Week 18: AI in Employment and Society

 

Welcome to Week 18 of the AI Literacy Course.

Imagine that an organisation wants to introduce an AI assistant for handling customer cases.

The supplier promises faster work and lower costs. However, employees will need to check the answers, correct mistakes, explain decisions, manage complaints, and remain responsible when something goes wrong.

Has the AI system saved work—or moved work and responsibility to different people?

This lesson examines how AI can change tasks, working conditions, power, skills, and the distribution of benefits.

The central rule is:

Define the problem, map the work, measure benefits and harms, involve affected people, and keep the power to pause or stop.

Learning goals

By the end of this lesson, you should be able to:

• analyse tasks instead of making simple predictions about whole occupations;

• distinguish automation, augmentation, and hidden work;

• assess productivity and job quality together;

• recognise algorithmic management and high-impact employment uses;

• identify data, accessibility, and worker-rights concerns;

• design a bounded pilot with measurable stop conditions;

• examine who receives the benefits and who bears the costs.

1. Begin with the problem

A workplace should not begin with:

“Where can we use AI?”

It should begin with:

“What problem are we trying to solve?”

The problem might be:

• repetitive work causing physical or mental strain;

• long waiting times;

• difficulty finding information;

• inconsistent decisions;

• increasing errors;

• an inaccessible service;

• poor coordination;

• insufficient staffing.

Then ask:

1. Is this the correct description of the problem?

2. What evidence supports it?

3. Could the process be improved without AI?

4. Would additional staff, training, clearer rules, or simpler software work better?

5. Who experiences the problem?

6. Who should help define the solution?

A poorly designed process does not become fair or effective merely because AI is added.

2. Start by analysing tasks

Jobs contain combinations of:

• routine tasks;

• complex decisions;

• communication;

• relationships;

• physical actions;

• emotional labour;

• exceptions;

• professional judgment;

• responsibility.

AI may change some parts while others remain human.

For example, an administrative employee might use AI to draft routine text. The employee may still need to:

• understand the case;

• check source material;

• identify exceptions;

• correct errors;

• communicate with the affected person;

• explain the decision;

• remain accountable.

Exposure to AI does not mean certain job loss. It means that AI could affect some part of the work.

Possible outcomes include:

• tasks becoming faster;

• tasks being automated;

• new checking work;

• jobs being reorganised;

• new roles being created;

• fewer workers being required for some tasks;

• increased demand for other services;

• displacement or changed entry routes.

The outcome depends on technology, workplace organisation, demand, investment, policy, and decisions about staffing.

3. Map the work before changing it

For each task, ask:

• What is its purpose?

• Who performs it now?

• What information does it require?

• What professional knowledge is needed?

• Which exceptions occur?

• Which relationships or conversations matter?

• What happens if the task fails?

• Who is affected by an error?

• Who is accountable?

• Could the task be improved without AI?

This map should include work that may not appear in an official process description.

Employees often perform informal but essential work, such as:

• noticing unusual cases;

• helping colleagues;

• calming worried people;

• adapting communication;

• repairing earlier mistakes;

• coordinating between departments;

• protecting safety;

• explaining rules in accessible language.

An AI system may not recognise these responsibilities.

4. Automation, augmentation, and hidden work

Automation

Automation transfers part of a task to a system.

Example:

An AI system automatically sorts incoming messages into categories.

Augmentation

Augmentation gives people a tool intended to support their work.

Example:

An AI assistant suggests a draft response that an employee reviews and edits.

Hidden work

AI can create additional work that is missing from productivity estimates.

Examples include:

• checking outputs;

• correcting data;

• handling exceptions;

• writing and testing prompts;

• documenting decisions;

• answering complaints;

• reporting incidents;

• training colleagues;

• monitoring system updates;

• repairing harm caused by errors.

Automation and augmentation can occur in the same workflow. A system may automate the first step while creating more human work later.

Ask:

• Who gains time?

• Who receives new responsibilities?

• Who performs the checking?

• Is checking time included in the productivity calculation?

• Who carries the consequences of an error?

• Is the extra work recognised and paid?

5. Productivity is not the same as job quality

Productivity may mean:

• more cases completed;

• less time per case;

• lower cost;

• shorter waiting times;

• fewer errors;

• improved service quality.

These measures may conflict.

A system might reduce handling time while increasing:

• serious errors;

• complaints;

• stress;

• checking time;

• surveillance;

• unequal treatment.

Ask:

Productivity for whom, measured how, and at what cost?

Job quality also includes:

• autonomy;

• meaningful work;

• manageable workload;

• pay;

• health and safety;

• privacy;

• fairness;

• job security;

• learning and development;

• supportive relationships;

• ability to use professional judgment.

A responsible assessment measures productivity and job quality together.

6. Algorithmic management changes workplace power

Algorithmic management means using data-driven systems to direct, monitor, evaluate, or organise work.

These systems may:

• allocate tasks or shifts;

• set the pace of work;

• rank performance;

• recommend training or promotion;

• calculate bonuses;

• monitor communication;

• predict absence or resignation;

• recommend warnings or disciplinary action.

Algorithmic management does not always use generative AI. It may use rules, scores, optimisation software, or predictive models.

The system can influence:

• what managers see;

• which behaviour is rewarded;

• how workers are compared;

• which information is ignored;

• whether workers can challenge a result.

A score may appear objective even when it is based on incomplete data or questionable assumptions.

7. Human approval must be meaningful

A person at the end of an automated process does not automatically provide effective human review.

Meaningful review requires:

• access to relevant evidence;

• knowledge of the system’s limitations;

• sufficient time;

• appropriate training;

• authority to disagree;

• the ability to correct data;

• responsibility for the final decision.

If a manager approves every recommendation because of workload or pressure, human review has become a rubber stamp.

Workers and affected people need a clear route to:

• receive information;

• identify incorrect data;

• explain relevant circumstances;

• question a result;

• request correction;

• reach a real person.

8. Recruitment and employment decisions require strong safeguards

AI may be used for:

• targeted job advertisements;

• application screening;

• candidate ranking;

• video interviews;

• skills assessment;

• promotion;

• training decisions;

• performance evaluation;

• dismissal recommendations.

These uses can affect people’s opportunities and livelihoods.

Possible problems include:

• historical bias in training data;

• irrelevant proxy variables;

• inaccessible assessment formats;

• false inferences;

• unequal error rates;

• unclear scoring;

• limited opportunities to challenge a result.

Removing names does not remove every proxy. Postcode, school, language pattern, age, disability, employment gaps, or device use may still reproduce group differences.

Ask:

• Is the feature relevant to the job?

• Has the assessment been shown to measure what it claims to measure?

• Has accessibility been tested?

• Are errors distributed unequally?

• Can applicants request an alternative method?

• Can a person correct inaccurate information?

• Is the final decision meaningfully reviewed?

Vendor claims and demonstrations are not enough. Evidence must match the actual intended use.

9. Accessibility needs an alternative route

An AI-supported process may disadvantage people through:

• timed video interviews;

• speech or facial analysis;

• poor compatibility with assistive technology;

• weak recognition of certain accents or dialects;

• assumptions about eye contact or communication style;

• inaccessible instructions;

• limited language support;

• no way to request accommodation.

A process should not force everyone to interact with AI in exactly the same way.

Workers and applicants may need:

• accessible formats;

• additional time;

• human assistance;

• another assessment method;

• a non-AI route;

• a way to request reasonable accommodation.

Accessibility should be designed before deployment, not added only after someone is excluded.

10. Emotion recognition is especially sensitive

Some systems claim to infer emotion, honesty, attention, confidence, or personality from facial expressions, voice, movement, or behaviour.

Such claims may be unreliable and can produce unfair or intrusive judgments. Behaviour also varies across individuals, cultures, disabilities, languages, and situations.

Certain workplace emotion-recognition uses are prohibited under the EU AI Act, apart from limited medical or safety contexts.

Do not assume that a camera or microphone can reliably reveal what a person feels, intends, or believes.

11. Workplace data and surveillance need limits

Workplace systems may process:

• messages;

• voice and video;

• location;

• device activity;

• keystrokes;

• work speed;

• schedules;

• health information;

• biometric data;

• behavioural inferences.

Collecting more data does not automatically improve a system.

Monitoring can:

• change behaviour;

• reduce trust;

• increase stress;

• create inaccurate profiles;

• blur the boundary between work and private life;

• give managers an incomplete picture of performance.

Apply Lesson 15’s data-protection principles:

• define the purpose;

• use approved systems;

• minimise personal data;

• limit access and retention;

• provide transparency;

• protect security;

• create correction and incident routes.

Workers should not be expected to surrender unlimited data simply because monitoring is technically possible.

12. Skills development requires time and access

Workers may need:

• domain knowledge;

• AI literacy;

• source criticism;

• privacy awareness;

• bias awareness;

• task and workflow design;

• output verification;

• incident reporting;

• the ability to explain decisions;

• the ability to work safely without the tool.

Telling workers to “upskill” in their private time transfers the cost to people with the least time and resources.

A responsible skills plan includes:

• paid learning time;

• accessible training;

• relevant practice;

• necessary equipment;

• access to approved tools;

• support when something fails;

• recognised learning;

• repeated training after major changes.

The EU AI Act requires providers and deployers to take measures to ensure an appropriate level of AI literacy among people working with AI systems on their behalf. Exact training needs depend on the person’s knowledge, role, and context.

13. Maintain skills and operational resilience

If people stop practising essential skills, the organisation may become dependent on the AI system.

This creates risk during:

• system outages;

• cyber incidents;

• incorrect outputs;

• model updates;

• vendor failure;

• emergencies.

Ask:

• Which human skills must remain available?

• How will workers practise them?

• Can the organisation continue safely without the system?

• Is there a manual fallback?

• Who decides when to switch to it?

The goal is not to reject useful technology. It is to prevent avoidable dependence and loss of critical knowledge.

14. Worker participation improves decisions

People who perform the work understand:

• exceptions;

• informal coordination;

• safety risks;

• emotional labour;

• accessibility needs;

• hidden checking work;

• actual failure consequences.

Workers and their representatives should be involved early enough to influence:

• the problem definition;

• whether AI is needed;

• tool selection;

• data limits;

• pilot measures;

• training;

• complaint routes;

• stop conditions.

Follow applicable consultation duties, national law, workplace arrangements, and collective agreements.

Participation after purchase can still identify problems, but important choices may already be difficult or expensive to change.

Consultation is meaningful only when someone has the authority and responsibility to respond.

15. Use WORK to evaluate a proposal

W — Work and purpose

Identify:

• the problem;

• the task;

• the current workflow;

• decisions and relationships;

• exceptions;

• data;

• failure consequences;

• non-AI alternatives.

O — Outcomes and rights

Measure:

• accuracy;

• time;

• service quality;

• workload;

• autonomy;

• privacy;

• accessibility;

• fairness;

• safety;

• trust;

• ability to challenge outcomes.

R — Rebalance gains and responsibilities

Ask:

• Who saves time?

• Who performs new checking work?

• Who receives the financial benefit?

• Who bears errors and transition costs?

• Who receives paid training?

• Who remains accountable?

K — Keep worker voice and test safely

Involve affected workers and representatives before major choices are fixed.

Use:

• a bounded pilot;

• clear roles;

• limited data;

• a comparison baseline;

• complaint and incident routes;

• measurable stop conditions;

• a rollback plan.

16. Design a bounded pilot

A pilot should test a clear claim.

Example:

“For eight weeks, the AI assistant will help five trained employees draft responses to routine, low-risk enquiries. Employees will approve every response before it is sent.”

A responsible pilot defines:

• purpose;

• duration;

• participants;

• approved tasks;

• excluded tasks;

• data limits;

• training;

• human responsibility;

• comparison baseline;

• success measures;

• complaint route;

• incident process;

• stop conditions;

• rollback plan.

Compare the AI-supported workflow with:

• the current workflow;

• a simpler process improvement;

• another non-AI alternative where relevant.

Do not assume that the existing process is good merely because it is the baseline.

17. Measure benefits and harms

Useful measures may include:

• time per task;

• total checking time;

• accuracy;

• serious-error rate;

• correction rate;

• workload;

• stress;

• autonomy;

• accessibility;

• privacy incidents;

• unequal error distribution;

• service quality;

• complaints;

• worker and service-user trust;

• ability to complete the work during an outage.

A pilot should not be judged by productivity alone.

A small time saving does not justify serious safety, privacy, discrimination, or accessibility problems.

18. Define stop conditions and rollback

Stop conditions should be agreed before the pilot begins.

Possible pause or stop conditions include:

• serious privacy or security incident;

• unacceptable error rate;

• a dangerous or high-impact error;

• substantially higher error for one group;

• inaccessible process with no alternative;

• workload increasing beyond an agreed limit;

• workers being unable to correct outputs;

• repeated complaints;

• use outside the approved purpose;

• failure of the manual backup process.

The pilot also needs a rollback plan.

The plan should explain:

• who can pause the system;

• how workers will be informed;

• how unfinished cases will be handled;

• how the previous workflow will resume;

• what evidence will be preserved;

• how affected people will receive support;

• what must change before another test.

The power to stop must be practical, not merely written in a policy.

19. Benefits and transition costs should be visible

Productivity gains may support:

• less repetitive work;

• safer tasks;

• improved services;

• shorter waiting times;

• higher wages;

• reduced working time;

• learning and development.

However, workers may experience:

• displacement;

• increased monitoring;

• more intense work;

• unpaid checking;

• skill loss;

• job insecurity.

AI systems also depend on labour that may be less visible, including:

• data labelling;

• content moderation;

• evaluation;

• maintenance;

• customer support;

• outsourced platform work.

Ask:

Who receives the gains, and who pays for the transition?

The answer is shaped by organisational decisions, bargaining, education, social protection, procurement, taxation, competition, and public policy—not by technology alone.

Classroom activity: Responsible workplace AI pilot

Use a fictional scenario provided by your teacher. Do not use real workplace personal data, confidential documents, or performance information.

Complete this pilot canvas:

Area	Questions
Problem	What needs improvement? Is AI necessary?
Task	Which steps, decisions, exceptions, and relationships are involved?
Data	What information enters, leaves, or is inferred?
Human role	Who checks, decides, explains, and remains accountable?
Benefits	What measurable improvement is expected?
Job quality	What happens to workload, autonomy, safety, skills, and privacy?
Inclusion	Who may face barriers or unequal errors?
Participation	Which workers or representatives should be involved?
Measures	What baseline and indicators will be used?
Stop conditions	What result would pause or end the pilot?
Rollback	How will work continue safely without the system?
Decision	Proceed, revise, choose another method, or reject?

Present your decision as a short written analysis, audio response, or presentation.

AI may help brainstorm alternatives using fictional information. The final assessment and decision must be your own.

Reflection questions

1. What problem is the organisation trying to solve?

2. Could the problem be addressed without AI?

3. Which tasks would be automated, supported, or newly created?

4. What hidden checking or repair work might appear?

5. How will productivity and job quality be measured together?

6. Who may experience privacy, accessibility, or unequal-treatment risks?

7. Can workers and applicants understand and challenge important outcomes?

8. What learning and support will occur during paid time?

9. What measurable result would stop the pilot?

10. How would work continue after the system is paused?

Key vocabulary

Automation:
Transferring part of a task to a technological system.

Augmentation:
Using technology to support people in performing work differently.

Hidden work:
Additional checking, correction, coordination, training, or repair work created by a new system.

Algorithmic management:
Using data-driven systems to allocate, monitor, evaluate, or direct work.

Automation bias:
Trusting a system’s output because it appears objective or precise.

Job quality:
The conditions shaping autonomy, workload, safety, meaning, privacy, development, security, relationships, and fairness at work.

Contestability:
The ability to understand, question, challenge, and correct an outcome.

Stop condition:
A pre-agreed result or incident that requires a pilot to pause or end.

Rollback plan:
A plan for returning safely to an earlier or alternative workflow.

WORK:
Work and purpose, Outcomes and rights, Rebalance gains and responsibilities, and Keep worker voice and test safely.

Summary

In Week 18, we learned that AI’s effects on employment begin with changes to particular tasks, workflows, relationships, and responsibilities.

AI may automate some work, support other work, and create new hidden work. Exposure does not guarantee job loss, but transition costs and displacement risks must not be ignored.

Responsible workplace AI requires more than productivity measures. It also requires attention to:

• workload;

• autonomy;

• privacy;

• accessibility;

• fairness;

• skills;

• safety;

• service quality;

• worker participation;

• the distribution of gains.

Use WORK:

• Work and purpose;

• Outcomes and rights;

• Rebalance gains and responsibilities;

• Keep worker voice and test safely.

A responsible pilot has a baseline, several measures, meaningful human responsibility, complaint and incident routes, measurable stop conditions, and a rollback plan.

Define the problem, map the work, measure benefits and harms, involve affected people, and keep the power to pause or stop.

Official guidance

• European Commission: EU AI Act

• European Commission: Employment systems under the AI Act

• European Commission: AI literacy

• OECD: AI and work

• OECD: Algorithmic management in the workplace

Lesson 18 Interactive Quiz: AI in Employment and Society

Choose one answer for each question. Then select Check my answers. This practice quiz does not collect names or scores.

1. A department has long waiting times and immediately proposes buying an AI assistant. What should it do first?

A. Purchase the tool before another organisation does B. Ask the vendor to define the department’s problem C. Investigate the cause of the delays and compare AI with simpler alternatives D. Measure success only by how many AI features are activated

2. AI drafts customer responses, but employees must correct errors and handle more complaints. What does this show?

A. Automation can create hidden checking and repair work B. The task has been completely removed from human work C. Correction time should be excluded from productivity measures D. Customer complaints prove that employees need less authority

3. A pilot reduces handling time by 15%, but employees report greater stress and more serious errors. What is the best assessment?

A. The pilot succeeded because speed is the only relevant measure B. Worker experience should be ignored because it is not numerical C. Serious errors are acceptable whenever average time falls D. Productivity gains must be assessed with error severity and job quality

4. Software assigns shifts and ranks workers using performance and availability data. What is this an example of?

A. Generative writing assistance only B. Algorithmic management C. Data anonymisation D. Worker participation

5. A recruitment process requires every applicant to complete a timed video interview analysed for confidence and emotion. What is the strongest concern?

A. Video interviews are always more accurate than human interviews B. All applicants behave identically when facing a camera C. The process may be inaccessible, intrusive, unreliable, and require an alternative route D. Emotion scores remove the need to assess job-related skills

6. A manager approves every AI-generated performance warning because there is no time to examine the evidence. What is missing?

A. Meaningful human review B. Faster automatic approval C. More precise-looking scores D. Removal of all complaint routes

7. When should affected workers and their representatives be involved in an AI proposal?

A. Only after the system has made a harmful decision B. After purchase, when the main decisions can no longer change C. Only when the supplier requests feedback D. Early enough to influence the problem, tool, data limits, pilot, and stop conditions

8. Which proposal is the strongest bounded pilot?

A. Deploy the system across the organisation with no end date B. Test one low-risk task for eight weeks with trained users, data limits, a baseline, and several measures C. Let each worker use the tool differently and compare no results D. Measure only whether participants say they like AI

9. A pilot causes a serious privacy incident. Which response shows that its stop and rollback plans are effective?

A. Continue until the planned end date so the results remain complete B. Ask workers to use the system more carefully without changing anything C. Pause the system, follow the incident process, preserve evidence, and resume the safe alternative workflow D. Delete visible records and avoid reporting the incident

10. Which question set best applies the WORK framework?

A. What work changes, which outcomes and rights matter, who gains or bears responsibility, and how can workers test and stop the system safely? B. Which tool has the most features, and how quickly can it replace the current process? C. How can all benefits be measured as speed while excluding worker experience? D. How can the organisation avoid documenting risks and stop conditions?

Reflection: A responsible workplace pilot measures benefits and harms, includes affected workers, and preserves a practical way to pause or stop.