Education Chatbots

Nudge the Task, Enroll the Family, and Never Let the Bot Read the Record

Quick answer: There is one randomized trial of an education chatbot that measured the outcome a school actually cares about, and it is worth more than every vendor page on the subject. In summer 2016 Georgia State University randomized 7,489 admitted freshmen, and roughly half received a text-message assistant called Pounce, built by AdmitHub. Students who had already committed to the university and received the outreach were 3.3 percentage points more likely to enroll that fall, which the authors scale to about 116 extra students per 3,500-student class, for a platform cost of $7 to $15 per student a year before staff time, against $100 to $200 per student in earlier counselor-outreach programs. The authors name two innovations, the data integration and the AI, and do not separate their effects; our reading is that the record did most of the work. Only students who had not filed the FAFSA got the FAFSA message, and the authors warn that "messaging students about tasks already completed could inadvertently confuse them." The bot worked because it knew the record.

That is also the design problem. In a US school, the record the bot needs in order to be useful becomes an education record under FERPA the day the person attends, the person on the other end may be under 13 and outside the terms of every messaging app you would deploy on, and the message you most want to send ("your application is missing a transcript") is applicant data before enrollment and a disclosure after it. This guide takes those three constraints in order, then walks the enrollment funnel with the template categories that fit each step, and ends with the five things an education bot hands to a person every time.

Step 1. Read the trial as a spec, not as a headline

The paper is Page and Gehlbach, "How an Artificially Intelligent Virtual Assistant Helps Students Navigate the Road to College," AERA Open, 2017. The design details are the useful part.

Pounce sent its first message in April 2016 to students admitted for the fall, and the message asked permission: "Hi {Student Name}! Congrats on being admitted to Georgia State! I'm Pounce—your official guide. I'm here to answer your questions and keep you on track for college. (Standard text messaging rates may apply.) Would you like my help?" The sample was restricted to admitted students "with an active U.S. cell phone number who provided consent for text message communication in their GSU application." Consent came before the first text, in the application, and the first text asked again.

The content was task nudges on more than 90 enrollment topics, "including intent to enroll submission, FAFSA completion, scholarship and loan acceptance, orientation registration and attendance, and immunization form submission," and the nudges were driven by data: "By integrating data from the university's student information and customer relationship management systems, Pounce could send students messages that were personalized to students' immediate needs for those domains where they were failing to make progress or raised questions. For example, only students who had yet to file the FAFSA would receive FAFSA-related outreach." The typical treated student received 43 messages over the summer and sent about 14 back.

The escalation path was email. Questions Pounce could not answer went "to university admissions counselors via email," staff replies went back by text, and AdmitHub "reviewed these responses and incorporated them into Pounce's knowledge base." Only 13.5 percent of students ever triggered it, the average student sent "fewer than one message that required staff intervention," 85 percent replied to the bot at least once, and 6.6 percent opted out.

The effect was concentrated where the record was richest. Among students already committed to GSU, treatment raised final-transcript submission by 4 percentage points, orientation attendance by nearly 3, cut immunization holds by nearly 5 and FAFSA verification holds by 3, and raised loan acceptance by 6 to 7. Enrollment at GSU rose 3.3 percentage points in that group. Among the three-quarters of the sample who had been admitted but had not committed, the enrollment estimate is 0.6 points and not statistically significant, though a few task measures still moved (immunization holds fell by 4 points there too). The bot did not talk anyone into a college. It stopped people who had chosen one from losing it to paperwork.

Read as a specification, that is four rules: ask consent twice; nudge tasks the record says are open, and only those; route what the bot cannot answer to a named person and fold the answer back in; expect the return on families who have already decided, not on cold leads. The rest of this guide is about doing that legally with the tools an SMB actually has.

Step 2. Decide who the bot is talking to, because the platforms already have

Before the funnel, the age gate. WhatsApp's Terms of Service, in the Age clause: "You must be at least 13 years old to register for and use our Services on your own (or such greater age required in your country or territory). If you are under this age, your parent or guardian can create a parent-managed account for you if it is available in your country or territory." A parent who creates one "is subject to these Terms and responsible for your child's activity on our Services." Instagram's minimum is also 13. So a WhatsApp or Instagram bot for a primary school, a kids' coding club or a junior language program is, by the platform's own rule, a bot whose account holder and responsible party is a parent.

US federal law draws the same line for websites and online services. The FTC's COPPA FAQ says the term "broadly covers any service available over the Internet, or that connects to the Internet or a wide-area network," and lists "mobile applications that connect to the Internet," "smart speakers, voice assistants, voice-over-Internet protocol services" among covered online services; a WhatsApp or web bot is plainly inside that, and a bot that lives only in carrier SMS is not named, so treat it the same way as a matter of practice. The COPPA Rule, 16 CFR Part 312, defines a child in 312.2 as "an individual under the age of 13," defines personal information to include "a telephone number," "a photograph, video, or audio file where such file contains a child's image or voice," and "a persistent identifier that can be used to recognize a user over time and across different websites or online services," and in 312.5 requires an operator collecting it from a child to obtain verifiable parental consent first.

The rule reaches commercial operators. Its definition of operator "does not include any nonprofit entity that would otherwise be exempt from coverage under Section 5 of the Federal Trade Commission Act," and a public school district, in our reading, is outside the FTC Act as a governmental body; a for-profit tutoring franchise is inside it. The vendor, which operates the online service and collects the data, is a commercial operator in its own right, and the FTC's FAQ adds that operators "should not state in Terms of Service or anywhere else that the school is responsible for complying with COPPA." The same FAQ describes a route for schools: where a district contracts for online programs "solely for the benefit of their students and for the school system," the schools "may act as the parent's agent and can consent under COPPA to the collection of kids' information on the parent's behalf," limited "to the educational context," and only if the operator gives the school the same direct notice it would give a parent and lets the school review and delete the children's data. In our reading, an admissions bot that markets to prospective families is not that context.

The practical division is simple and worth writing at the top of the bot's specification:

That last row is the one universities get wrong in the other direction. Under 34 CFR 99.3, an "eligible student" is one "who has reached 18 years of age or is attending an institution of postsecondary education," and 99.5 says that when a student becomes eligible "the rights accorded to, and consent required of, parents under this part transfer from the parents to the student." There is an exception: 99.5(a)(2) and 99.31(a)(8) let an institution disclose to the parents of a student who is a dependent for federal tax purposes without the student's consent. A bot cannot establish tax dependency from a WhatsApp handle, so for the bot the rule is unchanged: a university bot that tells a parent whether their child's deposit has cleared is disclosing to someone whose entitlement it has not verified, and the answer is to hand the parent to the office.

Step 3. Treat FERPA as the bot's disclosure specification

FERPA applies to "an educational agency or institution to which funds have been made available under any program administered by the Secretary" of Education (34 CFR 99.1), which covers public schools, nearly all colleges and universities because their students bring federal aid, and any private school that receives Department funds directly. A private tutoring center or an online course business is usually outside it. Borrow the discipline anyway; the families you serve will assume it.

Four definitions do the work. Student means "any individual who is or has been in attendance at an educational agency or institution and regarding whom the agency or institution maintains education records" (99.3), and 99.5(c) adds that an applicant to another part of the same institution has no rights in the application records "unless the student is accepted and attends that other component of the institution." So the admitted-but-not-yet-attending population that Pounce texted were not FERPA students at the time; their data was applicant data, governed by the university's own privacy policy and state law, and it became education-record data on attendance. Education records are records "directly related to a student" and "maintained by an educational agency or institution or by a party acting for the agency or institution" (99.3). Once the person attends, the chatbot's transcripts and the data it reads from your student system are education records, and the vendor holding them is a party acting for you.

Personally identifiable information includes the obvious identifiers and also "information requested by a person who the educational agency or institution reasonably believes knows the identity of the student to whom the education record relates" (99.3). A parent asking "did Maya hand in her form?" is requesting PII. Disclosure means "to permit access to or the release, transfer, or other communication of personally identifiable information contained in education records by any means, including oral, written, or electronic means, to any party except the party identified as the party that provided or created the record" (99.3). A WhatsApp reply is a disclosure.

Disclosure needs "a signed and dated written consent" that specifies the records, the purpose and the recipient (99.30(a) and (b)), and the consent "may include a record and signature in electronic form" that authenticates the person (99.30(d)), unless an exception in 99.31 applies. Two exceptions matter to a bot. The first lets the vendor exist: a "contractor, consultant, volunteer, or other party to whom an agency or institution has outsourced institutional services or functions may be considered a school official" if it "performs an institutional service or function for which the agency or institution would otherwise use employees," "is under the direct control of the agency or institution with respect to the use and maintenance of education records," and is bound by the redisclosure limits in 99.33(a) (99.31(a)(1)(i)(B)). In practice that is the clause your vendor contract satisfies before the bot reads the record, consent under 99.30 being the other route.

The institution also "must use reasonable methods to ensure that school officials obtain access to only those education records in which they have legitimate educational interests," and the regulation lets it meet that with "physical or technological access controls" or with an administrative policy it can show is effective (99.31(a)(1)(ii)). Our rule, stated as ours, is to meet it in the integration: give the bot task flags, not the record, because a bot is a technological control whether or not you treat it as one. The second exception, directory information, lets a school publish a student's name, address, phone, email, photo, date of birth, major, grade level, enrollment status and dates of attendance without consent, but only after public notice and an opt-out period (99.37), and "dates of attendance" explicitly "does not include specific daily records of a student's attendance" (99.3). Whether a child was in school today is never directory information.

From those texts, the disclosure specification for the bot has three tiers, and the tier is decided by the question, not by who seems to be asking:

1. Public information to anyone. Programs, fees, the calendar, admissions requirements, what to bring on the first day, how to apply. This is the knowledge base work, and the travel chatbot guide's rule for money questions applies to fees and refund terms: quote the dated policy text or hand off; never let the model summarize a refund policy in its own words.
2. Task status to the authenticated account holder. "Your application is missing a transcript," "the deposit for the spring term is due Friday," "the immunization form has not been received." This is the Pounce mechanism. Before the person attends it is applicant data under your own privacy policy and state law; from the first day of attendance it is a disclosure of an education record, so build it from the start as if it were one: on the consent collected at application and on the vendor's school-official status. The channel identity (a phone number, an Instagram handle) is not authentication: family members share phones, numbers are reassigned, and Instagram DMs can arrive from anyone, and 99.31(c) requires an institution to "use reasonable methods to identify and authenticate the identity of parents, students, school officials, and any other parties to whom the agency or institution discloses" record information. Send the nudge to the number on the consent record, name the task, and keep the content to what the task needs.
3. Record content only inside the student system's own login. Grades, attendance by day, disciplinary matters, financial aid detail, health information. The bot links to the portal; it does not read from it and it does not read it aloud. This is where a bot that "integrates with the SIS" needs its integration scoped to task flags rather than to the record, which is our reading of how to satisfy 99.31(a)(1)(ii) with a technological control.

Step 4. Walk the enrollment funnel, one message category at a time

An education business runs two workloads on the same number. In admissions season, which for most is a few weeks, inquiries arrive faster than a small office can answer and the winner is often whoever replies first, at night, in the parent's language. Out of season, the work is the Pounce work: reminding the families who already said yes to do the next thing. The bot's job changes between the two, and so does the message category it may use.

Two mechanics from elsewhere on the site carry over unchanged. Template categorization is the WhatsApp message template entry's subject, including the silent reprice of a utility template that Meta judges to be marketing, which a "deposit due Friday, and ask about our summer program" message will trigger. The nurture arithmetic, how many replies your office can actually supply in the window an inquiry opens, is the lead nurturing automation guide's subject, and admissions season is the case where the reply supply runs out first. A drip sequence after an open day is the marketing half, on its own opt-in, so that a parent's opt-out from the newsletter does not silence the fee reminder.

On channel: the platform terms in step 2 make WhatsApp and Instagram parent channels for anything under 13, and our WhatsApp channel guide and Instagram channel guide cover setup. SMS, which is what Pounce used, needs the consent record and the opt-out handling our SMS chatbot setup guide describes, and it reaches the student directly at university age. Ask on the enrollment form which language the family wants messages in, and honor it; the multilingual chatbot guide covers what "supports 30 languages" does and does not mean.

Step 5. Write the five hand-off rules before the first flow

An education bot has more mandatory hand-offs than most, and the reason to write them first is that each one is a rule the bot must obey even when it could plausibly answer.

1. Any message about a student's safety goes to a named person at once. Self-harm, abuse, bullying, a child who says they are alone or afraid. The bot's only reply is that a person is coming, the transcript goes with it, and the escalation has a human on the receiving end at every hour the bot is on. Write the fallback too: if no person acknowledges within a set number of minutes, the bot says so plainly and gives the emergency number and a crisis line for your country, because a small office will not staff every hour and the bot must not pretend otherwise. If your staff are mandated reporters under state law, the escalation path is also the reporting path. This is not a place for a model to counsel, and your escalation playbook should list who receives it at 2 a.m., because that is when it will arrive.
2. Record questions from anyone who is not the authenticated account holder. "Is my nephew enrolled?" "Did my daughter's ex-boyfriend apply?" The bot does not confirm that a named person is or is not a student. Directory information may be public under 99.37 after notice and opt-out; the bot has no way to know who opted out, so it does not go there.
3. Admissions decisions. "Did I get in?" is answered by the office, in the office's words, on the office's timeline. The bot says when decisions go out and how.
4. Money in dispute. A refund claim, a fee waiver request, a hardship case. Quote the dated terms, then hand off with the transcript. The travel chatbot guide has the tribunal decision that prices getting this wrong.
5. Anything the knowledge base does not contain. In the Pounce trial 13.5 percent of students sent a message the system had to route to staff, and the answers were folded back in. Our measured hallucination rates across the platforms we have reviewed run from 6.7 percent at best to 25 at the anchored top of the range; a bot that invents a scholarship deadline has done real harm, and our guide to reducing hallucinations covers the scope-first triage that keeps it from trying.

The handoff design guide covers the mechanics of the transfer itself. The education-specific addition is that rules 1 and 2 are absolute: they do not degrade with confidence, they are not overridden by a good question, and they are the first two lines of the system prompt.

What our fifteen reviews record about education

Four mention it, none tested it. Tars positions itself for "regulated verticals (banking, insurance, healthcare, education, government)" and ships education starter flows in its template gallery, which is the closest thing to a vertical fit in the catalog; its form-bot specialization matches the application-and-documents step above better than it matches an inquiry chat. AiSensy lists education among six verticals on its about page and has an Education filter in its template library. SendPulse has an "Online Education" category among its chatbot flow templates and, outside the chatbot, a course builder with student management. Wati's template library has an education category, in the alt text of a screenshot. No review built an enrollment flow, tested an age gate, or asked a bot a records question, and the search strings in Sources will reproduce that.

What the corpus does offer is transferable. The hallucination figures in step 5 come from Scenario D of our methodology; the handoff scores from Scenario E; and the platforms whose reviews report per-source citation (Chatbase, Tidio, Intercom) are the ones to trial for the public-information tier, because a fees answer with the policy page attached is the answer a parent will forward to the other parent.

FAQ

Do chatbots work for student enrollment?

The one randomized trial we know of says yes, narrowly. Georgia State's Pounce assistant raised fall enrollment by 3.3 percentage points among admitted students who had already committed, by nudging specific open tasks drawn from the student system, and had no significant effect on students who had not committed. Expect the return on families who have decided and are at risk of losing the place to paperwork, not on cold leads.

Can a school chatbot talk to students under 13?

Not on WhatsApp or Instagram, whose terms set 13 as the minimum age for an account of one's own, and not on a website or app without verifiable parental consent under COPPA if the operator is commercial; treat SMS the same way in practice. For anything below 13 the bot talks to the parent; the child's name in that conversation is information about a third party.

Is a chatbot allowed under FERPA?

Yes, if the vendor qualifies as a school official under 34 CFR 99.31(a)(1)(i)(B): it performs a function you would otherwise staff, it is under your direct control as to the records, and it is bound by the redisclosure limits; or with the signed, dated written consent that 99.30 describes. The institution must also use reasonable methods to limit school officials to records they have a legitimate interest in; our recommended method is to scope the integration so the bot sees task flags, not the record. Whether the bot may say a particular thing to a particular person is then a disclosure question, answered in step 3.

What can an education chatbot answer without consent?

Public information: programs, fees, the calendar, admissions requirements, logistics. Anything about a specific student is personally identifiable information from an education record once the school knows who is asking about whom, and needs either the account holder's consent on file or a 99.31 exception. Directory information is disclosable only after public notice and an opt-out window, which the bot cannot verify, so treat it as off limits.

Can the bot tell a parent their child's grades or attendance?

Not in the chat. Grades and daily attendance are education records; "dates of attendance" as directory information explicitly excludes daily attendance. Link the parent to the student portal, where the school's own login authenticates them, and keep the bot to task status ("the permission form has not been received") sent to the number on the consent record.

Which chatbot platform is best for schools?

Our reviews cannot say; none tested an education flow. Tars is the only reviewed platform that names education in its positioning and ships education templates; AiSensy, SendPulse and Wati have education template categories. Trial for the three tiers separately: public answers with citations, task nudges from your student system, and a hand-off that reaches a person at night.

Related guides

• Lead qualification playbook — the generic inquiry funnel this guide specializes for admissions.
• Lead nurturing automation guide — the reply-supply arithmetic that admissions season breaks first.
• Appointment booking chatbot — availability, the double-booking race and timezones, which a trial-class booking inherits.
• Healthcare chatbot guide — the HIPAA counterpart to step 3: what has to be true before a bot touches a regulated record.
• WhatsApp message template — categories, review and the silent reprice that step 4 warns about.
• Chatbot escalation playbook — the queue and the triggers the safeguarding rule depends on.
• Chatbot security and PII handling — what the bot stores, where, and for how long.
• Reduce chatbot hallucinations — scope-first triage, the method behind hand-off rule 5.

Sources

• Page, L. C., and Gehlbach, H., "How an Artificially Intelligent Virtual Assistant Helps Students Navigate the Road to College," AERA Open, volume 3, issue 4, 2017 (doi 10.1177/2332858417749220), read in full at journals.sagepub.com on 3 September 2026. Quoted or reported: the sample definition (N = 7,489 admitted students "with an active U.S. cell phone number who provided consent for text message communication in their GSU application"); the introductory message; the "more than 90 enrollment topics" list; the data-sharing passage and its FAFSA example; the text-to-email routing passage; Table 3 engagement figures (96.2 percent received messages, 43.2 messages received on average, 85.1 percent responded at least once, 13.9 messages sent on average, 13.5 percent sent a message requiring staff response, 6.6 percent opted out); the Table 4 and Table 5 effects for committed students (final transcript +4 points, orientation +3, immunization hold −5, FAFSA verification hold −3, loan acceptance +6 to +7, loan counseling +6, as the authors round them); Table 6 (enrollment at GSU +3.3 points among committed students, p < .05; +0.6 points among non-committed and +1.2 points overall, neither significant; two-year enrollment −1.3 points among committed students, p < .01); Table 4's immunization-hold column for non-committed students (−4.0 points, p < .001); the "approximately 116 accepted students" scaling and its decomposition (46 from two-year institutions, 70 from no enrollment); the cost passage ($7 to $15 per student per year for the platform, $100 to $200 for counselor outreach, about $53,000 a year at $15 per student); and the warning that "messaging students about tasks already completed could inadvertently confuse them." The 3,500-student cohort figure is the authors' description of GSU's freshman class. Georgia State is a large public university; the transfer of these results to a private school or a tutoring business is a judgment, stated as such below.
• 34 CFR Part 99, Family Educational Rights and Privacy, read at ecfr.gov on 3 September 2026 (title 34 current to 31 August 2026): 99.1(a) (applicability to institutions receiving Department funds), 99.3 (definitions of directory information including the daily-attendance exclusion, disclosure in full including its "except the party" tail, education records, eligible student, personally identifiable information paragraph (g), and student), 99.5(a)(1) (transfer of rights), 99.5(a)(2) with 99.31(a)(8) (disclosure to parents of a dependent student as defined in section 152 of the Internal Revenue Code), 99.5(c) (no rights in application records "unless the student is accepted and attends"), 99.10 (right to inspect and review), 99.30(a), (b) and (d) (consent and its electronic form), 99.31(a)(1)(i)(B) (the school-official conditions) and (a)(1)(ii) in both its sentences (the reasonable-methods requirement, and the administrative-policy alternative for an institution "that does not use physical or technological access controls"), 99.31(c) (reasonable methods to "identify and authenticate the identity" of parties receiving disclosures), 99.33(a) (redisclosure), 99.37(a) (directory information notice and opt-out). The statement that pre-attendance applicants are not FERPA "students" follows from the 99.3 definition and 99.5(c); the point at which a particular institution's records attach is a question for its counsel.
• 16 CFR Part 312, Children's Online Privacy Protection Rule, read at ecfr.gov on 3 September 2026 (title 16 current to 1 September 2026; source line "78 FR 4008, Jan. 17, 2013, as amended at 90 FR 16977, Apr. 22, 2025," timeline change 23 June 2025): 312.2 definitions of child, operator (including the nonprofit exclusion and the "on whose behalf such information is collected or maintained" clause), "online contact information," personal information items (5), (7) and (8), and "obtaining verifiable consent"; 312.5 (parental consent) and 312.10 (data retention and deletion) by section title. The statement that the vendor is a commercial operator in its own right rests on the primary prong of the operator definition (a person who "operates a website located on the internet or an online service and who collects or maintains personal information").
• Federal Trade Commission, Complying with COPPA: Frequently Asked Questions (ftc.gov business guidance, page updated 10 July 2026), read 3 September 2026: the "online service" passage quoted in step 2 (the definition sentence and the list including "mobile applications that connect to the Internet," "smart speakers, voice assistants, voice-over-Internet protocol services"), the schools passage (schools "may act as the parent's agent and can consent under COPPA to the collection of kids' information on the parent's behalf," limited to "the educational context," with the direct-notice and review-and-delete conditions), and the sentence that operators "should not state in Terms of Service or anywhere else that the school is responsible for complying with COPPA." SMS is not named as a covered online service anywhere in those passages, and the body says so.
• WhatsApp Terms of Service (whatsapp.com/legal/terms-of-service, effective date shown as 4 January 2021, with the current age clause and parent-managed-account language), read 3 September 2026: the Age paragraph and the two parent-managed-account bullets under Other. The Instagram minimum age of 13 is stated from Meta's published policy and was not re-read for this guide.
• Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), searched 3 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. Education matches: grep -liE '\beducation\b|universit|\bschool|\bstudent|\bedtech\b' sample-reviews/*-review.md returns 4 (aisensy, sendpulse, tars, wati); per-file counts from grep -ciE on the same pattern are 4, 2, 10, 1. Passages quoted: tars-review.md lines 156, 378 and 663 (vertical positioning) and 176, 382 and 431 (the template gallery's "education enrollment" starter flow); aisensy-review.md line 157 (about-page verticals) and the template-library captions at lines 668 and 720; sendpulse-review.md line 237 (course builder with student management) and the caption at line 618 ("Online Education" flow-template category); wati-review.md line 567 (template library alt text). Hallucination figures: tidio-review.md line 205 ("6.7% (1/15) hallucinations," measured on a single-page index) and aisensy-review.md ("15-25%," anchored), the two ends of the range reported in the travel chatbot guide's Sources, which lists all thirteen files. The statement that no review tested an education flow is an argument from absence across the fifteen files and is stated that way in the body.
• Ahrefs Keywords Explorer, US overview, queried 3 September 2026 — the demand, difficulty, CPC, global-volume and parent-topic figures in this page's keyword note.
• Chatbotscape evaluation methodology, including the six-scenario protocol under which the Scenario D hallucination and Scenario E handoff figures were recorded. /methodology (continuously updated).

About this guide

Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or admissions lead of a school, tutoring or language business, or a university office deciding what a bot may do and say. It names no best platform for education, because our reviews have not tested one, and it reads the trial and the regulations it is built on rather than summaries of them.

Methodology

The journal article and the two federal regulations were read in full on the day of drafting and every quotation carries a table, section or paragraph reference so that a reader can check it. The corpus figures are our own published measurements and anchored assessments, with the evidence class of each stated in Sources. Every corpus search was executed in the form the page states. No education deployment was run for this guide and no figure on it is a measurement of a school.

The editorial judgments on this page, listed rather than flagged line by line:

1. Transferring a public university's randomized result to private schools, tutoring centers and course businesses. The mechanism (record-driven task nudges to families who have already committed) is what we transfer; the effect size is not.
2. Reading the Pounce effect as driven more by the student-system integration than by the conversation. The authors name both as innovations and do not decompose the effect; the attribution is ours.
3. Reading FERPA's definitions as a three-tier disclosure specification for a chatbot. The tiers are our construction; the definitions are quoted.
4. Recommending that the reasonable-methods requirement in 99.31(a)(1)(ii) be met by scoping the integration to task flags. The regulation also accepts an effective administrative policy; the preference for a technological control is ours.
5. Treating a phone number or a social handle as insufficient authentication for record content. 99.31(c) requires reasonable authentication methods without naming any; the shared-phone and reassigned-number cases are why we say a channel identity is not one.
6. Treating an SMS-only bot as if COPPA's "online service" definition reached it. Neither the rule nor the FTC's FAQ names SMS as a covered online service; the practice is ours.
7. Making the safeguarding hand-off absolute rather than confidence-based, with a timed fallback to emergency contacts. This is an editorial rule, not a legal one.
8. Publishing the corpus gap (four mentions, zero tests) instead of a shortlist.

See our methodology for how platform facts are verified.

Last updated

4 September 2026.