Two kinds of computer helper
The guesser
Today’s popular AI chatbots work by pattern-matching—they produce plausible answers, sometimes confidently wrong ones, and they usually send your question to a corporate cloud.
The one that shows its working
A reasoning engine follows strict, checkable rules over verified facts—like an accountant who can produce the ledger, not just the number. Run locally against data you are entitled to use, it answers from evidence rather than guesswork.
The family, briefly
"AI" is not one thing—it is a family of techniques with very different manners. Telling them apart is the difference between a tool you can trust and one you can't:
| Kind | What it actually does | Watch out for |
|---|---|---|
| Symbolic reasoning | Applies stated rules to known facts—like a clerk with a checklist. | Only as good as its rules; rigid, but every step is checkable. |
| Machine learning & deep learning | Finds statistical patterns in vast numbers of examples. | Approximate by nature; it cannot tell you why. |
| Generative AI | Produces plausible text, images and code from learned patterns. | Fluent is not true—it predicts what sounds right. |
| Perception systems | Image recognition and natural-language processing—turning the messy world into data. | Confident mistakes on anything unusual. |
And inference simply means drawing a conclusion from information—whether by rules (checkable) or by a trained model (its best guess). The honest helper on this site is the checkable kind.
Where what you type goes
It is worth understanding how the popular services actually work. They are built from data—enormous amounts of it—and when you use one, the service typically keeps a copy of what you provided, citing legal requirements: records for compliance, safety and disputes. Fair enough, as far as it goes. But the terms usually stretch that into a broad licence to use, analyse and sometimes share what you typed—the legal requirement becomes the door through which the licence walks.
Local infrastructure separates the two. If a court orders access, a local system can produce exactly what the order directs—the specific information, for the specific purpose, consistent with the court's directive—without anyone ever having granted a licence to use or sell it to other parties. Lawful access and broad licence are different things; this design meets the first without conceding the second.
The blueprints are already free
The standards that make this possible were built through the World Wide Web Consortium, whose members—including the largest technology companies—are bound by a royalty-free patent policy: the foundational patents for this kind of infrastructure are permanently available to build on, without licence fees. What has been missing is not permission but product: working local implementations. Building them creates the market pressure that makes the big vendors interoperate rather than dictate.