Skip to main content

An open standard, v1

Heuristics for trustworthy AI interfaces

We have a shared standard for usability and a shared standard for accessibility. There is no equivalent for whether an AI interface is trustworthy: does it show its uncertainty, let people review and correct output, and keep humans in control of what matters?

AXIS is a concise, opinionated set of eight heuristics for AI product interfaces, plus an auditor you can run against any AI feature in a few minutes.

The eight heuristics

Each heuristic states a rule, why it matters, and what it looks like done well. Together they cover one question: can a person safely rely on what this interface gives them? None of it is asserted from nowhere; every heuristic cites the published research it synthesizes, from Microsoft's Guidelines for Human-AI Interaction to Google's People + AI Guidebook. See the full grounding.

01

Visible confidence

Signal how sure the system is, and mark the parts it is least sure about.

Why it matters

AI output arrives with uniform, flat confidence; a guess reads exactly like a fact. Surfacing uncertainty tells people where to slow down and check, instead of trusting every sentence equally. A system that admits what it doesn't know earns more trust than one that never hesitates.

What good looks like

Low-confidence spans are visually distinguished, and the system has a genuine “I'm not sure” state instead of a confident wrong answer.

Grounding

Microsoft HAX G2 · Google PAIR (Explainability + Trust) · Apple HIG, Generative AI · NIST AI RMF (explainable & interpretable)

02

Reviewable output

Let people see and inspect what the AI produced before it is used or sent.

Why it matters

If output is applied automatically or buried behind a summary, users have no moment in which to catch errors. Review has to be a designed step in the flow, not something users improvise after the fact.

What good looks like

A clear review state that shows exactly what was generated, in full, before anything is done with it.

Grounding

Apple HIG, Generative AI · IBM design principles for generative AI (CHI 2024) · Shneiderman, Human-Centered AI

03

Cheap correction

Allow editing or overriding output in place; never force a full restart to fix one thing.

Why it matters

If fixing a small error means regenerating everything, correction becomes expensive, so people stop correcting. They either accept flawed output or abandon the tool. The cost of a fix should be proportional to the size of the mistake.

What good looks like

Inline editing and per-item accept/reject controls, not a single all-or-nothing “regenerate” button.

Grounding

Microsoft HAX G9 · Google PAIR (Feedback + Control) · Nielsen's usability heuristic 3 (user control and freedom)

04

Graceful recovery

Provide a clear way back when the AI is wrong: undo, retry, or escalate.

Why it matters

These systems fail; that is a design constraint, not an edge case. A product with no recovery path turns a model error into a user dead-end, and dead-ends are where trust goes to die.

What good looks like

Obvious undo and retry actions, plus a route to a human or a manual mode when the AI can't get it right.

Grounding

Microsoft HAX G8 and G9 · Google PAIR (Errors + Graceful Failure) · Nielsen's usability heuristic 9 (error recovery)

05

Human control on consequential actions

Require explicit confirmation before anything irreversible or high-impact.

Why it matters

Autonomy is fine for low-stakes steps and dangerous for high-stakes ones. The level of human control should be matched to the consequence of the action, not to how confident the model happens to be.

What good looks like

A deliberate, explicit confirmation before actions that spend money, send messages, or can't be undone.

Grounding

Shneiderman, Human-Centered AI · Microsoft HAX G16 and G17 · Apple HIG, Generative AI · ISO/IEC 42001 (human oversight)

06

Honest expectations

Don't imply certainty or ability the system lacks; label AI-generated content.

Why it matters

Overselling the AI sets users up to over-trust it, and the crash that follows is worse than modest expectations up front. Clear labelling and honest framing protect both the user and their long-term trust in the product.

What good looks like

AI-generated content carries a visible label, and interface copy avoids promising accuracy the model can't guarantee.

Grounding

Microsoft HAX G1 and G2 · IBM Carbon for AI (labelling AI content) · Google PAIR (Mental Models) · NN/g research on sycophancy and over-trust

07

Preserved context

Don't silently drop what the user provided; keep their inputs visible and honoured.

Why it matters

When a system quietly loses context the user gave it earlier, it produces confident output that ignores real constraints, and because nothing looks wrong, the failure is hard to spot. Users shouldn't have to re-check that their own inputs were used.

What good looks like

The user's key inputs stay visible through the flow, and the output demonstrably reflects them.

Grounding

Microsoft HAX G12 (remember recent interactions); the silent-loss failure mode itself is AXIS's own contribution

08

Shown provenance

Make output verifiable: show sources, inputs used, or how it was produced.

Why it matters

Verifiability is the basis of trust in professional settings. If users can't check where something came from, they can't responsibly rely on it, no matter how good it looks.

What good looks like

Citations, source references, or a visible trail showing what the output was based on.

Grounding

Microsoft HAX G11 · Google PAIR (Explainability + Trust) · NIST AI RMF (accountable & transparent) · NN/g guidance on verifying AI citations