AI at work and in society

What you'll learn

• Why tasks, not whole jobs, are the right unit for thinking about automation
• The difference between automation and augmentation, and why it matters to you
• Why the jobs most exposed are not the ones people assume
• What algorithmic management is, and what it changes about working life
• Why accountability and recourse are the questions to ask about automated decisions
• The real resource cost of AI systems, and why it is easy to overlook
• What market concentration means when a few providers supply everyone
• Why cheap, plausible content changes the problem of trust rather than just the volume of lies
• What the digital divide looks like when access is unequal
• Which skills are likely to hold their value

Key terms and definitions

Core concepts

Jobs are bundles of tasks

The usual question — "will AI take my job?" — is badly framed, and the better framing is genuinely more useful.

A job is a bundle of tasks. A nurse takes observations, explains a diagnosis, writes notes, reassures a frightened patient, spots that something is wrong before any instrument says so. A solicitor reads documents, drafts standard clauses, advises a client on a decision, argues in front of a judge.

AI systems are good at some tasks in almost every bundle and poor at others. So the common outcome is not a job disappearing but a job changing shape: the share of time spent on each task shifts, and what the role is mainly for changes with it.

This matters practically. A job where one task was ninety per cent of the work is genuinely exposed. A job made of many different tasks, several needing physical presence or responsibility for a decision, changes without vanishing.

Automation and augmentation

Two different things happen when a system takes on a task.

• Automation — the system does the task instead of a person. Sorting documents, transcribing recordings, drafting standard text.
• Augmentation — the system helps a person do the task better. A radiologist reviewing scans with a system that flags areas worth attention; a translator editing a machine draft.

Which one you get is not a property of the technology. It is a choice made by whoever introduces it, and the same tool can be deployed either way. A system that drafts a report can replace the writer or give the writer a faster first draft to improve.

Augmentation tends to produce better results where the stakes are high, because the person stays responsible for the judgement. It also tends to be more demanding to do well, because the person has to remain skilled enough to catch what the system gets wrong — which is exactly what deskilling erodes.

Which work is actually exposed

The expectation for most of the last century was that routine physical work would automate first. What has happened with AI is partly the reverse.

Tasks currently most affected tend to be:

• Producing routine text — standard correspondence, summaries, first drafts
• Classifying and sorting information at volume
• Early-career work that consists largely of the above

Tasks currently least affected tend to involve:

• Physical work in unpredictable settings — care, trades, maintenance
• Responsibility for a decision somebody must answer for
• Relationships and trust built over time
• Work where being wrong is expensive and verification is hard

There is a particular problem in that list. Much early-career work exists partly to train people: a junior does routine drafting and, in doing it, learns judgement. If the routine part is automated away, the training route can disappear with it — and the senior roles still require the judgement that route produced. That is a real structural worry, and it is more interesting than "jobs will go".

Algorithmic management

Less discussed than automation, and more immediate for many workers: software increasingly allocates, monitors and assesses work. Scheduling shifts, routing deliveries, ranking performance, flagging who is falling behind.

Three things change when it does.

• Monitoring becomes continuous, where before it was occasional.
• Decisions become opaque. Being told a system assigned you fewer hours is harder to question than being told by a manager who can explain why.
• Responsibility becomes diffuse. "The system decided" invites nobody to defend the decision.

None of this is inevitable, and that is the point. Each is a design and management choice, which means each can be made differently.

Accountability and recourse

When an automated system contributes to a decision about a person — a job application screened out, a benefit refused, a place allocated — two questions matter more than how the system works.

Who is accountable? A system cannot be responsible for anything. Somebody chose to deploy it, somebody set the threshold, somebody decided what happens to those it rejects. Accountability sits with people, and a decision nobody will answer for is the warning sign.

What recourse exists? Can the person affected find out a system was involved, learn roughly on what basis, and have the decision reconsidered by a human who can actually change it? A review that cannot overturn the outcome is not recourse.

A useful principle: the higher the stakes and the harder the decision is to reverse, the stronger the case for a person deciding, rather than a person rubber-stamping.

The resource cost

AI systems are often described as though they were weightless. They are not. Training large models and then answering millions of requests consumes real electricity, requires water for cooling, and depends on manufactured hardware with its own supply chain.

Two honest qualifications. The cost per individual use is small, and the comparison that matters is against the alternative, which also has a cost — a journey not made, a document not printed. But aggregate demand is large enough that it affects where data centres are built, how much power they draw and what the local effects are, which is why it is now part of planning and policy arguments rather than a footnote.

The reasonable position is neither "this is destroying the planet" nor "it is just software". It is a real industrial activity with real inputs, worth counting rather than assuming away.

Concentration of a few providers

Building a frontier model requires computing resources, data and expertise available to very few organisations. The consequence is market concentration: a small number of providers supply the systems that a great many schools, hospitals, firms and governments come to rely on.

That matters for reasons that have nothing to do with the quality of the models:

• A provider's decision about what its systems will and will not do becomes a decision affecting everyone
• Prices and terms can change for users with nowhere else to go
• An outage or a withdrawn product disrupts organisations that built around it
• The values built into widely used systems get applied far beyond where they were debated

Dependence on infrastructure is not new — electricity and payment networks are concentrated too. What is newer is dependence on systems that make judgements, where the provider's choices shape outcomes rather than just delivering a service.

Trust, not just misinformation

Convincing text, images, audio and video are now cheap to produce. The obvious consequence is more false material. The deeper consequence is different.

When anything could be fabricated, genuine evidence loses some of its force. A real recording can be dismissed as a fake; a real photograph proves less than it did. This is a shift in the status of evidence, and it is harder to address than any individual falsehood.

What holds up under that pressure is not detection technology, which struggles. It is provenance — knowing where something came from — and the habit of trusting sources with something to lose by being wrong rather than trusting material because it looks convincing.

Unequal access

The digital divide has three layers, and only the first is usually noticed.

• Access — devices, reliable electricity, affordable data
• Capability — knowing how to use these tools well enough to benefit
• Quality — free tiers and paid tiers differ, so even users of the "same" tool are not equally served

Two opposite effects run at once, which is why confident predictions here are unwise. AI can narrow gaps: patient explanation at any hour is something a well-resourced student always had access to through tutoring. It can also widen them, because the people best placed to benefit are those who already have the equipment, the connection and the background knowledge to judge the output.

What is likely to hold its value

Not a prediction about particular jobs, which nobody can make reliably — but about what tends to stay scarce.

• Judgement under uncertainty: deciding what should be done when the answer is not lookupable
• Verification: knowing enough to tell a good answer from a plausible one
• Communication with people who need to be understood, persuaded or reassured
• Physical skill in unpredictable settings
• Responsibility: being someone who will answer for a decision
• Learning quickly, because the tools will keep changing

The common thread is that each requires understanding a field well enough to supervise work in it. Which is an argument for learning the fundamentals of something properly, not against it.

Worked examples

Example 1: Task-level analysis (4 marks)

Explain why "will AI replace teachers?" is a poorly framed question.

• A job is a bundle of tasks, and systems are strong at some and weak at others (1 mark)
• Marking routine work or drafting materials is more exposed than managing a class (1 mark)
• Tasks involving relationships, responsibility and physical presence are far less exposed (1 mark)
• So the likely outcome is the role changing shape rather than disappearing (1 mark)

Example 2: Automated decisions (4 marks)

An employer screens applications automatically and rejects some without human review. Identify the main concerns.

• Accountability: a system cannot be responsible, so it must be clear who chose to deploy it and set the threshold (1 mark)
• Recourse: applicants should be able to learn a system was used and have the decision reconsidered (1 mark)
• A review that cannot overturn the outcome is not genuine recourse (1 mark)
• The higher the stakes and the harder the decision is to reverse, the stronger the case for a person deciding (1 mark)

Example 3: Evaluating a claim (3 marks)

"AI will close the gap between rich and poor schools." Evaluate this.

• It may narrow gaps, since patient explanation on demand was previously available mainly through tutoring (1 mark)
• It may widen them, since benefit depends on devices, connection and the background knowledge to judge output (1 mark)
• Free and paid tiers of the same tool also differ, so equal access to a product is not equal benefit (1 mark)

Common mistakes and how to avoid them

• Asking whether a job will disappear. Ask which tasks in it are exposed.
• Treating automation as the only option. Augmentation is a choice, made by whoever deploys the system.
• Assuming manual work is most at risk. Routine text and classification are more exposed than care or trades.
• Ignoring the loss of training routes. Automating junior work can remove how judgement was learned.
• Overlooking algorithmic management. Monitoring and allocation affect more workers than replacement does.
• Asking how a system works instead of who answers for it. Accountability and recourse are the practical questions.
• Treating AI as weightless. It uses electricity, water and hardware, and aggregate demand is large.
• Ignoring concentration. Few providers supplying everyone is itself a risk.
• Thinking the problem is only more lies. Genuine evidence also loses force.
• Assuming access equals benefit. Capability and tier quality matter too.

Using this in practice

When you meet a claim about AI changing work or society:

1. Which tasks, specifically, are being talked about?
2. Automation or augmentation — and who made that choice?
3. Who is accountable if the outcome is wrong?
4. What recourse does an affected person have?
5. How reversible is the decision, and how high are the stakes?
6. What does it cost in energy, hardware and dependence on a provider?
7. Who benefits, and who is left out at each layer of access?
8. What is the human contribution that remains, and am I building it?

Quick revision summary

• Think in tasks, not jobs: roles change shape more often than they vanish
• Automation replaces a person at a task; augmentation helps them do it — which happens is a choice
• Routine text and classification are more exposed than care, trades or decision-making responsibility
• Automating early-career work can remove the route by which judgement was learned
• Algorithmic management makes monitoring continuous, decisions opaque and responsibility diffuse
• For automated decisions, ask who is accountable and what recourse exists — stakes and reversibility set how much human control is needed
• AI has a real resource cost in electricity, water and hardware, significant in aggregate
• Market concentration means a few providers' choices affect everyone who depends on them
• Cheap convincing content weakens the force of genuine evidence, so provenance matters more than detection
• The digital divide has layers of access, capability and tier quality, and AI can narrow or widen gaps
• What stays scarce: judgement, verification, communication, physical skill, responsibility and learning quickly