We do not train on your data, and here is what that costs us

Most AI products in 2026 train their models on the inputs you give them.

Some disclose this clearly. Most bury it in the terms of service. A few opt-out flags exist in enterprise tiers but not in consumer accounts.

The default is: your conversations are training data.

AI Twin's position on this is simple.

We do not train models on your data. Full stop.

Not because it would be illegal. Not because it would be hard.

Because the moment we start, the product stops being trustworthy in the way it was built to be.

This is not a costless position. We pay real prices for it.

I want to lay out what those prices are. The field is full of products that claim privacy-first without acknowledging the trade-offs.

Cost one: slower personalisation

Products that train on user data can fine-tune for individual usage patterns at a level AI Twin does not attempt.

The competitor uses your conversations to make subsequent conversations work better. We do not.

Our personalisation comes from typed memory you have explicitly shared. Not from quietly learning from your behaviour.

This means we improve more slowly for individual users than products that learn from all inputs.

Slower. That is the word.

Cost two: no network effects from data aggregation

Many AI products improve as more users feed them more data. Each user's interactions improve the model for the next user.

AI Twin does not have this loop.

The improvements we ship come from product development. Not from data accumulation.

Our improvement velocity is bounded by our team's velocity. Not by our user count.

Which means if we are slow as a team, the product is slow.

We accept that. We just have to be a fast team.

Cost three: cheaper inference is not available to us

Products that train custom models on their user base can run cheaper inference on those models.

AI Twin uses third-party models (Claude, GPT, Gemini, others) without sending them training data.

We pay per-token rates rather than amortising training costs across user data.

This means our gross margin is lower than it would be if we trained custom models on user inputs.

Worse unit economics. There is no way around that.

Why we hold the line anyway

Three reasons.

The product would no longer be the product we wanted to build.

The trust premium compounds. Each user who joins us specifically because we do not train on their data brings other users who care about the same thing. Slowly. But the curve goes up.

The regulatory direction is one-way. Training on personal data without explicit consent is going to get more expensive to defend over the next five years. Not less.

So we hold the line.

Where this leaves us

We ship slower.

We compete on trust, not on capability inflation.

We pay more per inference.

We make less money per user in 2026 than the products that monetise data flow.

That is the deal.

A note on what this is not

This is not virtue-signalling. We are not the only company holding this position, and we are not the loudest about it.

It is also not a marketing strategy.

It is just what we believe the product needs to be. The fact that some users care about it is helpful. It is not why we are doing it.

If everyone in the market shifted tomorrow and started training on user data with explicit consent and clear opt-outs, our position would still be the same.

We just would not be talking about it as much.

What you can do with this

Most product positioning is about what a company does.

This position is about what we do not do.

The companies who will earn long-term trust in AI are the ones who can show their work on the trade-offs.

The ones who never mention the trade-offs are usually the ones avoiding them.

Worth knowing which kind your tools are.

That is the test.