UX Brighton 2026 starts in
Buy tickets

What happens when AI can do your job?

Florian Herrengt · UX Brighton 2026

florianherrengt.com · read my blog

Watch the short intro · Open the slides · Download the PowerPoint · GitHub repo

The four stories · Never correct the AI · Reading list · Sources

What the evidence tells us

The Stanford AI Index compares AI and human performance on specific tests. To read the chart from the talk, look at what each test measures before drawing conclusions about a job. A model can do well on a task and still need help when the brief is unclear, evidence is missing or several decisions depend on one another.

UK software job ads were already falling before ChatGPT launched. On Indeed, software development ads had fallen by more than a quarter from their April 2022 peak by the time ChatGPT launched that November. In October 2026 they were about 38% below their pre-pandemic level. Indeed tracks ads at all levels of experience.

The number of people working as programmers and software developers generally rose from 2006 to early 2023. It then fell by about 70,000 into 2024, the largest fall in the period NFER studied. The way the survey grouped jobs changed in 2021, so comparisons across that date are less precise.

Some tech employers told NFER they had hired too quickly as the UK came out of the pandemic. They also pointed to weaker demand, rising costs, more work moving overseas and early effects of AI. The interviews cannot tell us how much each factor contributed.

LinkedIn tracked a separate measure. In April 2026, UK entry-level hiring was down 28% year on year for graphic designers and 27% for software engineers. Across industries, entry-level hiring broadly tracked the wider market. The report cannot tell us whether AI contributed to the declines.

Yale’s October 2026 analysis found no clear link between measured AI use and changes in US employment or unemployment.

In Stanford’s US payroll sample, employment among workers aged 22 to 25 in jobs more exposed to AI was 19% below where it would have been had it kept pace with their less exposed peers. Most of the gap came from fewer hires. The study could not establish how much AI caused, and a separate study found a much weaker link between AI exposure and junior hiring after accounting for remote work.

The four stories

Redesign the factory

Start with the result you need. If you were building this from scratch with AI at the centre, how would you let it do the whole job for you? Work out what it needs to know, what it can do on its own and how it will check the result.

Look back at the old process once you have a design. Carry over the steps that still give the system information it needs, prevent a failure or leave a decision with the person who needs to make it.

Look at what the red flag protects

For every check or approval, ask what you are trying to stop from happening. Can you change the system so the error cannot happen, or can you catch it automatically before anyone uses the result? If a person still needs to check, ask what they can see or decide that the system cannot. Ask again when the system changes.

When the system stops for review, show what it checked, what evidence it found and what the reviewer needs to decide. Let them correct the result or send it back so the system can pick up where it stopped.

The 1865 law also gave people with horses a way to stop the locomotive. If one of them raised a hand, the driver had to stop at once. Think about who can stop your system if it gets something wrong.

Know what problem you solve

Durant used the shops and salespeople from his carriage company to sell Buicks.

You could switch from Sketch to Figma and pick up the same job where you left off. AI can change how much of the work you do yourself, but you’re still there to solve the same problem.

Understand the machine

Knowing a little HTML and CSS helps you design for the web because you understand how a page is built. You can think through what happens when the screen size or the content changes and talk to developers about how the design should behave.

Learn how AI works in the same way. Pick something you can’t explain yet and find out what causes it. Run the same request a few times to see what varies on its own. Then change one thing, predict what will happen and see whether you were right. If the result surprises you, find out what the system did before you move on.

Seeing those steps makes it easier to work out whether a problem starts with the information the model gets, the model itself or the way the app uses its answer.

You are paid to make good decisions

AI makes it easy to produce work nobody asked for. The slop is quick to make, but someone still has to read it, check it and work out whether any of it helps.

Use that extra power to make better decisions or to make more of them with the same care. You can check an assumption you would have had to take on faith, or get enough evidence to make a call you would have put off.

Never correct the AI

When you fix an AI result, save the input that led to the mistake and record what should have happened. Use it as a test each time you change the system.

Ask AI to write and run code when you can describe the task as a clear rule. When a task needs judgement, ask which parts of the result you can verify with a fixed rule and what you should compare them against.

Give the check a source you trust and define what it takes to pass. If you only have an image, use OCR to read the text before checking it. Send unclear readings for review.

Before you rely on a check, see if it accepts a result you know is right and catches one you know is wrong. Run it on every result and hold any that fail.

Reading list

How technology changes work

AI use and design work

Jobs and hiring

UX and service design

Building and testing AI systems

Sources

Benchmarks and jobs

The four stories

Checking AI results

From my blog

There are things you could just stop doing and things would move faster – Most “Best Practices” are cargo cults

Build systems you can run, debug, restore and replace – Good engineers make boring choices

You can make your own numbers look incredible while reducing the throughput of the entire team – AI is removing the middle class of software engineering

Read the blog