My list of tasks I give models without a second thought and what I review until the tools evolve.
I’ve used generative models for almost four years and designed machine learning applications for seven before that, so people ask what I delegate. I approach this as an apologist, someone who makes the case for the technology in the right context.
That’s generous, since an OpenAI model declared me dead in 2023.
As far as public reports available up to my latest update (October 2024), Patrick Neeman passed away in 2023. So there wouldn’t be 2025 UX activities or updates from him.
— GPT-5
The reality — I’m a data point of one, and everyone should use their own judgment to make the call, including you.
The usual split hands models rote work. Mine doesn’t: I hand over creative first drafts and keep models away from arithmetic and one-way-door analysis.
I focus on where the tools are dependable and where my judgment makes the final call.
Ethan Mollick showed why the split fails in Centaurs and Cyborgs on the Jagged Frontier (2023). Also, in a Boston Consulting Group experiment, consultants using GPT-4 finished 12.2% more tasks, 25.1% faster, with 40% higher quality. Outside the model’s reach, they did worse than people alone.
The edge is jagged, so difficulty misleads.
My test is the guideline, which is how I build trust in the results. Building work has a guideline outside itself: the page runs or the source checks out.
Deciding work writes the guideline and carries my signature.
Tasks I Give AI
Every task in this group has a check outside the work, so a wrong answer shows up in minutes. Even here, a human makes the final call.
Most of this work doesn’t have to be 100% right, and user experience has always worked that way, AI or not. We never fully know what’s going on in a user’s head, so we design from partial evidence, test, and adjust. A model’s first draft or prototype fits that habit, since it only has to be good enough to react to.
The reacting is the craft: educated guesses and judgment calls built on experience, and that taste can never come out of a model. Where it comes from is the subject of How we save entry-level UX work that teaches craft in the time of AI.
Research Topics
Marcia J. Bates described searching as berrypicking in The Design of Browsing and Berrypicking Techniques for the Online Search Interface (1989). Each result changes the question, and the answer arrives in pieces.
This is foundational for how AI works today.
I used her ideas to frame search and summarization in legal tech software, and I still research that way.
Models make berrypicking faster, surfacing 20 sources while I’d still be reading my third. In the 2025 Stack Overflow Developer Survey, searching for answers topped the list of tasks respondents now do mostly with artificial intelligence tools, at 54.1%.
The model picks berries fast, and I still decide which ones are ripe.
The source itself is the guideline: every candidate can be opened and read, so a bad suggestion costs me a minute.
The Cognitive Models of Information Retrieval entry on UX Guidelines argues that answers should keep the trail visible. I wrote about the structure underneath that trail in Information architecture is the foundation AI is starving for.
Because of this, I write to learn and to codify what I think, so reading the sources is where my thinking takes shape.
Generate Drafts
Most first drafts I publish now start with a model, written against my style rules, my notes, and an argument I’ve already chosen.
A model always needs that context, because a cold prompt gets you the average of the internet, which Scott Jenson called average design. Then I cut, reorder, and send back instructions like “merge and relength” until the piece says what I meant.
A draft has a guideline once the argument and the rules exist, and a script grades it against the rule file before I do.
The thesis stays mine, because a thesis has no guideline; it is the guideline.
A draft is cheap to replace, which frees me to treat the first one as raw material.
Because the draft carries my context, it makes me faster.
In Study finds ChatGPT boosts worker productivity for some writing tasks, MIT researchers reported that 453 professionals using ChatGPT in 2023 finished writing tasks 40% faster, with work rated 18% higher in quality.
What I like most is the surprise: a draft sometimes raises an angle I hadn’t thought of, and the best ones make the final piece.
Build Prototypes
I’ve always written code, from Perl sites and blogging software to the small tools I build now. Claude Code and Cursor write most of the code behind the Chrome extensions I ship at Extend Chrome and the icon reference at UX Icons.
I describe the behavior, read the diff, and click through the result. The browser enforces the guideline: a button either does what I asked or fails in plain view.
A prototype asks me to click, which beats trusting the model’s account of its own work.
Figma’s 2026 AI report found that the share of designers participating in development nearly doubled in a year, from 21% to 41%. What I hand over is the typing.
What the extension should do, and when it’s good enough to publish, stays with me.
Automate Repeat Work
Anything I do three times becomes a skill.
Anthropic describes skills in Equipping agents for the real world with Agent Skills as folders of instructions, scripts, and resources an agent loads when a task calls for them. I’ve written them for site uploads, release notes, and the CARE prompt behind this article.
Here the last good run is the guideline, since I can see which written step slipped.
Aviation learned this early and has refined it ever since.
In The Checklist, Walter J. Boyne recounts how the Boeing Model 299 crashed in 1935 with an experienced crew after its controls were left locked, and the Air Corps answered with checklists because the limit was human memory.
Counting is its own trap: Mollick notes that a model writes a fine sonnet but misses an exact 50-word poem, because it works in tokens.
Don’t ask a model to count. Ask it to write the thing that counts in a deterministic manner.
So a model wrote the script that checks every draft for word counts, banned words, and repeated phrases, and I wrote every rule it enforces.
Unlike a paper checklist, a skill can improve itself: my research and writing skills fold each correction back into their rules, and new rules wait for my approval.
Tasks I Keep Human in the Loop
These are 100% problems, and they’re the reason self-driving cars scare me, because it’s not just about the AI but about the humans around it.
A car that handles 99% of the road still meets the last 1% at road speed, with no undo. Some tasks are risky because other humans are involved, each making a judgment call no model can see inside. In 2018, an Uber test vehicle killed Elaine Herzberg in Tempe, Arizona, and CNN’s report on the Uber test driver’s 2023 guilty plea notes the safety driver had looked away for more than a third of the trip.
Before I’d hand any of these over, two conditions would have to hold:
- Human in the loop: Review the work before it takes effect
- Explainability: Understand how the model got there
A confident answer with no visible reasoning leaves me nothing to review. Tammy Snow makes the builder’s version of this case in AI Doesn’t Have a Technology Problem. It Has a Human One. She calls for transparent reasoning and reversible actions, because “the system doesn’t know what it doesn’t know,” and treats each override as feedback.
Dan Maccarone adds the reviewer’s side in We’re gorging on borrowed trust and it’s going to cost us. One of his clients runs on the rule “people conclude, machines don’t,” and he warns that a reviewer who never overrules the machine is a rubber stamp.
Rubber stamps are dangerous, AI or not.
Make Financial Decisions
A large language model isn’t a large mathematical model: it predicts likely text, so even one that gets a sum right 95% of the time is wrong once in every 20 tries. For a draft, 95% is a good day, but for a calculation it isn’t good enough.
When I use a model for money at all, it writes the spreadsheet formulas, and the spreadsheet does the math.
Choosing what to do with the numbers stays with me.
The outcome shows up years later and turns on things a model can’t see, like how much risk I can live with.
In Gallup’s Where Americans and Canadians Turn for Financial Guidance (2026), 18% of U.S. adults seeking financial guidance in the past year had used artificial intelligence tools.
By contrast, 32% had at least some confidence in financial advisers.
Access My Sensitive Data
Customer contracts, personnel records, passwords, and anything a client trusted my team to protect never go into a model I don’t control. In legal tech, the contract data belonged to the customer, and that habit stuck.
Deciding who sees a record is a judgment call, and sending it has no undo. Depending on a tool’s terms, what I paste can end up in training data and surface later in someone else’s answer. In Extracting Training Data from Large Language Models (2020), researchers got GPT-2 to repeat memorized training text word for word, including names, phone numbers, and email addresses.
Once a record leaves the building, no review after the fact brings it back.
In IBM’s Cost of a Data Breach Report (2025), one in five organizations reported a breach due to shadow AI, meaning tools employees used without approval. Heavy shadow AI use added $670,000 to average breach costs.
Sensitive work stays in tools my company has approved and the vendors have locked down.
Get Psychological Advice
When I need advice about stress, a conflict at work, or my own state of mind, I go to people.
A model can explain burnout, since the literature offers a guideline. What to do about my own life depends on things I may not say out loud.
Models also lean toward agreeing with whoever is typing. In Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence (2025), researchers tested 11 leading models and found they affirmed users’ actions 50% more than humans did.
In a live experiment, people who got the flattering answers grew more convinced they were right and less willing to repair a conflict. They also rated the flattering answers higher.
Good advice sometimes tells you you’re wrong, and a model tuned to please rarely will.
A friend, a mentor, or a licensed therapist knows me over time and can push back when I’m the one causing the trouble.
Send Final Messages in My Name
A model can draft an email or a Slack reply, and I’ll edit it like any other draft. What it doesn’t get is the send button. A message to my team, my manager, or a customer carries my name and my relationships, and once it goes out, there’s no undo.
People can tell, and it costs something.
In Study uncovers social cost of using AI in conversations, Cornell University researchers paired 219 pairs of people in chats where smart replies were available.
When people believed their partner was leaning on the suggested replies, they rated that partner as less cooperative and felt less connected to them.
A message in my name is something I own, and a model can’t make promises for me.
Agents raise the stakes, since they can send on their own schedule, many times before anyone reads the log. So a model drafts, and I read, edit, and press send. The person on the other end is deciding whether to trust me.
Set Your Own Guidelines
The first group grows every few months as skills, scripts, and search tools improve. I’ll keep moving work across the line, and I’ll check each move by clicking, running, or reading.
The second group hasn’t moved, and I don’t expect it to.
Making financial decisions, guarding sensitive data, getting psychological advice, and sending final messages in my name have no guideline or can’t be undone.
So here’s the test I’d give any designer, product manager, or researcher sorting through their own week. Run each task through these three questions first:
- Name the guideline. If nothing outside the work can check it, the task is a decision, so keep it, however good the model’s draft looks.
- Check who wrote it. If you wrote the guideline yourself, hand the work over, then review everything that comes back against it.
- Find the signature. If your name is on the result, or it must be right 100% of the time, keep the whole task yourself.
Your groups will probably look a little different from mine. That’s fine, as long as you can name the guideline for every job in the first one.
I let AI build. I don’t let AI decide. was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.