TASK
What is delegated?
Drafting, searching, calculating, recommending, deciding or acting have different exposure.
AI vulnerability
See where AI may reshape your work
The practical risk is not only job replacement. Capability can also weaken through routine delegation, over-trust and invisible responsibility.
What it means
02AI vulnerability is about tasks, not job titles
AI vulnerability is the degree to which a piece of work can be done by AI without a person losing anything that matters — and the degree to which handing it over quietly costs you something you needed to keep.
That second half is the part most people miss. A task can be safely delegated and still leave you worse off, because the capability behind it fades when you stop using it. The risk is rarely a sudden replacement. It is a slow trade: speed today for judgment later.
Nobody's job is a single task. Every role is a bundle of dozens, and AI does not take the bundle — it takes some threads and leaves others. Which threads go, and what is left holding the work together, is what this page is about.
The question is not "will AI take my job?" It is "which parts of my work am I about to stop being able to do?"
The frame
03Vulnerability is contextual
A skill is not simply automated or safe. Vulnerability changes with the task, consequence, data, verification and the human capability that remains active.
WHY
What is at stake?
Convenience, learning, money, safety, rights and relationships require different oversight.
WHO
Who can verify?
Capability matters: a person cannot meaningfully supervise reasoning they do not understand.
OWN
Who is accountable?
Responsibility must remain visible even when the work is AI-assisted.
Pathways
Four pathways: how capability becomes vulnerable
Automation exposure
01A task becomes easy to delegate, reducing the need to perform it manually.
- repeatable
- structured
- low-context
Deskilling exposure
02A person stops practising the underlying reasoning and becomes less able to work independently.
- atrophy
- dependency
- shallow recall
Over-trust exposure
03Fluent output is accepted without enough evidence, challenge or source checking.
- bias
- fabrication
- false confidence
Accountability exposure
04It becomes unclear who made the choice, checked the result or owns the consequence.
- opacity
- handoff
- responsibility gap
Research pulse · R7
One study from this field
Four experiments with 3,562 participants found GenAI collaboration improved immediate performance, but gains did not persist on later solo tasks; intrinsic motivation also fell and boredom increased.
Wu et al. (2025), Scientific Reports.
What it looks like
05Six ordinary moments
None of these are disasters. That is exactly why they are worth noticing.
The email you no longer write
01You used to think through how to say the difficult thing. Now you describe the situation and edit what comes back. The message goes out fine. Six months later, a conversation goes badly and you realise you have not practised finding the words in a long time.
DESKILLING
The formula you stopped understanding
02The spreadsheet works. You asked for it, you pasted it in, the numbers look right. When someone asks why the Q3 figure moved, you can point at the cell but you cannot explain the logic — and you are the one who signed off on it.
ACCOUNTABILITY
The summary you forwarded
03A long report, a clean summary, on to the next thing. The summary was accurate about what the report said and silent about what the report left out. You passed on a confident version of something you never actually read.
OVER-TRUST
The code that runs until it doesn't
04It shipped, it worked, everyone moved on. Three weeks later it fails in production at 2am and the person on call — you — is reading a function they have never really understood, under time pressure, for the first time.
DESKILLING
The lesson that is fine
05The plan is competent. The activities work. The students are fine. But it is not built on your read of this particular class, and next year you will start from the tool again rather than from what you learned this year.
AUTOMATION
The recommendation nobody can explain
06A decision gets made — a candidate, a supplier, a budget line — supported by an analysis everyone trusts and no one can reconstruct. When it is challenged, the room discovers that the reasoning does not live anywhere a person can reach.
ACCOUNTABILITY
Where it concentrates
06How exposure differs by kind of work
Exposure is not a ranking of which jobs are safe. It is a description of which tasks in a kind of work are most easily delegated, and which human capability has to stay awake as a result.
| Kind of work | Most exposed tasks | What has to stay yours |
|---|---|---|
| 01Administrative and coordination | Scheduling, formatting, note-taking, routine correspondence, data entry | Knowing which exceptions matter, and who needs to be told what |
| 02Writing, content and communications | First drafts, variations, summaries, headlines, repurposing | Judgment about what is true, what is yours to say, and what should not be said at all |
| 03Customer service and support | Tier-one answers, triage, tone matching, ticket summarising | The escalation nobody scripted, and the moment a person needs to be believed |
| 04Analysis, research and reporting | Literature scanning, data cleaning, chart generation, first-pass synthesis | Framing the question, choosing what counts as evidence, noticing what is missing |
| 05Software and technical work | Boilerplate, tests, refactors, documentation, familiar-shape problems | System design, debugging under pressure, knowing why the architecture is what it is |
| 06Teaching and training | Materials, differentiation drafts, quiz generation, feedback phrasing | Reading the room, judging readiness, deciding what this learner needs next |
| 07Design and creative production | Variations, mockups, asset generation, iteration on a set direction | Taste, the original brief, and knowing when the obvious answer is the wrong one |
| 08Legal, finance and compliance | Document review, clause comparison, reconciliation, first-pass flagging | Interpretation, professional responsibility, and the call that carries consequence |
| 09Healthcare and care work | Documentation, coding, scheduling, guideline lookup, drafting summaries | Clinical judgment, examination, and the presence a person needs from another person |
| 10Skilled trades and physical work | Estimating, scheduling, diagnostics support, documentation | Everything performed with the hands, and the judgment built by doing it |
This is a map of task exposure, not a forecast about people. Two people with the same job title can have very different exposure depending on what they actually do all day — and on how much of it they have handed over already.
Oversight
07Match oversight to consequence
The same tool behavior can be reasonable in one setting and unsafe in another. Verification effort should rise with consequence and irreversibility.
Low consequence · easy to reverse
Use AI freely for options, formatting and low-stakes experimentation.
brainstorm · outline · rewrite
Higher consequence · easy to reverse
Use AI to compare options, then verify claims and test before committing.
budget draft · work plan · public message
Low consequence · hard to reverse
Pause for consent, privacy and reputational effects even when the immediate stakes feel small.
personal data · image sharing · identity claims
High consequence · hard to reverse
Use qualified human review and authoritative evidence. Do not delegate final judgment.
safety · rights · health · legal
Examples illustrate a decision method. They are not professional medical, legal or financial guidance.
Human in the loop
08Keep a human in the loop: six decision gates
- 01
Intent gate
Is the goal legitimate, clear and aligned with the people affected?
- 02
Data gate
Can this information be shared? Is privacy, consent or confidentiality involved?
- 03
Evidence gate
Which claims, sources, calculations or assumptions require independent checking?
- 04
Consequence gate
What could happen if the output is wrong, biased, leaked or misunderstood?
- 05
Accountability gate
Who makes the final decision, signs off and can explain the reasoning?
- 06
Learning gate
What should the person still be able to do without the tool after this task?
Research pulse · R8
One study from this field
A review of 35 studies found that explainability can increase acceptance without reliably improving accuracy. Active engagement and independent verification are stronger safeguards.
Romeo & Conti (2025), AI & SOCIETY.
Modes
09Choose your mode by consequence
| Mode | Appropriate use | Human responsibility |
|---|---|---|
| USE | Explore and vary: ideas, outlines, formatting and reversible low-stakes work. | Set intent and select. |
| ASSIST | Support skilled work: drafting, comparison, coding, analysis and planning. | Understand, test and adapt. |
| VERIFY | Consequential claims: evidence, recommendations, calculations and public communication. | Check independently and document. |
| OWN | Final human judgment: rights, safety, health, legal status, consent and accountability. | Qualified person decides. |
Context beats blanket policy: the same AI system may move between modes inside one workflow.
Research pulse · R9
One study from this field
A 2025 review in medicine identified risks to judgment, examination and professional autonomy from AI-supported work, calling for longitudinal monitoring and explicit safeguards against skill erosion.
Natali et al. (2025), Artificial Intelligence Review.
How to read a research claim
- 01Who was studied? And are they enough like you for it to transfer?
- 02What was actually measured? Speed, quality, confidence and retention are different things, and studies often measure the easiest one.
- 03Compared to what? A result with no control, or a weak one, is a description rather than a finding.
- 04How long after? Almost everything works immediately. Ask what survived a month.
- 05Who else found it? One study is a hypothesis. A replication is evidence. A meta-analysis is a conversation.
Research on human skills, AI capabilities and future-readiness is expanding quickly, and some of what looks settled today will not be in a year. Reading two papers that disagree teaches you more than reading one that agrees with you.
Anti-deskilling protocol
10Keep the capability alive
- 01
Attempt first
Write a hypothesis, outline, calculation or plan before asking for assistance.
- 02
Request friction
Ask for critique, counterarguments, missing evidence and questions — not only an answer.
- 03
Test the output
Use another source, a known example, a calculator, a peer or a small experiment.
- 04
Explain it back
Reconstruct the logic and key facts from memory in your own words.
- 05
Make the call
Choose what to use, change or reject. State the reason.
- 06
Log the learning
Record one thing gained, one uncertainty and one capability to practise next.
Minimum standard: if you cannot explain why the result is credible and appropriate, you are not yet ready to rely on it.
Comparison
11Vulnerable pattern vs stronger pattern
| Context | Capability shrinks | Capability grows |
|---|---|---|
| 01Learning | Submit an AI summary you cannot explain. | Read first, create questions, compare the AI summary and explain the concept without it. |
| 02Work | Forward a polished recommendation without checking assumptions. | Identify decision criteria, verify key claims and record the accountable sign-off. |
| 03Personal life | Treat a confident answer as professional advice. | Use AI to prepare questions, then consult authoritative sources or a qualified person when stakes are high. |
Change the workflow, not only the attitude. Add a first attempt, a verification step and an accountable decision.
Personal audit
12Where is your capability exposed?
Complete the audit for one recurring AI-assisted task. The aim is to redesign the workflow, not to stop useful technology.
01I can complete the essential task without AI when necessary.
02I understand what information the system received.
03I know which claims or calculations require verification.
04I can explain the reasoning in my own words.
05I have considered bias, missing context and affected people.
06The level of oversight matches the consequence.
07The final accountable person is named.
08I notice signs of dependency or loss of confidence.
09I deliberately practise the underlying human skill.
10I review failures and update the workflow.
Loading your saved answers…
Questions
More questions →Closing
10The goal is not to avoid AI. It is to remain capable with it.
Keep enough skill to challenge the output, enough judgment to match oversight to consequence, and enough agency to own the final decision.
Volume 03
AI vulnerability
The printed A5 field guide — same content, typeset for paper.