Artificial Intelligence and the Impact of Computing
What machine learning actually is, why AI systems become biased, and how to think clearly about technology that is changing faster than the rules around it.
What AI actually is
Artificial intelligence means computer systems performing tasks that would normally require human intelligence, such as recognising objects, understanding language or making decisions.
Machine learning
Most modern AI uses machine learning. Instead of a programmer writing rules, the system is shown a very large number of examples and finds patterns in them itself.
To build a system that recognises cats:
- Collect hundreds of thousands of images, each labelled cat or not cat, which is the training data
- The system adjusts its internal values, trying to get more labels right
- It is then tested on images it has never seen, to check it has learned something general rather than memorising
- It can then classify new images
The system has no idea what a cat is. It has found statistical patterns that correlate with the label cat. That is why it can be confidently and bizarrely wrong.
Where you already meet it
- Recommendations on streaming services and social media
- Voice assistants and speech to text
- Face unlock on a phone
- Spam filters
- Sat nav route prediction
- Translation
- Text and image generation
Step 1 of 5 A normal program follows rules a human wrote down.
Bias, error and responsibility
Why AI systems become biased
An AI system learns from data, and data comes from the world, which contains bias. If a hiring system is trained on twenty years of a company's decisions, it learns the patterns in those decisions, including any discrimination in them.
Real examples include facial recognition working significantly less accurately on darker skin because the training data contained mostly light skinned faces, and recruitment tools downgrading applications that mentioned women's sports clubs.
The system is not prejudiced. It is doing exactly what it was built to do: reproduce the patterns in its training data. That is what makes it dangerous, because the output looks neutral and mathematical.
Correlation is not understanding
The skin cancer example above is the classic case. The system found something that correlated with the answer without having any understanding of the actual problem, and it would fail completely on photographs taken a different way.
Who is responsible
If a self driving car causes a crash, who is at fault? The owner, the manufacturer, the programmer, the company that supplied the training data, or the regulator who approved it? There is no settled answer, and the law is well behind the technology.
The same question applies when an AI system refuses someone a loan, flags a student for cheating, or recommends a longer prison sentence.
A generated answer is produced because it is statistically likely, not because it is true. Systems that generate text will state false things with complete confidence, which is why anything important must be checked against a real source.
- The training data is unbalanced: 18 apples and only 2 pears.
- The model copies the patterns it finds in its data.
- So it is right 98% of the time for apples but only 40% for pears.
The wider impact of computing
Employment
Technology removes some jobs and creates others. Automated checkouts, warehouse robots and translation software have all replaced human roles.
The genuine problem is not the total number of jobs. It is that the jobs created often need different skills and appear in different places from the jobs lost, so the specific people affected are not the people who benefit.
Privacy
More of life is recorded than ever before: location, purchases, searches, messages, movement through streets covered by cameras. Individually each piece seems harmless. Combined, they describe a person in extraordinary detail.
Environment
- Training a single large AI model can use as much electricity as several households use in a year
- Data centres run continuously and need enormous cooling
- Device manufacturing consumes rare earth metals, and mining damages habitats
- E-waste contains toxic materials and is often exported to countries with weaker protections
Against this: video calls replace flights, digital documents replace paper, smart heating cuts energy use, and AI is being used to design more efficient materials and predict extreme weather.
How to think about it
The useful question is never "is this technology good or bad". It is:
- Who benefits?
- Who bears the cost?
- Who decided, and were the people affected consulted?
- What would need to be true for this to be acceptable?
Answer those four and you have a genuinely thoughtful argument rather than an opinion.
- Mining rare metals damages land.
- Making a device uses energy, water and chemicals.
- Using it needs electricity, including at the data centres.
- Throwing it away creates e-waste, with toxic metals in landfill.
- Recycling recovers materials, which go back into making new devices.
- Reduce, reuse and recycle, in that order of importance.
Key terms you must be able to define
- Artificial intelligence
- Computer systems performing tasks that would normally require human intelligence.
- Machine learning
- An approach where a system finds patterns in large amounts of example data rather than following written rules.
- Training data
- The labelled examples a machine learning system learns patterns from.
- Bias
- Systematic unfairness in an AI system's output, usually caused by unrepresentative training data.
- Algorithm
- A set of step by step rules for solving a problem or making a decision.
- Automation
- Using machines or software to perform work previously done by people.
- E-waste
- Discarded electronic equipment, often containing toxic materials.
- Accountability
- Being answerable for a decision, which becomes difficult when a machine made it.
Flashcards
Say the definition out loud before you flip the card. Retrieving it from memory is what builds recall, reading it again is not.
Reach the top grade on this topic
The top level here is about thinking clearly, not having strong opinions.
Explain the mechanism of bias. Not "AI can be racist", but "the system reproduces the patterns in its training data, so if that data reflects past discrimination the system will reproduce it while appearing neutral and objective".
Give both sides. AI both consumes enormous energy and helps design more efficient systems. Both are true and a good answer says so.
Name who is affected. Not "some people lose their jobs", but "checkout staff, who are often in low income areas with few alternative employers".
Reach a judgement. After both sides, say what you think should happen and why.
- Explain how machine learning works in four steps
- Explain why an AI system can be biased without anyone intending it
- Discuss the impact of automation on employment, giving both sides
- Explain why an AI generated answer should be checked against a real source
Mistakes that cost marks here
- Saying AI 'understands' or 'thinks'. It finds statistical patterns.
- Saying AI is biased because the programmers were prejudiced. The usual cause is unrepresentative training data.
- Giving only negative environmental impacts.
- Treating an AI generated answer as a source. It is confident, not verified.
- Writing a one sided answer on impact, which cannot reach the top level.
An AI system trained to spot skin cancer became extremely good at it, until researchers discovered it had partly learned that photographs containing a ruler were more likely to be cancerous, because doctors put a ruler in the picture when they suspected something serious.
Knowledge check
Exam-style questions
Five written questions in the style of the real paper. Write a full answer, then let the marker check it against the mark scheme.
- LThe system is given a large amount of example data
- LIt finds patterns in that data rather than following rules written by a programmer
- LIt can then apply what it has learned to new data it has not seen before
Machine learning is an approach to artificial intelligence in which a system is given a very large amount of example data rather than being given rules by a programmer. The system adjusts its internal values so that it becomes better at producing the correct answer for those examples, effectively finding statistical patterns in the data by itself. Once trained, it can apply those patterns to new data it has never seen before, for example classifying a photograph it was not trained on. Importantly, it has no understanding of what it is looking at, only patterns that correlate with the right answer.
- LThe system learns from training data
- LIf that data is unrepresentative or reflects past unfairness, the system learns that
- LThe system reproduces those patterns in its output
- LThis happens without anyone intending it, and the output looks objective
An AI system does not have rules written for it. Instead it learns patterns from its training data, so whatever is in that data shapes how it behaves. If the training data is unrepresentative, for example a facial recognition system trained mostly on light skinned faces, the system will simply be less accurate on the groups that were underrepresented. If the training data reflects past human decisions that were unfair, such as years of recruitment decisions at a company, the system learns those patterns too and continues to apply them. What makes this particularly dangerous is that nobody has to intend it: the bias arrives through the data rather than through anyone's prejudice, and because the output comes from a computer it can appear objective and mathematical, so it is trusted more than a human decision would be.
- LAdvantage: marking is much faster
- LAdvantage explained, such as students getting feedback sooner or teachers having more time
- LDisadvantage: it may not understand an unusual but correct answer
- LDisadvantage explained, such as unfair marks or students learning to write for the machine
An advantage is speed. An AI system can mark thousands of pieces of work in the time it takes a teacher to mark a handful, which means students receive feedback while the work is still fresh in their minds rather than a fortnight later, and it frees a great deal of teacher time for actual teaching. A disadvantage is that the system does not understand the subject, it recognises patterns, so an answer that is correct but phrased in an unusual or original way may be marked down simply because it does not match what the system learned to expect. Over time this risks teaching students to write for the marking system rather than to think clearly, which is the opposite of what education is for.
- LAutomation removes some existing jobs
- LGives a specific example such as checkout staff or warehouse workers
- LAutomation also creates new jobs
- LGives examples such as maintenance, programming or data roles
- LThe new jobs often need different skills from the jobs lost
- LThe people who lose out are often not the people who benefit
- LConsiders benefits such as lower prices, safer work or higher productivity
- LReaches a supported conclusion with a suggestion such as retraining or support
Increased automation clearly removes some existing jobs. Self service checkouts have reduced the number of cashiers needed in supermarkets, warehouse robots have replaced pickers, and automated systems have taken over much routine work in factories and call centres. At the same time automation creates new roles, including engineers who design and build the systems, technicians who maintain them, programmers who develop the software and analysts who work with the data they produce. The overall number of jobs in an economy has historically not collapsed because of automation. The real difficulty is that the jobs created are not the same jobs, do not need the same skills and are frequently not in the same places. A supermarket cashier made redundant in a small town does not simply become a robotics technician, because that requires qualifications, training and often relocation, and the areas that lose the most routine work are often those with fewest alternative employers. So the costs fall on specific people who did nothing wrong, while the benefits go to shareholders and to consumers generally. Those benefits are genuine: automation lowers prices, increases productivity and removes people from work that is dangerous, repetitive or damaging to health, and few would argue for putting humans back into the most hazardous factory roles. Overall, automation should not be resisted, because the productivity and safety gains are real and the alternative is simply falling behind. However, it cannot be treated as automatically fine either. Governments and employers have a responsibility to fund retraining, to give reasonable notice, and to support the regions most affected, because otherwise the pattern is that society as a whole gains while a specific and largely powerless group pays for it.
- LIt produces text that is statistically likely rather than verified as true
- LIt can therefore state incorrect information confidently
- LAnswers should be checked against a reliable source
A text generating AI produces each part of its answer because it is statistically likely to follow what came before, based on patterns learned from enormous amounts of text. It has no mechanism for checking whether what it is producing is actually true, and it does not know what it does or does not know. As a result it can state completely incorrect information, including invented statistics, references and quotations, in exactly the same fluent and confident tone it uses for correct information, which makes errors very hard to spot. Anything important that comes from such a system therefore has to be checked against a reliable independent source before it is used or repeated.