Keep the Evidence Visible: AI-assisted UX research: workflow, privacy, and trust

Use AI in your research without losing the research. Keep the Evidence Visible is the fully revised second edition of ChatGPT for SaaS (2023), rewritten for 2026 and broadened from SaaS to UX research everywhere. It is a practical guide for people who know UX work but are not AI experts: designers, researchers, product managers, and research-ops teams who are already pasting transcripts into a chatbot and can feel that something is wrong. AI can draft, summarize, compare, and code your research in seconds. It can also sound confident when it is wrong, hide uncertainty, and turn one participant into "users." This book teaches the one discipline that keeps AI honest: an AI-assisted finding is only as trustworthy as the evidence chain behind it. If you can see the path from raw data to observation to interpretation to recommendation, the work is reviewable. If it collapses into one fluent paragraph, you have a nice-sounding artifact and no research. What you will learn - A simple mental model, evidence layers, confidence levels, data classes, and named artifacts, that keeps every claim checkable and every decision human. - A five-step workflow for AI-assisted analysis that stops the model leaping to conclusions before the evidence is checked. - How to prompt like you are briefing a sharp but context-free colleague, with a full **prompt pattern library** you can lift and adapt. - How to use AI across the whole study: before (clarifying questions, choosing methods, recruiting and sample size), during (consent, transcription, capturing clean data), and after (single-session summaries, codebooks, inter-coder checks, honest synthesis). - How to protect participant data under GDPR when an AI tool is in the loop: lawful basis, consent that matches the processing, minimization and pseudonymization, retention, international transfers, and how to actually get a tool approved. - The AI failure modes that bite research teams, over-generalization, bias, synthetic users, and prompt injection - and the specific defense for each. >Grounded, current, and transparent Every model, vendor, and regulation claim is marked as a dated 2026 snapshot and was checked while preparing this edition. The book is built around one running example and worked case studies, including a failure library that shows the method stopping a bad study, not just polishing a good one. Two appendices give you reusable assets you will actually open: a prompt pattern library and a research-data handling checklist. This is book is about the habits, vocabulary, and templates that outlast the next ten model releases, so you can make AI useful without pretending that model fluency removes the discipline of research. If you want faster research that still holds up in a room full of skeptics, start here.

Pages: 134, Paperback, Independently published