How Cocovox's AI Works

Which model is running this?

Cocovox uses Anthropic's Claude — one of the most carefully-built AI models available today. It's the same technology used by researchers, medical teams, and educators who need thoughtful, safe responses instead of the fastest possible ones.

Claude was trained with a technique called Constitutional AI — a set of principles the model is asked to follow before it answers. That makes it more willing to say "I don't know," more likely to ask a follow-up question, and less likely to bluff. For a product that talks to children, those qualities matter more than raw speed.

What it does well

• Listens patiently. Your child can take thirty seconds to think, repeat themselves, or change their mind. Claude doesn't rush them.
• Models clear speech. When a word is tricky, it offers a gentle model your child can echo — never a correction they'll feel shamed by.
• Asks useful follow-up questions. Instead of guessing at what your child meant, it asks — the way a good tutor would.
• Keeps sessions age-appropriate. Vocabulary, examples, and tone are matched to the learner's profile. A 7-year-old and a 14-year-old get different conversations.
• Admits its limits. If it didn't hear clearly, it says so. If a question is outside what it should answer, it says so. It's trained to be honest instead of confident.

What it doesn't do

• It doesn't diagnose speech or language disorders. Your SLP does that. Cocovox gives a practice check, never a clinical assessment.
• It doesn't start from a blank page each time. Cocovox keeps short notes about your child's learning — topics that were hard, breakthroughs, goals they mentioned — and uses them so the next session can pick up where they left off. Some of these notes are shown to your child on their My Learning Profile page, where they can delete them. Others, including some the tutor uses, are not shown there.
• It doesn't pretend to understand. If the microphone missed a word, it asks your child to repeat instead of guessing. No fake smiling-and-nodding.

How humans stay in charge

• Every conversation is logged. A parent with a linked account can open their child's chats in the Parent Dashboard and read what was said. The list shows each chat's title, when it happened, and how many messages it had. Some chats are kept private.
• Scores do appear in a few places. Teachers can see AI grading suggestions, and they decide the grade. From middle school up, learners can see an XP leaderboard, and the weekly parent email includes a quiz average. Elementary learners have no leaderboard in their menu.
• SLPs can flag concerning patterns. Clinicians using Cocovox inside an IEP program see patterns across sessions and can adjust goals.
• Word checks, and what they can't catch. Your Family Dashboard says the same:
  • Messages with words on our blocked list, like swearing and slurs, aren’t sent to Coco, and your child is asked to say it a kinder way.
  • If something your child types sounds like they may be hurt or unsafe, Coco’s reply starts by pointing them to a trusted adult and a help line, and a note is saved for our team to review. The note says what kind of worry it was, not the words.
  • When Coco finishes a typed reply, the app checks it for words on the same list and swaps it out if it finds one. Part of a reply can show while Coco is still writing it.
  • After Coco replies to your child, typed or spoken, we check the reply. If it’s flagged, a note is saved for our team to review. This check doesn’t change what your child saw or heard.
  • The word checks are for typed chats. They look for words and phrases, and no check can catch everything.
• Some AI-drafted things are marked. The AI-written summary in the in-app weekly digest and AI grading suggestions for teachers carry a small badge. The weekly email adds a note under its AI-written summary. Each one tells you a machine wrote it and a human should read it before treating it as truth. Other AI-drafted messages may not be marked yet, so read them the same way.
• If a teacher connects Cocovox to Claude, it only sees counts — never names. This link is rolling out, and it is off unless it has been switched on. Where it is on, teachers can link their Cocovox account to Claude to ask about their class. This is the same Claude that already powers Cocovox — and over this link it receives less, not more: only counts and patterns (like "9 students are behind on fractions"), never a student's name, and small groups are hidden entirely. Anything Claude drafts (a quiz, a lesson) arrives in Cocovox as a draft the teacher must approve before a single student sees it.