# AI UX Evidence Library

> Plain-English, source-checked takeaways from research papers on designing AI products. Each question has a one-line answer, two ratings (evidence quality and fit with LLM products), concrete design rules tagged as from-the-study or our-reading, study details with verified numbers and quotes, trade-offs, and full citations.

Every number and quote is checked against the paper text. Each entry is also available as clean Markdown by adding .md to its URL.

## Design questions

- [Who benefits most from AI tools at work?](https://aiuxevidence.com/e/who-benefits.md): Less experienced and lower-skilled workers gain most from AI assistance; top performers gain little, and everyone does worse on tasks the AI handles badly.
- [Can you trust citations in AI search answers?](https://aiuxevidence.com/e/citations.md): Not by themselves: in an audit of four generative search engines, only about half of generated sentences were fully backed by their citations.
- [Do inline AI suggestions like Copilot cost users time?](https://aiuxevidence.com/e/inline-suggestions.md): Yes: in a study of GitHub Copilot, the biggest single slice of programmers' time went to thinking about and verifying its suggestions.
- [How should AI show confidence or uncertainty to users?](https://aiuxevidence.com/e/show-confidence.md): Showing uncertainty helps people trust the AI the right amount; plain first-person hedges also raised accuracy, but a score alone didn't.
- [How do you stop users from blindly trusting AI answers?](https://aiuxevidence.com/e/make-users-think.md): Small speed bumps, like having people decide before they see the AI's answer, cut blind acceptance of AI answers (overreliance), though people liked them less.
- [Can AI writing assistants change what users believe?](https://aiuxevidence.com/e/nudging-opinions.md): Yes. An opinionated AI writing assistant shifted both what people wrote and what they said they believed afterwards.
- [Do AI explanations help users trust AI appropriately?](https://aiuxevidence.com/e/explanations-trust.md): Not by default: AI explanations often raise acceptance of AI answers whether they're right or wrong. They help when checking the explanation is easier than doing the task.
- [Should users be able to edit AI-generated output?](https://aiuxevidence.com/e/let-users-edit.md): Yes. People abandon algorithms after seeing them err, but even a small ability to adjust the AI's output keeps them using it.
- [When should an AI agent ask for confirmation before acting?](https://aiuxevidence.com/e/ask-before-acting.md): An AI should ask before irreversible or uncertain actions, and act on its own only on confident, easy-to-undo steps.
- [Is a blank chat box good UX for AI products?](https://aiuxevidence.com/e/blank-prompt.md): No. Non-experts struggled to write prompts for an LLM, overgeneralized, and treated the model like a person, so a bare chat box leaves the hard part to users.
- [How do you set user expectations for an AI that makes mistakes?](https://aiuxevidence.com/e/set-expectations.md): Tell users up front how accurate the AI is and how it fails; in experiments, that plus tuning toward the right error type raised acceptance of an imperfect AI.
- [How should you onboard users to an AI feature?](https://aiuxevidence.com/e/onboarding.md): Before relying on an AI, experts wanted to know its strengths, limits and what it was built to optimize, more than how its model works.

## Method

- [How we rate the evidence](https://aiuxevidence.com/how-we-rate): rubric for evidence quality and fit with LLM products

## Optional

- [Full text of all entries](https://aiuxevidence.com/llms-full.txt)
- [Homepage](https://aiuxevidence.com/)
- [Newsletter](https://aiuxevidence.beehiiv.com)