{ "video": { "id": "L3RuP_q8Bwc", "title": "The UX of AI: Making AI-Powered Apps Your Users Don't Hate - Kathryn Grayson Nanz, Progress Software", "duration": 2159, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "Online" }, "analysis": { "video_id": "L3RuP_q8Bwc", "title": "The UX of AI: Making AI-Powered Apps Your Users Don't Hate - Kathryn Grayson Nanz, Progress Software", "one_liner": "A front-end/design advocate argues AI's real differentiator is no longer model performance but UX, and gives concrete interface patterns for five user problems: trust, clarity, control, transparency, and meaningful benefit.", "summary": "Kathryn Grayson Nanz argues that the knowledge gap between the people building AI features and the people using them is the widest she's seen for any technology, and that users get only a few bad experiences before they disengage permanently. Using the Macintosh System 1→System 6 evolution as a metaphor — we're currently around 'System 3' — she says we must borrow familiar UI patterns and adapt them, because AI can only remix existing patterns and AI-native interface conventions don't exist yet to remix. The bulk of the talk walks through five categories of user problem drawn from her user research (trust, clarity, control, transparency, meaningful benefit) with specific patterns for each: citations and action plans, streaming and visible reasoning, stop buttons and version history, granular permissions and cost estimates, and prompt templates plus next-step actions.", "key_points": [ "Five categories of user challenge, drawn from her user research, structure the whole talk: trust, clarity, control, transparency, and meaningful benefit — trust being 'by far, biggest hurdle'.", "Macintosh analogy: System 1 used literal turtle/rabbit speed icons and low/high volume pictures; System 6 dropped them for terms like 'rate of insertion point blinking' and 'RAM cache'. AI UX is 'probably somewhere around system 3' — past the turtle icons, but users are not yet experts.", "Don't defer this to AI: 'AI can only really remix things that already exist' and AI-interface patterns don't fully exist yet, so AI-generated UI 'tends to look pretty darn average' — a fine starting point, a bad ending point.", "Don't claim your AI is the trustworthy exception (most teams aren't training their own models anyway) — instead build 'trust, but verify' affordances: citations as tooltips (a snippet of relevant quote), inline links (for external sources), or a side panel for research work, which positions the assistant 'as more of a librarian than inherently a subject matter expert'.", "Citations serve a second purpose beyond vetting: they let users repurpose output into a report, email, campaign or presentation while keeping a trail of accuracy attached to their own name and credibility.", "For agentic work, show an action plan and get approval before the agent starts (as Claude and ChatGPT do), plus toggles for always-allow so repeated flows don't need babysitting.", "Mark AI-generated content as AI-generated — App Store disclaimer, visible or digital watermark, even an asterisk — because users' reactions range from enthusiasm to dismissing it as 'slop', and marking removes the sense that you tried to sneak it in.", "Streaming text does double duty: it disguises latency and lets users assess a partial response, stop it early, or form follow-ups. It also prevents the 'XKCD my code's compiling' disengagement where a 1-minute wait becomes a 5-to-10-minute side quest.", "Visible chain of thought lets users 'pinpoint where things went off the rails' between point A and F and give more specific corrections, mirroring how humans explain their rationale.", "Control patterns: a prominent 'emergency brake' / big red button that is not buried in a menu or behind a remembered command, plus version history scaled to the task — none for simple chat, undo/redo across ~10 steps for document editing, checkpoints/save states for advanced work. Git-style version control is more than a user-facing app needs.", "Granular revert matters more now 'as the cost of tokens is starting to tick up' — users should keep the good parts of a mixed-quality output instead of saying 'try again' and starting over.", "Permissions are not binary: reference a database vs. delete tables, read emails vs. send from the address, run scripts, search local files — each needs direct sign-off, plus a history view and the ability to revoke. A system that remembers must also be able to forget, with the user having final say.", "Users should know the time and money/token/credit cost before approving an action; a rough estimate is enough to let them revise. And when an agent takes over the browser, show a banner/sidebar/outline so the user knows they're no longer driving and doesn't interrupt it.", "Putting a blank text box in front of a user and saying 'ask AI' offloads a lot of work onto them; provide examples, templates, suggested prompts and guided workflows, then next-step action buttons and direct integrations (word processor, linked repo, ticket tracker) so output isn't 'more than a novelty'." ], "takeaways": [ "Stop selling your AI as trustworthy and instead ship verification affordances — link every claim back to source material via tooltips, inline links, or a research side panel, so users can check and can safely reuse the output under their own name.", "Never let an agent act invisibly: show the plan before it runs, show the reasoning as it runs, highlight what changed, badge AI-generated content, and show a clear banner whenever the agent is in control of the user's browser or session.", "Give a big, unhidden stop button and a version history sized to your task (nothing for chat, ~10-step undo/redo for document editing, checkpoints for advanced work), with granular revert so users don't burn tokens regenerating everything to fix one part.", "Model permissions as a spectrum, not a yes/no — per-scope (this folder, read-not-send), per-duration (once vs. always), with a visible history and revocation — and show time/cost estimates at the approval moment, even rough ones.", "Don't assume AI literacy: replace the empty 'ask AI' box with examples, templates, suggested prompts and guided workflows, and add next-step buttons/integrations that move the output into the tools the user actually works in." ], "topics": [ "ux", "design", "agents", "trust", "human-in-the-loop", "permissions", "streaming", "chat-interfaces" ], "tools": [ "Progress Software", "Macintosh (System 1, System 3, System 6)", "Microsoft Copilot (research agent)", "Microsoft Teams", "Claude", "ChatGPT", "Slack", "XKCD", "Nielsen's heuristics", "Git" ], "quotes": [ { "text": "Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should.", "at": "02:37", "url": "https://www.youtube.com/watch?v=L3RuP_q8Bwc&t=157s" }, { "text": "In situations where we cannot promise the truth, we have to go above and beyond to earn it.", "at": "12:03", "url": "https://www.youtube.com/watch?v=L3RuP_q8Bwc&t=723s" }, { "text": "AI is just another technology, and our users deserve clarity over dramatics.", "at": "19:09", "url": "https://www.youtube.com/watch?v=L3RuP_q8Bwc&t=1149s" }, { "text": "That means that the differentiator for the AI-powered software we build isn't performance anymore. It's the quality of the experiences that we can build around them.", "at": "34:56", "url": "https://www.youtube.com/watch?v=L3RuP_q8Bwc&t=2096s" } ], "words": 7323 }, "summary_url": "/#L3RuP_q8Bwc", "transcript": { "html": "/transcripts/L3RuP_q8Bwc.html", "txt": "/transcripts/L3RuP_q8Bwc.txt", "vtt": "/transcripts/L3RuP_q8Bwc.vtt" } }