The 40th principle: who verifies the other 39?

Last week, Taras Bakusevych published 39 Principles for Designing Human–AI Interaction — and if you design AI products for a living, you should read it before you read this. It is the most complete applied synthesis of the field I've seen: Horvitz's mixed-initiative work, Microsoft's 18 guidelines, Google PAIR, IBM's generative AI principles, all compressed into one framework a product team can actually use.

Three things elevate it above the usual listicle.

On the second point he's describing, from the model side, exactly what our Heuristic 07 — Appropriate Assertiveness demands from the product side: pushback is professional, and blind compliance is a failure mode. On the third, it's the argument of Can I Prove What the Agent Did? compressed into four words.

This is not a takedown. It's a completion.

Because all 39 principles rest on one unexamined assumption, and that assumption is expiring.

The unexamined assumption

Every principle assumes a human on the other side of the glass.

Read the framework again and notice the shape of it: a human types, a system responds, an interface mediates. Suggest or act. Show the diff. Ask when uncertain. Design the wait. That was a safe axiom in 2019, when Microsoft published its guidelines. It was still reasonable in 2024. In 2026, it's a hole. Two questions the article never asks:

These aren't edge cases. They're the two fronts where trust work is moving fastest — and where the design community is, so far, mostly absent.

Blind spot 01

The agent is now the user.

Bakusevych's principles govern how your product's AI behaves toward humans. But the traffic hitting your brand increasingly isn't human. Shopping agents comparing your offer against three competitors. Research agents summarizing your category for a buyer who will never visit your site. Procurement agents reading your documentation to decide whether you make a shortlist. It's the shift we mapped across YC's Spring 2026 batch: the agents got jobs, and one of those jobs is evaluating you.

Almost every principle in the framework has a mirror image on the brand side — and the mirror is unbuilt.

At auxfirst we call this discipline AUX — agentic experience design. Because when both sides of an interaction can be synthetic, "represent the AI's nature honestly" stops being a UX nicety and becomes the handshake protocol of commerce. Concretely: picture a checkout where the "user" is a shopping agent acting for a customer. Which of the 39 principles applies? Nearly all of them — inverted. The agent needs your capability claims, your provenance, your refusal paths, your escalation design. It just needs them in a form no Figma file has ever specified.

Blind spot 02

Principles without proof.

Here is the harder problem. All 39 principles are voluntary interface choices. Every single one can be claimed in a keynote and absent from the product. The framework has no answer to the only question a buyer, a regulator, or a partner agent actually asks: who checked?

We've seen this movie. Web accessibility had principles for a decade — thoughtful, well-sourced, widely ignored. It took WCAG conformance levels, audit practices, and legal teeth to turn principles into a market. Accessibility didn't become real when designers agreed it mattered. It became real when conformance became verifiable.

AI interaction design is at the pre-WCAG stage. Bakusevych has written an excellent set of principles. What the field lacks is the conformance layer: named trust tiers, testable requirements, attestation that a system actually does what its interface implies. That's the gap we're building toward — in the open with TrustKit and the Trust Harness, and in a working draft we call the Trustworthy Agent Protocol (TAP): normative requirements instead of aspirations, trust tiers instead of vibes, continuous monitoring instead of a one-time badge. The formula is simple:

The clock

The regulatory deadline is louder than the article admits.

One date is conspicuously missing from the piece: August 2026. That's when the EU AI Act's transparency obligations bite — one month after publication. And under those obligations, a chunk of the framework quietly changes category:

In the EU, these are no longer design advice. They're compliance requirements with enforcement attached. Which means the 39 principles aren't just a better-UX argument anymore — they're a compliance roadmap the article doesn't realize it drew. The teams that treat trust design as a legal perimeter will spend less and ship calmer than the teams that treat it as polish.

The ceiling

Where the ladder stops.

One more quiet limit. Bakusevych's "Calibrated Trust" — the strongest category in the framework — ends at appropriate reliance: the user trusts the system exactly as much as it deserves. Necessary. Not sufficient. In our Trust Architecture, appropriate reliance covers the first two stages. The commercial value lives higher.

A framework that stops at reliance treats trust as a UX property. Treated as an asset, it's a growth engine — and it's the only durable moat when every competitor has access to the same models. That's the argument we've made since the manifesto: a safe agent isn't automatically a trusted one, and a relied-upon agent isn't yet an advocated-for one.

---