The canon

The published frameworks side by side: who publishes each, what it actually contains, what it is genuinely good for, where it is thin, and what its publisher sells. This is the page to reread before a meeting. Every characterisation here was taken from the publisher’s own page; where a detail could not be verified, it is absent rather than softened.

Naming a publisher’s interest is not a dismissal. Microsoft, Google and Apple all sell AI products and platforms; NN/g sells training and consulting; Shape of AI is one consultant’s catalogue. Reading that alongside the work is part of reading it correctly. A vendor is often the only primary source for its own product, and the HAX guidelines in particular are the best-evidenced thing in this field.

Guidelines for Human-AI Interaction (HAX Toolkit)

Microsoft — Amershi et al., CHI 2019

Dates. Presented at CHI 2019, Glasgow, 4–9 May 2019. No revision date published; the toolkit page still points at the 2019 paper.

What it contains: Eighteen numbered guidelines, grouped by a four-phase interaction timeline. The toolkit around them is three further artifacts: the HAX Design Library (patterns and examples), the HAX Workbook (a team exercise for prioritising which guidelines to implement) and the HAX Playbook (common failures for NLP applications). The guidelines page itself does not print the list — the full table is in the paper.
What it is good for: The single most quotable artifact in this field, and the only one with a published evidence base you can describe: 168 candidate recommendations consolidated to 20, cut to 18 by eleven researchers inspecting thirteen products, then validated by 49 practitioners against twenty products across ten categories. If you cite one thing accurately, cite this.
Where it is thin: Derived entirely from prediction and recommendation systems — Table 2 lists e-commerce recommendations, navigation, music recommenders, activity trackers, autocomplete, feed filtering, email importance filtering, voice assistants, photo organisers and web search. Nothing generative, nothing agentic. Its failure vocabulary is false positives and false negatives, which does not describe a committed action. Nothing about data. Section 7 says all of this in the authors’ own words.
The publisher’s interest: Microsoft sells AI products and platforms and publishes design guidance about the category it sells. Mitigating: the work came out of Microsoft Research, was peer-reviewed at CHI, and names its methodology in enough detail to be criticised.

People + AI Guidebook, third edition

Google — People + AI Research (PAIR)

Dates. First published 8 May 2019. A second edition followed, introduced at Google I/O. Third edition updated April 2025 — the generative-AI update.

What it contains: Six chapters following a product lifecycle: User Needs + Defining Success, Data + Model Evolution, Mental Models + Expectations, Trust + Explanations, Feedback + Controls, Errors + Graceful Failures. Plus patterns, downloadable workshop kits, case studies and a glossary. CC BY-NC-SA 4.0, so a workshop kit can be run inside a client engagement with attribution.
What it is good for: The decisions made before an interface exists — whether this should be an AI feature at all, where to draw the automate-versus-augment line, and everything about data, none of which the HAX guidelines touch. Also the framing worth stealing: trust is something to calibrate, with over-trust treated as a design failure, not something to maximise.
Where it is thin: Its error material describes wrong outputs, not committed actions. And the first edition is still live on the same domain, undated and unmarked, at /old-gb/ and at /chapter/… URLs that are the ones in the site’s own sitemap — so a citation with /chapter/ in it is a citation of a superseded document.
The publisher’s interest: Google sells AI products and platforms; the third edition arrived as Google was shipping Gemini, and the worked examples are Google-flavoured throughout. Mitigating: PAIR sits in Google Research rather than a product org.

Human Interface Guidelines — Generative AI

Apple

Dates. New page 9 June 2025. Revised 8 June 2026 — refining and reverting results, feedback during generation, choosing a model type.

What it contains: Nine sections: Best practices, Transparency, Privacy, Models and datasets, Inputs, Outputs, Continuous improvement, Platform considerations, Resources. Note there is no “machine intelligence” grouping in the HIG — this page and Machine learning sit under Technologies, and there are no HIG pages for Apple Intelligence, Writing Tools, Genmoji or Image Playground.
What it is good for: The only guidance in the canon that addresses a system taking an action: “Consider consequences and get permission before performing irreversible or potentially problematic tasks” — avoid automating destructive actions, ask for confirmation before a significant action on someone’s behalf. Also the hardest line on disclosure anywhere: “Never trick someone into thinking they’re interacting with or viewing content authored by a human if they’re actually interacting with AI.”
Where it is thin: That action guidance is a single bullet, framed entirely as permission before a task starts. Nothing about a run already underway, partial completion, or side effects already committed. And it is platform guidance, not research — no methodology, no evidence base, no sample.
The publisher’s interest: Apple is a platform vendor documenting its own products; every worked example is an Apple feature. This guidance also sits adjacent to App Review, which makes it closer to policy than to advice — a different kind of document from the other two, and worth saying so when it gets quoted alongside them.

Human Interface Guidelines — Machine learning

Apple

Dates. Consolidated into one page 2 May 2023; art added 24 October 2023; minor clarity updates 8 June 2026.

What it contains: A taxonomy — place your feature on five axes: critical or complementary, private or public, proactive or reactive, visible or invisible, dynamic or static — followed by nine patterns: Explicit feedback, Implicit feedback, Calibration, Mistakes, Corrections, Multiple options, Confidence, Attribution, Limitations.
What it is good for: The Confidence pattern, which is the sharpest thing any vendor says about showing confidence and cuts directly against a per-field confidence UI: “If you’re not sure how your confidence values correlate with the quality of your results, it’s not a good idea to convey confidence to people.” The Mistakes pattern is also unusually practical: match the corrective tooling to the seriousness of the consequence, and be especially careful with proactive features, where people have less patience for errors they did not invite.
Where it is thin: Explicitly the non-generative page — it points at the Generative AI page for generative work, and its examples are QuickType, Face ID, Photos, Apple Music and Memoji. Same era and same assumptions as the HAX guidelines.
The publisher’s interest: Platform vendor, same as above.

AI UX research

Nielsen Norman Group

Dates. A rolling stream of articles, individually dated, from 2023 to the present.

What it contains: Two very different kinds of thing under one masthead. Small qualitative usability studies that are NN/g’s own work — chatbot studies at n = 9, early generative-AI behaviours at n = 8, an agent study at n = 6. And summaries of other people’s large academic studies, which is where every big number on the site comes from.
What it is good for: Concrete observed behaviour that no framework contains: users treating chatbots as search bars and shortening to keywords as confidence grows; disliking “great question!” sycophancy; rarely clicking citation links while still trusting the presence of citations. Useful precisely because it is behaviour rather than principle.
Where it is thin: The famous 66% productivity figure is not an NN/g study — it is an average of three external controlled studies on a 2023-vintage model, two of which measured first-time use. Quoting it as “NN/g found” is wrong, and knowing that is a cheap way to be the most careful person in a meeting. Where the sample is NN/g’s own, it is usually single digits. Treat the findings as hypotheses, not effect sizes.
The publisher’s interest: NN/g sells training courses, consulting and paid reports, advertised on the same pages as the free research. The incentive runs toward findings that imply a need for expert help.

Shape of AI

Emily Campbell — an individual, not an organisation

Dates. No launch date, no last-updated, no changelog. The only year anywhere is a 2025 copyright line. Whether it is actively maintained is unverified.

What it contains: Around sixty UX patterns for AI interfaces, filed under seven types: Wayfinders (getting a first prompt written), Prompt actions, Tuners (refining the input), Governors (“human-in-the-loop features to maintain user oversight and agency”), Trust builders, Identifiers, and Dark matter (“potentially nefarious, but certainly ambiguous patterns”). Free, CC BY-NC-SA.
What it is good for: Vocabulary. It is the fastest way to acquire the words a room is already using, and the Governors category is the closest published name for what a review gate is. Naming a pattern is often all a conversation needs.
Where it is thin: No stated methodology and no citations to research. It is curation from observed products — a record of what shipped, not evidence about what worked. Undated, so you cannot tell how current any entry is. Treat it as vocabulary and never as authority; citing it as evidence is the mistake that makes a designer sound well-read and wrong.
The publisher’s interest: One person, who also consults on AI design and product strategy, with a newsletter and a community attached and courses signposted as coming. Small and disclosed — a different order of interest from a platform vendor, and still worth naming.

All eighteen HAX guidelines, verbatim

Titles and qualifying sentences as printed in Table 1 of the CHI 2019 camera-ready paper. The toolkit website does not print this list. The caption to Table 1 describes the guidelines as “roughly categorized by when they likely are to be applied during interaction with users” — so treat the phase boundaries as approximate and argue with what a guideline says, not with where it was filed.

Initially · G1–G2 · 2 of 18

G1Make clear what the system can do.: Help the user understand what the AI system is capable of doing.
G2Make clear how well the system can do what it can do.: Help the user understand how often the AI system may make mistakes.

During interaction · G3–G6 · 4 of 18

G3Time services based on context.: Time when to act or interrupt based on the user’s current task and environment.
G4Show contextually relevant information.: Display information relevant to the user’s current task and environment.
G5Match relevant social norms.: Ensure the experience is delivered in a way that users would expect, given their social and cultural context.
G6Mitigate social biases.: Ensure the AI system’s language and behaviors do not reinforce undesirable and unfair stereotypes and biases.

When wrong · G7–G11 · 5 of 18

G7Support efficient invocation.: Make it easy to invoke or request the AI system’s services when needed.
G8Support efficient dismissal.: Make it easy to dismiss or ignore undesired AI system services.
G9Support efficient correction.: Make it easy to edit, refine, or recover when the AI system is wrong.
G10Scope services when in doubt.: Engage in disambiguation or gracefully degrade the AI system’s services when uncertain about a user’s goals.
G11Make clear why the system did what it did.: Enable the user to access an explanation of why the AI system behaved as it did.

Over time · G12–G18 · 7 of 18

G12Remember recent interactions.: Maintain short term memory and allow the user to make efficient references to that memory.
G13Learn from user behavior.: Personalize the user’s experience by learning from their actions over time.
G14Update and adapt cautiously.: Limit disruptive changes when updating and adapting the AI system’s behaviors.
G15Encourage granular feedback.: Enable the user to provide feedback indicating their preferences during regular interaction with the AI system.
G16Convey the consequences of user actions.: Immediately update or convey how user actions will impact future behaviors of the AI system.
G17Provide global controls.: Allow the user to globally customize what the AI system monitors and how it behaves.
G18Notify users about changes.: Inform the user when the AI system adds or updates its capabilities.

Dating the agent era

Four dates, so that “this material predates tool-calling agents” is a checkable claim rather than a dismissal.

4–9 May 2019 — the eighteen guidelines are presented at CHI 2019.
6 October 2022 — ReAct is posted to arXiv; ICLR camera-ready follows 10 March 2023.
22 October 2024 — Anthropic ships computer use in public beta.
25 November 2024 — the Model Context Protocol is announced.

Three traps, in one place

The superseded PAIR edition. The 2019 first edition is still live at /old-gb/ and /chapter/…, undated and unmarked, and those are the URLs in the site’s sitemap. Cite /guidebook/ or you are citing a document Google replaced twice. The tell is the chapter names: “Explainability + Trust” is the old one, “Trust + Explanations” the current.
The 66% figure. Not an NN/g study. An average of three external controlled studies on a 2023-vintage model, two of which measured first-time use only.
Paraphrase drift. HAX G11 is “Make clear why the system did what it did.” Microsoft’s own generative-AI page renders it “Make it clear why the system did what it did.” Harmless there; the mechanism is not. Quote from the paper, not from a page quoting the paper.

Every framework on these pages links to the page its publisher actually wrote. If a lesson characterises a guideline in a way its own source does not support, check the resource list and tell your teaching agent — a mischaracterised framework makes every disagreement built on it worthless.