedded analytics code. That is an allegation in a complaint. It has not been tested, and nothing above rests on it. I note it only because a parent who goes and reads these documents this month will run into it, and I would rather you hear it here with the word alleging attached.

The survey I relied on contradicts itself. The prose on page 16 says "38% of kids age 9 to 17 have used AI-powered learning tools." The chart directly beneath that sentence gives its base as "Kids who ever use or interact with AI (n=1,032)," not the full sample of 1,204. Those are different populations. I have used the narrower one. Finish the sentence, then check the chart. That is the same instruction as the rest of this issue, and it applies to a report by an organization I respect.

I got a name wrong, and I am leaving the correction in rather than quietly fixing it. The survey I cite above lists examples of AI learning tools as "Gauth AI, Khanmigo, or Socratic Owl." I read Socratic Owl as Socratic, the Google app, and started writing about the wrong product.

Socratic Owl is a separate app with no relationship to Google, published by LUMINATION AI LTD, released on iPhone in June 2024 and on Android in 2025. Socratic by Google was something else, and it is gone: socratic.org now redirects to Google Lens, and the apps have been pulled from both the App Store and Google Play.

What makes this more than my own carelessness is that the confusion is being written into the record by people acting in good faith. Google's own help center for Socratic is still online, still describing features, and its header link still points at socratic.org, which no longer goes there. And in March of this year a Wikipedia editor repairing Google's dead App Store link on the Socratic page substituted the store listing for Socratic Owl, with the edit summary "Bad Link."

Two products, one name, and the one a parent would recognize is the one that no longer exists. I made an assumption because the name sounded familiar. That is the exact failure this entire issue is about, and obviously I fall victim to it as well.

Every quote above was read directly off the live document, not through any summary of it, on September 9, 2026, and checked again on September 12. Legal terms change without notice, so a quote is good only on the day it was read. When we get something wrong in a published issue, it will be listed here too.

Sources

Gauth, operated by TikTok USDS Joint Venture LLC

Privacy Policy, last updated January 22, 2026. gauth.com/policy

Terms of Service, last updated January 22, 2026. gauth.com/terms

Executive Order 14352, "Saving TikTok While Protecting National Security," signed September 25, 2025; 90 FR 47219, published September 30, 2025

https://www.federalregister.gov/documents/2025/09/30/2025-19139/saving-tiktok-while-protecting-national-security

Official PDF: https://www.govinfo.gov/content/pkg/FR-2025-09-30/pdf/2025-19139.pdf

Khan Academy Privacy Policy, last updated January 30, 2026

https://www.khanacademy.org/about/privacy-policy

MagicSchool Privacy Policy, effective March 19, 2026

https://www.magicschool.ai/privacy-security/privacy-policy

Survey data

The Common Sense Media Census: AI Use by Tweens and Teens, 2026. Common Sense Media, published June 8, 2026; fielded by SSRS March 18 to 26, 2026. The 85% figure is on pages 6 and 21. The 38% figure, and the chart underneath it, are on page 16.

https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens-1.pdf

Announcement: https://www.commonsensemedia.org/press-releases/common-sense-media-releases-inaugural-annual-study-on-ai-use-by-tweens-and-teens

The class action

Lominchar v. TikTok USDS Joint Venture, LLC, No. 2:26-cv-09700 (C.D. Cal., filed August 31, 2026)

Complaint: https://www.courthousenews.com/wp-content/uploads/2026/09/lominchar-v-tiktok-usds-joint-venture-llc-complaint.pdf

Coverage, September 1, 2026: https://www.courthousenews.com/tiktoks-us-entity-faces-class-action-over-ai-homework-helper/