The public conversation runs between two poles: everything is about to be automated, or nothing will really change. Neither describes what is actually happening.
A more useful question replaces the job with the task.
Jobs are bundles of tasks, and the tasks are affected unevenly. What is most exposed: text and image work that tolerates error and carries no accountability. What is least exposed: anything physical, anything requiring someone answerable for the outcome, and anything where the difficulty is people rather than output. Most roles lose some tasks and keep the rest.
Why "will my job disappear" is the wrong question
Almost no job is one activity. A bookkeeper reconciles accounts, chases clients, spots anomalies, and explains bad news to a business owner. A designer produces artwork, interprets vague briefs, argues for decisions, and manages revisions.
Automation has historically taken tasks, not whole jobs — and then the job reshapes around what remains. Cash machines did not eliminate bank branches; they changed what branch staff spent their day doing.
So the practical question is: which of my tasks are cheap to automate, and what fraction of my week are they?
What is genuinely exposed
Three conditions together, not individually:
It is text or images in, text or images out. No physical component.
Errors are tolerable and easy to catch. A wrong sentence in a product description is noticed and fixed. A wrong dosage is not that kind of error.
Nobody is accountable in a formal sense. No signature, no licence, no liability.
Where all three hold — routine marketing copy, basic translation, first-draft graphics, simple summarising, boilerplate code, transcription — the work is already changing, and the change is visible in what freelance markets pay.
What is not
Physical work. Robotics is a separate and much slower problem. Plumbing, care work, trades, logistics handling.
Accountable work. A model can draft a legal opinion, a diagnosis, a structural calculation. It cannot be answerable for one. Where a professional signs their name and carries liability, the work does not transfer — it gets assisted.
Work whose difficulty is people. Negotiating between parties who want incompatible things. Delivering bad news. Persuading someone who does not want to be persuaded. The hard part was never the document.
Work requiring trust and continuity. Clients who hire a person, not an output.
Anything needing genuinely current, local, verified knowledge. Models are unreliable on the specific and recent — which is exactly where much professional value lives.
Between producing an output and being responsible for it. AI is very good at the first and structurally incapable of the second. Roles that are mostly production are exposed. Roles that are mostly judgement, accountability and relationships are not — even when the production part of them changes completely.
The pattern that keeps repeating
In most fields where these tools have been adopted seriously, the same thing happens:
The first draft gets cheap. The judgement gets more valuable.
A marketing writer who produced five posts a day now produces twenty drafts and spends their time deciding which are any good, fixing what is wrong, and knowing what the client actually meant. Output up, and the skill has moved from writing to editing and judgement.
That is genuinely disruptive for anyone whose value was volume of production. It is broadly positive for anyone whose value was knowing what good looks like.
What to actually do
Skip the forecasts. Run the experiment on yourself.
- List your tasks for a week, honestly, with rough hours.
- Try the tools on each one. Not a demo — your real work.
- Sort into three piles: it does this well; it helps but needs heavy correction; it cannot do this.
- Add up the hours in pile one. That is your actual exposure, measured rather than guessed.
This takes an afternoon and tells you more than any report, because it uses your job rather than a category.
Then:
- If pile one is large, start deliberately moving toward the tasks in pile three. That is a multi-year move, and you have time if you begin now.
- If pile one is small, learn the tools anyway for the leverage, and stop worrying.
- Either way, be the person who uses them well. In most fields the near-term risk is not being replaced by AI; it is being outpaced by a colleague who has learned to use it and produces three times as much.
On the predictions
Treat all confident numbers sceptically, in both directions.
Forecasts of technological unemployment have a long and poor track record — consistently wrong about which jobs, and about timing. Many current ones come from organisations selling either the technology or protection from it.
What can be said with reasonable confidence: the mix of tasks inside many jobs is changing, faster in some fields than others, and the people most affected are rarely the ones the headlines name.
Frequently asked questions
Which jobs are most exposed to AI?
Those made mostly of text and image work with a tolerance for error and light accountability — routine copywriting, basic translation, first-draft graphics, simple summarising. Exposure tracks tasks, not job titles.
Which jobs are least exposed?
Work involving physical presence, legal or professional accountability, negotiation of conflicting interests, or responsibility for a decision. A model can draft a recommendation but cannot be answerable for it.
Will AI create jobs as well?
It has so far in previous automation waves, but the new jobs rarely go to the same people or places as the lost ones. That mismatch is the real problem, and averages hide it.
What should I actually do about it?
Use the tools on your own work for a month and see what they genuinely do well. That gives you a more accurate picture of your exposure than any forecast, including this article.
Are the job-loss predictions reliable?
Treat all of them sceptically, in both directions. Forecasts of this kind have a poor historical record, and many are produced by organisations with something to sell.
Found an error? Email us and we will fix it and note the change at the bottom of this article. Hello@daily-atlas.com
