Moral and ethical issues: computers in the workforce, automated decision making and artificial intelligence

Overview

Recall questions on arguments about computers destroying or creating jobs, job polarisation, automating a warehouse, automated decision making and its objections, the COMPAS and 2020 grading cases, and the benefits, risks and responsibility of artificial intelligence.

What you need to know

Answer each one before you reveal it.

Questions from this subtopic

Each one with the model answer underneath.

1. 1

   What is the case that computers destroy jobs, and what is the case that they do not?

   Model answer

   The destruction case: computers automate tasks that were previously paid work, and they do it faster and more cheaply, so demand for the workers who did those tasks falls. Frey and Osborne, in a widely cited 2013 study for the Oxford Martin School, estimated that around forty-seven per cent of United States employment was in occupations at high risk of computerisation over the following one to two decades. Automation also depresses wages and bargaining power in the occupations it does not eliminate. The counter-case: technology has repeatedly created more work than it has destroyed, because it lowers prices, raises demand and creates categories of job that did not exist. An OECD study by Arntz, Gregory and Zierahn in 2016 re-ran the analysis by tasks rather than whole occupations and put the share of jobs at high risk across OECD countries at around nine per cent, arguing that most occupations contain tasks that resist automation. James Bessen's work on cash machines is often cited: the number of bank tellers employed in the United States rose after cash machines spread, because branches became cheaper to run so banks opened more of them and tellers moved to sales and advice work. Both sides are working from models, not observations of the future, and the estimates differ chiefly because of the unit of analysis.

2. 2

   Even where the total number of jobs holds up, economists argue the shape of employment changes. Explain what is meant by job polarisation.

   Model answer

   Polarisation, sometimes called hollowing out, is the observation that employment has grown at the top and the bottom of the wage distribution while shrinking in the middle. The explanation offered by researchers including Goos and Manning for the United Kingdom and Autor for the United States is that computers substitute most readily for routine tasks, meaning tasks that can be fully specified as a set of rules, and routine tasks were concentrated in middle-paying occupations such as clerical work, bookkeeping and assembly-line production. High-paid work often involves non-routine analytical and managerial tasks, where computers complement rather than replace the worker. Low-paid service work such as cleaning, care and food preparation involves non-routine manual tasks that are surprisingly hard to automate, and is also hard to offshore. The consequence discussed in the literature is not mass unemployment but a widening gap, with fewer of the middle-income routes that once let people without degrees earn a stable living. Machine learning complicates the picture, because it can handle some non-routine tasks that could never be written as explicit rules, which is why forecasts made before 2015 may understate the exposure of professional work.

3. 3

   Set out the arguments for and against an employer replacing a team of warehouse pickers with an automated storage and retrieval system.

   Model answer

   For: throughput rises and running costs fall, so the goods reach customers cheaper; machines work continuously without breaks, sick leave or night-shift premiums; error rates and stock discrepancies typically fall; and the picking job itself is physically punishing, involving repetitive lifting and long distances walked, so removing it removes a source of musculoskeletal injury. New roles appear in maintenance, systems support and exception handling. Against: the displaced workers are unlikely to be the ones who get the new roles, because the new roles need different qualifications, and warehouses are often concentrated in areas with few alternative employers, so the local cost is borne by people who capture none of the gain. Remaining workers may find the pace set by the machine, with algorithmic targets and monitoring reducing autonomy. There is a capital cost and a new single point of failure. There is also a distributional question that is not technical at all: the productivity gain has to go somewhere, and whether it goes to lower prices, higher pay or higher profit is decided by market power and policy, not by the technology. Reasonable people disagree about whether the answer is to slow automation, to retrain, or to change how the gains are shared.

4. 4

   Why do organisations use automated decision making, and what can it do better than a person?

   Model answer

   Speed and scale: a lender can assess millions of applications in seconds, which no human panel could do at a price customers would pay. Consistency: the same inputs always produce the same output, whereas human decisions vary with fatigue, the time of day and who is deciding, and studies of human decision makers repeatedly find that inconsistency. Cost, which can widen access, since cheap assessment makes small loans or small insurance policies viable. Ability to weigh many variables at once and to detect statistical patterns a person would not notice, which is why fraud detection is largely automated. Auditability in principle: a rule-based system can be inspected and its reasoning reproduced exactly, which is harder with a person's judgement. And removal of some human biases: a system does not know it dislikes the applicant's accent. Those are real advantages and any balanced answer has to state them before turning to the objections.

5. 5

   What are the main objections to letting an algorithm decide something important about a person?

   Model answer

   Bias inherited from data: a model trained on past decisions learns the pattern of those decisions, including any discrimination in them, and can reproduce it at scale while appearing neutral, because a proxy variable such as postcode can stand in for a protected characteristic. Opacity: with complex models neither the subject nor often the operator can say which factors drove a particular decision, so it cannot be meaningfully challenged. Absence of an exception route: rules cannot anticipate every case, and a person can recognise that a case is unusual where a classifier simply assigns it to the nearest category. Automation bias: humans nominally in the loop tend to defer to the machine's output rather than genuinely reviewing it. Scale of error: a wrong human decision harms one person, a wrong rule harms everyone it is applied to, simultaneously. Diffusion of accountability between the developer, the purchaser and the operator. And feedback loops: predictive systems that direct attention to a place or group generate more recorded incidents there, which the model then reads as confirmation. UK data protection law responds to some of this by generally prohibiting solely automated decisions with legal or similarly significant effects unless a specific ground applies, with a right to obtain human intervention.

6. 6

   The COMPAS risk-assessment tool used in parts of the United States criminal justice system is the standard case study in algorithmic fairness. What was the disagreement about?

   Model answer

   In 2016 the news organisation ProPublica analysed risk scores produced by the COMPAS tool and reported that among defendants who did not go on to reoffend, black defendants were roughly twice as likely as white defendants to have been labelled high risk, and that among those who did reoffend, white defendants were more likely to have been labelled low risk. It concluded the tool was biased. The tool's developer, then Northpointe, replied that the tool was calibrated: within any given risk score, the proportion who actually reoffended was about the same for both groups, which is the fairness criterion a user of the score most naturally cares about. Subsequent academic work by Kleinberg, Mullainathan and Raghavan and separately by Chouldechova showed that both parties were right on their own terms, and that when the underlying rate of the outcome differs between groups it is mathematically impossible for a classifier to be calibrated and also to equalise false positive and false negative rates, except in degenerate cases. The lesson usually drawn is that fairness is not one property to be engineered in, but a set of incompatible properties among which someone has to choose, and that the choice is a political and ethical one rather than a technical one. The case is also contested on wider grounds, including whether risk prediction should feature in sentencing at all.

Other ways to work this

The same content, four other ways.

What is inside Levo

Six ways to work the same material.