Frequently asked questions

Is tAI built on top of another company's model?

No. Every version, from the earliest release through tAI 4.2, is trained end to end on data collected and compiled by Artfical itself. That happens on infrastructure Artfical operates itself, rather than rented from a third-party cloud AI platform. No third-party model or corpus is used as a base for any version, at any point in that history. That's a deliberate, harder, and more expensive path than fine-tuning an existing base model would have been. It's described in full from a company-decision perspective at ai.artfical.com/about/models.

Can tAI take an irreversible action without my approval?

No, under any circumstance, across any product surface. Sending a message, publishing content, pushing to a default branch, and running a destructive command are all approval-gated at the product level. That's independent of how confident the model itself is in the correctness of the action it's proposing. That gate isn't something the model can reason its way past by being sufficiently certain. It's enforced outside the model, not by the model's own restraint. See Connector safeguards and Sandboxing for the full detail behind how that's actually enforced.

Has tAI been audited by an outside organization?

Not yet, by a standing, ongoing third-party evaluator. Some larger, longer-established labs have built relationships of that kind over time; we haven't yet. Every evaluation number published anywhere on this site, unless a page explicitly states otherwise, comes from our own internal harness. We say so directly on the relevant pages. We don't imply broader independent validation than currently, honestly exists. Overstating that would undermine trust in every other claim made throughout this section.

How is Turkish-language safety actually verified, rather than assumed?

Through a fully separate evaluation track, built from natively written Turkish test prompts for every one of the seven categories. It's not a machine-translated version of the equivalent English test set, applied as a shortcut. Native speakers with relevant subject-matter background are involved in constructing that test set directly. The reasoning behind why translation alone isn't sufficient is covered in full on the Turkish-language misuse page. So is the actual measured gap between English and Turkish results across recent versions. That gap has narrowed with each version, and the page reports the current figure directly rather than rounding it down to "closed."

Why does this site mention things that didn't work, like a rejected checkpoint?

Because a security page that only ever reports the flattering final number is easy to write. It's correspondingly hard to actually trust. We'd rather this section earn trust than read smoothly. An intermediate failure that was caught internally and genuinely fixed before it ever reached a real user is considerably more informative than omitting it entirely. It tells you more about how the underlying process actually works in practice than only describing the polished, successful outcome would. See tAI 4.2 for a specific, concrete example of exactly this kind of rejected checkpoint.

What should I do if I find a safety issue?

Go to Responsible disclosure and report it there directly. Include enough detail for us to actually reproduce what you found. That page covers exactly what counts as worth reporting. It covers how a report gets triaged and handled once it's submitted. It covers what kind of credit a confirmed report can lead to. It's worth reading in full before you submit anything, if you have the time. We'd still rather receive an under-detailed report than none at all.

Does tAI train on my conversations by default?

Where usage data is used to inform future training at all, that use is opt-out at the account level, meaning it's a choice you make rather than a default you have to go find and disable. Any usage data used that way goes through the same review and filtering step new training corpus data goes through before it's eligible for inclusion, described on the Data handling page. Turning that setting off removes you from that pipeline going forward, not just from what's visible in your own history. This is separate from the two training corpora, ArtficalAI and the Artfical Code Index, which are their own curated resource under a different sourcing process entirely, also described on that page. Your conversations are never quietly folded into either one without going through this explicit, opt-out-able path first. That setting is visible and changeable in Settings at any time, not something you have to contact us to change.

What happens if I export my data and then delete my account?

Those are two independent actions, and doing one doesn't require or affect the other. Exporting first gives you a copy of your conversation history to keep, described on the Data handling page. Deleting your account afterward still removes the underlying data on our end the same way it would if you hadn't exported anything first; having your own copy doesn't change what happens to the copy we hold. The two features exist for different reasons, portability and control, and using one is never a prerequisite for the other. You can export without ever deleting anything, and you can delete without ever having exported first. Doing both, in either order, is a completely normal and fully supported way to use the two features together.