The HTTPS of AI

Remember 1995. The web exists, online shops too, credit cards too. And yet almost no one buys. It is not that the technology is missing — it is that nothing proves the page on screen is really your bank's, that the card number will not be intercepted, that the merchant is who he claims to be. The web knows how to transport information; it does not yet know how to guarantee it.

Then an invisible layer arrives: the encryption of exchanges, the certificates that attest sites' identity, and their tiny manifestation — a padlock in the address bar. No one ever bought a padlock. Everyone bought because of it. Online commerce, online banking, online government: whole swathes of the economy were waiting neither for more bandwidth nor for better pages — they were waiting for an infrastructure of trust. When it arrived, everything unlocked within a few years.

HTTPS gave the internet its padlock. AI is waiting for its own.

Generative AI today is exactly where the web was in 1995. The capability is there — spectacular. The surface adoption too. But the uses that commit an organization — the signed report, the transmitted diagnosis, the documented decision, the invoiced deliverable — all run into the same absence: nothing proves that what the system asserts is true, where it comes from, or that one will be able to demonstrate it later. Executive committees each phrase it in their own way; it is always the same sentence: we cannot commit to what we cannot verify.

The parallel teaches three things. First, trust is not added to the product — it settles beneath it, as an infrastructure layer: invisible, continuous, systematic. Second, it does not come from the actors it governs: certificates were not issued by the merchants themselves, and the verification of AI cannot be rendered by the models it judges. Third, its economics are those of infrastructure: its value grows with every use, and those who hold it become the obligatory passage for everything built on top.

What the padlock certified was identity and transport. What AI must certify is deeper: the provenance of every claim, its confrontation with qualified sources, and the capacity to replay the verdict years later. Call it the cognitive trust layer. Whoever lays it does not sell one more tool — they unlock the uses everyone is waiting for.