1.5.2 - Workforce changes and automated decisions

How computers change work

Computers can automate tasks previously performed by people, augment people by supporting their work, and create new tasks such as maintaining systems. A warehouse may automate repetitive lifting, leaving workers to resolve unusual orders and maintain equipment. This can reduce injury exposure and increase consistent throughput, while displacing some existing roles. "Computers remove jobs" is too simple: particular tasks disappear, change or expand, and the gains and losses may reach different people.

An employer may benefit from lower cost, fewer repetitive errors or longer operating hours. Employees may gain safer work and useful tools, but face redundancy, retraining demands, deskilling or closer performance measurement. A customer may receive quicker service yet lose help from a person when a case does not fit the standard process. Remote collaboration can widen access to work, while dependence on devices, connectivity and digital skills can exclude some applicants.

Moral questions concern values about an individual's conduct, such as whether a manager should announce a planned replacement honestly. Ethical questions examine what ought to be done and whose interests should count, such as fair treatment during automation. Social effects concern wider relationships and distributions, including employment opportunities or unequal access. Cultural effects concern languages, practices and expectations of different communities. These aspects can overlap; explaining the affected person and causal link matters more than putting every effect into one rigid category. OCR clarification, printed p16

For a supermarket installing self-checkouts, fewer routine scanning duties may free staff to help customers, but only if the employer actually retains and reallocates those staff. Retraining offers a path to a new role; it is less convincing if no suitable roles or training time exist. Judge a proposal using the stated implementation, rather than assuming all promised benefits occur automatically.

Rules and automated decisions

Automated decision making uses a computer process to reach a decision, such as approving a routine claim or selecting a price. It does not always involve AI: a fixed rule can decide that an application meeting stated conditions proceeds to the next stage. The developer chooses the rule, input fields and consequences. Calling the output automatic does not remove responsibility for those design choices.

A rule-based process can apply the same criteria quickly across many cases, create an audit record and reduce variation between staff. However, consistently applying a mistaken rule consistently makes mistakes. A housing-repair service that rejects every request with a missing building code may exclude genuine urgent repairs when residents cannot find the code. Input validation can detect a missing field; it cannot by itself decide whether rejecting the request is a fair policy.

The importance of an error depends on its consequence. Suggesting an uninteresting song is different from denying access to employment or an essential service. Useful questions include whether information is complete and current, whether unusual circumstances can be considered, whether affected people understand the decision, and whether a mistake can be challenged before serious harm occurs.

Decision support recommends an outcome to a person who can meaningfully consider and change it. A staff member who merely presses "accept" without evidence, time or authority to disagree may not provide real oversight. Designing a system originally is also different from reviewing a particular person's decision when it is made.

AI learns patterns rather than certainty

Artificial intelligence (AI) describes systems that perform tasks associated with intelligence, such as recognising patterns, processing language or making predictions. In machine learning, a model is fitted using training examples, rather than a programmer explicitly writing every decision rule. A later input is processed using learned relationships. Not every AI system learns in the same way, and a trained model does not automatically understand a situation as a human would.

For example, a maintenance model could learn relationships between machine sensor readings and past failures. It can help engineers inspect likely problems earlier, potentially reducing outages. Its usefulness depends on suitable data and testing: a pattern that worked for one machine type may fail on a different type. A statistical association is not proof of a cause.

Generative AI produces new text, images or other content based on learned patterns. It can draft routine material or help a worker explore possibilities. Fluent output can still contain invented claims, outdated information or inappropriate recommendations. Verification is especially important when the output influences a consequential decision. A convincing explanation does not itself prove the underlying recommendation is correct.