How SoWo uses AI

Last updated: July 26, 2026

This explains how artificial intelligence works inside SoWo: what happens to your content, what we won't do with it, and what the AI gets wrong.

It's written for a teacher, a learner, or a parent to read. It isn't a contract — our Terms of Service and Privacy Policy are. But everything here is true, and where we're still building something we say so rather than rounding up.

The commitments

1. We don't train AI models on your content. Not models of our own, and we route your content only to providers whose published policies say they don't train on it either.

2. You own what the AI generates for you. We assign you our rights in it the moment it's created.

3. Checking AI output is your job, not ours. Our Terms make review the Facilitator's responsibility. SoWo does not review output for you, and nothing in the product stops unreviewed output reaching a learner. Read "What the AI gets wrong" below before you rely on any of it.

4. We tell you what we haven't finished. See the last section.

Where your content goes

We use a third-party routing service to reach AI model providers, rather than contracting with each one directly. We name the categories in our Privacy Policy, and if you want to know exactly which providers we use, email support@joinsowo.com and we'll tell you.

When you generate practice content: your prompt and the relevant context go over an encrypted connection to a model provider, which returns output. We store the prompt and output so you can see your history.

When you use speaking practice: we record your voice and send that recording the same way, so it can be transcribed and turned into pronunciation and language feedback. We store the recording so you can hear it back, and you can delete it at any time. We do not create a voiceprint and do not use your voice to identify you.

We configure our routing service to use only providers whose published data policies say they don't train on inputs or outputs, and we don't enable optional prompt logging. The routing service itself doesn't retain prompt or response content.

What providers may keep: model providers may retain content briefly for security and abuse detection — typically no more than 30 days — before deleting it. These commitments come from the providers' own published terms. We don't contract with each provider directly and can't independently verify their internal practices. We'd rather say that plainly than imply a guarantee we haven't signed.

Proficiency estimates

SoWo can generate an estimated proficiency level for a learner.

This is an automated estimate, not an accredited or standardized assessment. It's advisory to the Facilitator, who has to review it before relying on it. It should never be the only basis for a placement, grade, or certification decision.

When we read your content

Content flagged by our safety systems may be read by a small number of trained SoWo personnel, to determine whether it breaks our rules. That access is limited to the people who need it.

We don't otherwise read or listen to your content. We'd rather say that plainly than let you assume no human ever sees anything.

What the AI gets wrong

We'd rather tell you this directly than bury it in a disclaimer.

Fluency is not accuracy. Language models are exceptionally good at producing text that sounds like a confident native speaker. That's precisely what makes their errors hard to catch, and it's why a learner, by definition, can't catch them.

The specific ways this shows up in language learning:

• Invented idioms and collocations — plausible-sounding phrases that no native speaker actually uses
• Confident false grammar explanations — rules stated authoritatively that are incomplete or simply wrong
• Register and formality errors — output that's too casual or too formal for the situation, especially in languages where politeness is grammaticalized
• Dialect flattening — defaulting to one standard variety, and marking valid regional usage as an error
• Cultural mismatch — examples and scenarios that don't fit the learner's context
• Uneven quality across languages — performance is meaningfully worse in lower-resource languages

This is why our Terms put review on the Facilitator. Nothing in the product enforces it — the check has to come from you, and the failure mode is invisible from the inside.

And it means AI output should never be the sole basis for deciding placement, promotion, grades, certification, discipline, or admission for any person.

Who owns what the AI makes

You do. Our Terms assign you our rights in AI output the moment it's generated, and that isn't conditioned on your compliance with anything.

Two honest limits. We can only assign what we have, so we don't promise the output is original or protectable by copyright. And because generative models produce similar output for similar prompts, yours isn't exclusive to you.

Ownership transfers rights. It isn't a promise the output is any good — see the section above.

Where we are today

SoWo is in early release. Three things we're still building, stated plainly because you shouldn't rely on what doesn't exist yet.

Contractual zero-retention with our AI providers. Today we rely on providers' standard API terms, which don't use inputs for training. That's a real protection, but it isn't the same as a signed agreement requiring them to delete your content the moment output is returned. We're working toward that and we'll update this note when it's in place.

Automatic deletion. We delete your data when you ask us to, within 30 days. We don't yet purge inactive data automatically on a schedule. We're building it, and our Privacy Policy will carry specific retention periods once it's running. In the meantime, ask and we'll delete it.

School and district deployments. SoWo isn't currently offered for institutional K-12 use, and we don't serve learners under 13. Getting there needs a data privacy agreement and safeguards we haven't finished. We'd rather not offer it than offer it badly. If you're a school and want to know when: support@joinsowo.com.

Questions

support@joinsowo.com

Social World, LLC · Colorado, United States · mailing address available on request.

See also our Terms of Service, our Privacy Policy, and our Copyright Policy.