How ML/AI changes society — and how to write about it in the HSC
You watched 3Blue1Brown — "But what is a neural network?" (Deep Learning, Chapter 1). Think through each answer before you reveal it.
The task: Recognise handwritten digits 0–9 from a 28×28 pixel grayscale image — the MNIST dataset. Each image is just a grid of pixel brightness values.
Input layer — 784 neurons: One neuron per pixel (28 × 28 = 784). Each neuron's activation value holds that pixel's brightness as a number between 0 and 1. Bright pixel → activation close to 1. Dark pixel → close to 0.
Output layer — 10 neurons: One per possible digit (0 through 9). After the network processes an image, you look at which output neuron has the highest activation — that's the network's guess. If neuron 7 lights up the brightest, the network thinks it's a 7.
Hidden layers: Two hidden layers of 16 neurons each sit between input and output, transforming the raw pixel data into increasingly abstract representations.
Weights: A number on each connection between neurons. It controls how strongly a neuron in the previous layer influences this one. Weights can be positive (this input pushes the neuron to activate) or negative (this input suppresses it).
Biases: A number added to the weighted sum before the activation function. It controls how easy or hard it is for a neuron to activate — a high bias means it fires readily; a very negative bias means it only fires with strong input.
How many? The 784 → 16 → 16 → 10 network has roughly 13,002 weights and biases in total. Every single one of those is a knob the network can turn.
Training = finding the right 13,002 numbers. The network starts with random weights and biases, runs thousands of training images through, measures how wrong its guesses are, and nudges each parameter slightly in the direction that reduces the error. Do that enough times and the parameters converge on values that work.
Tie-in from Lesson 19: The sigmoid function in the video — the S-shaped curve that squashes any number into 0–1 — is exactly the same function that logistic regression uses. When you called predict_proba() in scikit-learn, sigmoid was running under the hood to turn the weighted sum into a probability.
The NESA syllabus asks you to assess the impact of automation on: safety of workers, people with disability, and the nature and skills required for employment. These three are closely connected — automation changes who can work, how safely, and at what kind of job.
Automation can make dangerous jobs safer — but it can also create new hazards.
| Positive impact | Negative / new risk |
|---|---|
| Robots handle toxic, explosive, or physically dangerous tasks (e.g. mine site vehicles, chemical plants) | Workers alongside unguarded robots face collision risk — "cobots" (collaborative robots) need strict safety protocols |
| AI fatigue monitoring cameras detect drowsiness in truck and train drivers in real time | Constant surveillance creates psychological stress — workers report feeling watched and distrusted |
| Predictive maintenance ML models spot machinery failures before they happen, preventing accidents | Fewer skilled maintenance workers are employed, centralising safety knowledge in the ML system itself — a single point of failure |
| Exoskeletons with AI-assisted movement reduce musculoskeletal injuries in warehouse and construction work | Workers may be pushed to work faster because they can wear an exoskeleton — increased pace can cause other injuries |
The standard narrative is "AI takes jobs." The reality is more nuanced: it displaces routine tasks while creating demand for new roles — but the timing and distribution of that disruption is uneven.
Data entry, basic customer service (chatbots), manufacturing assembly, truck and taxi driving (autonomous vehicles), radiologist screening (ML image analysis), paralegal document review.
Common pattern: if the task can be reduced to a pattern in data, ML can learn it.
ML engineers, data scientists, AI ethics officers, prompt engineers, AI trainers and evaluators (labelling data), model auditors, cybersecurity specialists defending AI systems.
These roles typically require higher qualifications — raising the barrier to entry for displaced workers.
The workers most likely to be displaced (lower-skilled, older, in regional areas) are also the least likely to have access to reskilling programs. Automation can therefore worsen inequality even while improving overall productivity. This is a key HSC discussion point — don't present automation as universally good or bad.
This is a two-sided story. AI-powered tools can dramatically expand what people with disability can do — but the same AI systems can exclude them when bias or poor design is baked in.
| Technology | Benefit | Risk of exclusion |
|---|---|---|
| Speech recognition (Siri, voice-to-text) | Enables communication for people with motor impairments, RSI, or who are blind | Trained primarily on standard accents — accuracy drops significantly for non-standard speech patterns, including some conditions affecting speech |
| Real-time captions (auto-transcription in Zoom, YouTube) | Enables Deaf users to participate in meetings and consume video content | Error rate on accented, fast, or technical speech can be high — misleading captions can be worse than no captions |
| Facial recognition (e.g. unlocking phones, border control) | Convenient authentication without requiring motor precision | Higher error rates on people of colour; some systems fail entirely on people who are blind or have facial differences |
| Hiring algorithms | Can reduce human fatigue bias — sometimes more consistent than tired recruiters | If trained on past hiring data where disabled candidates were underrepresented, the algorithm encodes and perpetuates that exclusion |
A warehouse company automates its picking line with robots. 80 pickers lose their jobs. 12 new robot technician roles are created. The technician roles pay 30% more.
Net effect: 80 jobs lost, 12 created — that's a net loss of 68 jobs in that workplace. Even if the new roles pay more, the displaced 80 workers aren't automatically the ones who get the 12 new jobs.
The assumption needed for a positive outcome: the displaced workers can be reskilled into the technician roles, or find equivalent work elsewhere. For older workers or those without technical education, this often doesn't happen.
Most disadvantaged: workers with the least transferable skills, lowest savings, or most dependents — i.e., those who were already most economically vulnerable. Automation concentrates the productivity gains at the top while spreading disruption through the workforce.
This doesn't mean automation is wrong — it means the policy response (retraining, transition support, tax on automation gains) matters enormously.
ML-powered quality control catches defects faster than humans. Supply chain optimisation reduces overproduction. Predictive maintenance avoids costly downtime.
Agriculture AI reduces pesticide and water use. Smart grids balance power demand. But training large models (e.g. GPT-4) uses electricity equivalent to hundreds of transatlantic flights.
AI produces large productivity gains but those gains flow mostly to companies and shareholders, not workers. Platform monopolies (Google, OpenAI, Meta) concentrate power in a handful of firms.
AI is simultaneously one of the most promising tools for addressing climate change and one of the fastest-growing sources of energy consumption. This is the environmental paradox of ML:
For the HSC: Always present both sides and then take a position supported by evidence. "Automation always helps the environment" and "automation always hurts the environment" are both wrong. The answer is "it depends on how it's designed, powered, and deployed."
Automation raises productivity — more goods and services can be produced with fewer workers. In theory, this should raise living standards across the board. In practice, how those gains are distributed depends on policy and market structure:
| Who captures the gain? | Why? |
|---|---|
| Capital owners (shareholders) | Automation replaces labour with capital. Profit from efficiency goes to shareholders, not former workers. |
| Tech companies | AI is a winner-take-all industry — a handful of models dominate, giving those companies enormous pricing power. |
| High-skill workers | Engineers who build and maintain AI systems command premium salaries. AI augments their productivity further. |
| Consumers | Prices for some goods fall. Streaming, search, and navigation are free at point of use — subsidised by ad revenue. |
| Displaced workers | Receive little or none of the productivity gain from the automation that replaced them. |
This section is directly in the NESA syllabus: "Explore by implementation how patterns in human behaviour influence ML and AI software development — including psychological responses and patterns related to acute stress response."
The key insight is that this runs in both directions: human psychology shapes how ML systems are built, and ML systems then reshape human psychology.
The most powerful design pattern in consumer AI isn't a technical breakthrough — it's a loop borrowed from behavioural psychology.
Variable reward schedules are the most addictive form of reinforcement. A slot machine doesn't pay every spin — it pays unpredictably. That unpredictability is what keeps people pulling the handle. Social media and short-form video platforms are engineered around the same principle.
| Platform mechanism | Psychological effect | How ML drives it |
|---|---|---|
| Infinite scroll — no natural stopping point | Removes the friction that would naturally end a session | The feed is ranked by ML to maximise engagement; each scroll is a new variable reward trial |
| Likes, reactions, comments | Dopamine release on social validation; anxiety when it doesn't come | ML determines which posts to amplify, directly controlling who receives feedback |
| Autoplay (YouTube, TikTok, Netflix) | Switches engagement from active choice to passive consumption | Recommendation ML picks the next video optimised to keep viewing — even if it leads toward more extreme content |
| Notification timing | Intermittent alerts keep people checking the app throughout the day | ML predicts when each user is most likely to re-engage and schedules notifications for that moment |
If you're building a recommendation system and your only objective function is "maximise time spent," you will build an addiction engine by default — not because you intended to, but because that's what the maths optimises for. The way human psychology shapes ML is that the training signal (clicks, watch time, likes) encodes what captures attention, not what's good for users. Engineers who don't think about this reproduce harm at scale.
Acute stress response — the fight-or-flight mechanism — is triggered when the body perceives threat or urgency. ML systems can trigger it repeatedly, even in non-life-threatening situations.
Delivery riders and Uber drivers are managed entirely by ML: acceptance rates, ratings, timing, route efficiency. The algorithm monitors and judges constantly with no human mediator. Studies show elevated stress hormones, anxiety, and depression at rates significantly above non-gig workers.
Humans review content that AI flags as potentially harmful — graphic violence, child exploitation, self-harm. Moderators develop PTSD at high rates. The AI handles volume; humans absorb the psychological cost of what passes through the filter.
Online exam proctoring software (e.g. ProctorU) uses ML to detect "suspicious behaviour." Students report high anxiety from the feeling of constant surveillance — eyes tracked, keystrokes monitored, room scanned. False positives can invalidate exams.
Productivity ML tools score employees on keystrokes per minute, mouse movement, time between responses. Amazon warehouse workers are timed on "time off task." The constant measurement creates sustained low-level stress — a chronic, not acute, version of the fight-or-flight response.
Machine learning doesn't exist in a cultural vacuum. The data it's trained on, the people who label it, and the communities it's deployed in all carry cultural context. When that context is ignored, ML systems cause real harm.
Many Indigenous communities assert that data about their people, language, land, and cultural practices should remain under their control — not available for commercial ML training.
The CARE Principles for Indigenous data governance were developed as a counterpart to FAIR data principles:
| Principle | Meaning |
|---|---|
| Collective Benefit | Data ecosystems should enable Indigenous Peoples to derive benefit from the data. |
| Authority to Control | Indigenous Peoples' rights and interests in Indigenous data must be recognised and their authority to control such data respected. |
| Responsibility | Those working with Indigenous data have a responsibility to share how that data is used to support Indigenous self-determination. |
| Ethics | Indigenous Peoples' rights and wellbeing should be the primary concern at all stages of the data life cycle. |
Real example: Voice recognition trained on Standard Australian English performs poorly on Aboriginal English and Indigenous language speakers — reducing access to AI-powered services for the communities who already have less access to technology. The training data choice was a cultural decision that had discriminatory outcomes.
How people think about AI and automation is deeply influenced by worldview and religious belief. These aren't just abstract philosophy — they shape policy, regulation, and public acceptance of ML systems.
| Framework or belief | How it shapes views on AI |
|---|---|
| Utilitarianism (maximise overall wellbeing) | Supports automation if it reduces suffering overall — even if some individuals are harmed. Tends to justify large-scale deployment if aggregate outcome is positive. |
| Deontological ethics (rules and rights matter regardless of outcome) | Argues that individuals have rights that can't be overridden by aggregate benefit. An algorithm that's "mostly right" can still violate someone's right to fair treatment. |
| Religious traditions (varies widely) | Some traditions raise concerns about creating artificial consciousness or making life-and-death decisions algorithmically. Others focus on stewardship responsibilities. Most major traditions emphasise dignity of the individual — a common ground with human rights frameworks. |
| Collectivist cultures | May be more open to surveillance and data sharing if framed as community benefit (e.g. pandemic contact tracing). Individualist cultures prioritise consent and privacy more strongly. |
For the HSC: You don't need to argue for a particular framework. You need to show that you understand different perspectives exist and can explain how they lead to different conclusions about AI. A strong response acknowledges the complexity rather than picking one "right" view.
This is one of the most HSC-examined topics in Software Automation. The NESA syllabus asks you to "investigate the effect of human and dataset source bias in the development of ML and AI solutions."
There are two distinct types to understand: dataset source bias (problems in the training data) and human bias (problems introduced by the people making the ML decisions).
ML models learn from data. If the data is biased, the model learns to reproduce and often amplify that bias.
| Bias type | What it means | Example |
|---|---|---|
| Historical bias | Training data reflects past discrimination, which the model treats as the "correct" pattern | Amazon's hiring AI trained on 10 years of résumés — mostly from men, because the tech industry historically hired mostly men. It learned to penalise résumés that included the word "women's" (e.g. "women's chess club"). |
| Representation bias | Some groups are underrepresented in the training data | Early facial recognition systems were trained predominantly on lighter-skinned faces. MIT researcher Joy Buolamwini found error rates of up to 34% on darker-skinned women vs <1% on lighter-skinned men. |
| Measurement bias | The feature being measured is a proxy that disadvantages some groups | COMPAS algorithm predicted criminal recidivism using features correlated with poverty (e.g. family members with criminal records) — which correlates with race due to systemic inequality. ProPublica analysis found it falsely flagged Black defendants as high-risk at twice the rate of white defendants. |
| Confirmation bias (human-introduced) | Labellers apply their own biases when annotating training data | If images of "professionals" in a training set are labelled by people who unconsciously associate professionalism with being male and white, those biases get encoded into any model trained on that data. |
Amazon built an ML system to rank job applicants. It was trained on résumés submitted over 10 years — most of which came from men, because tech hiring has historically skewed male.
Bias type: Historical + representation bias.
Outcome: The system penalised résumés that mentioned "women's" anything, and downgraded graduates of all-women colleges. Amazon shut the project down.
The lesson: "Past hiring decisions" is not neutral data. It encodes who was given a chance before.
Used in US courts to predict the likelihood a defendant would reoffend. Scores were used to inform bail, sentencing, and parole decisions.
Bias type: Measurement bias — features correlated with race through socioeconomic proxies.
Outcome: ProPublica (2016) found it classified Black defendants as high-risk at twice the rate of white defendants who did not reoffend. Real humans in real courts made decisions based on a biased algorithm — without knowing it.
The lesson: Algorithmic decisions in high-stakes contexts (law, healthcare, employment) need bias audits.
MIT Media Lab's "Gender Shades" study (Buolamwini & Gebru, 2018) tested commercial facial recognition systems from IBM, Microsoft, and Face++.
Bias type: Representation bias — training sets dominated by lighter-skinned faces.
The lesson: A system can achieve 99% overall accuracy while being dangerously wrong for specific subgroups. Aggregate metrics hide distributional harm.
Diverse and representative datasets — actively recruit data from underrepresented groups rather than relying on what's convenient.
Bias audits — test model performance across demographic subgroups before deployment, not just on aggregate accuracy.
Diverse teams — people with different lived experiences notice different failure modes. Homogeneous teams build systems that work well for people like themselves.
Impact assessments — before deploying ML in high-stakes contexts (criminal justice, healthcare, hiring), require a formal assessment of potential harm to specific groups.
Read each scenario. Name the bias type and explain the mechanism.
Bias type: Historical + measurement bias.
Historical banks may have "redlined" rural areas — routinely denying loans regardless of individual creditworthiness. The model treats past denial as evidence of default risk. "Rural postcode" becomes a proxy that has nothing to do with ability to repay.
The model is technically doing what it was asked (predicting defaults based on historical patterns) but those patterns encode past discrimination.
Bias type: Representation bias + human (confirmation) bias in labelling.
The moderators only flag what they recognise as toxic — their cultural context shapes the training signal. Content that would be flagged in any culture gets through because the annotators didn't recognise it. And content from their own cultural background may be over-flagged.
Fix: moderator teams should represent the communities whose content they are labelling.
Bias type: Historical + representation bias.
If the training data shows that software engineering roles were historically filled by men (because that was true, due to systemic barriers), the model predicts that men are more likely to click on or apply to those roles, and optimises accordingly. The algorithm amplifies the historical underrepresentation rather than correcting it.
This was documented in a real study of Google and Facebook ad targeting in 2019 — ML optimised for engagement in ways that reinforced gender segregation in job markets.
The impact and ethics material appears in HSC exams as extended response questions — typically 4, 6, or 8 marks. The mark allocation signals how much depth is expected. Short-answer bullet points do not earn full marks on extended responses.
Structure formula for extended response: State the impact → explain the mechanism → give a specific example → acknowledge the complexity (positive and negative, or uneven distribution of effect). Avoid generalisations like "AI affects everyone."
"Assess the impact of machine learning on the employment of workers in a specific industry."
Tip: choose one industry. "Across many industries" spreads your answer too thin for 4 marks. Name a specific technology and a specific job category.
In the transport industry, ML-powered autonomous vehicle technology is progressively displacing truck drivers while creating demand for new roles in vehicle maintenance and fleet monitoring software.
Truck driving is among the most common occupations in many countries, with a largely older, male workforce with limited formal technical qualifications. As companies like Waymo and Tesla deploy semi-autonomous trucks, the long-haul segment of the industry is beginning to automate — reducing the need for full-time drivers on fixed-route highway journeys while human oversight remains required in complex urban environments.
However, the impact is not uniformly negative. Fleet operations now require ML model engineers, remote safety monitors, and maintenance technicians for sensor and software systems — roles that did not exist in traditional logistics. The critical issue is the transition gap: the reskilling required for these new roles demands technical education that is not easily accessible to drivers mid-career. The productivity gains from automation therefore flow primarily to shareholders and consumers (through lower freight costs), while displaced workers bear the costs of displacement.
A balanced assessment acknowledges both the genuine efficiency gains and the policy challenge of ensuring those gains are equitably shared.
"Explain how bias in training data can lead to unjust outcomes when ML models are deployed in high-stakes decision-making. Use a specific example to support your answer."
Tip: name the specific bias type, explain the mechanism (not just "it's biased"), and address why this is a problem in high-stakes contexts specifically — not just any ML use.
Bias in ML training data occurs when the dataset used to train a model does not accurately represent the full population the model will encounter, or when it encodes historical patterns of discrimination as if they were correct signals.
One clear mechanism is historical bias: if an ML model is trained on decisions made by humans who were themselves biased, the model learns to reproduce those biased decisions at scale. A well-documented example is the COMPAS recidivism algorithm, used in US courts to predict the likelihood that a defendant would reoffend. The algorithm used features such as peer and family criminal records — factors correlated with socioeconomic status, which is in turn correlated with race due to systemic inequality. A ProPublica investigation found that the model falsely classified Black defendants as high-risk at nearly twice the rate of white defendants who did not go on to reoffend, and falsely classified white defendants as low-risk at twice the rate of Black defendants who did reoffend.
The harm is amplified in high-stakes contexts because the consequences of an incorrect prediction are severe and often irreversible — defendants may be denied bail, given longer sentences, or denied parole based on an algorithm they had no opportunity to challenge. Unlike a biased hiring algorithm, where a candidate can apply elsewhere, a person cannot undo time served in prison.
Mitigation requires auditing model performance across demographic subgroups rather than relying on aggregate accuracy, which can mask distributional harm. It also requires critically examining whether the features used as inputs carry proxied discrimination from historical inequalities.
"Describe TWO ways in which patterns of human psychological behaviour influence the design of ML-based recommendation systems."
Tip: "describe" means one point isn't enough — give the mechanism, not just the label. Two well-explained points score more than four shallow ones.
1. Variable reward scheduling: Humans respond most strongly to unpredictable rewards — a pattern exploited by social media and short-form video platforms. Recommendation ML systems are trained to maximise engagement metrics (clicks, watch time, shares). Because unpredictable content that occasionally exceeds expectations creates stronger dopamine responses than consistently average content, optimising for engagement causes the algorithm to surface emotionally intense or surprising content disproportionately. The psychological tendency toward variable reward therefore shapes the training objective of the ML system.
2. Loss aversion and fear of missing out (FOMO): Notification systems in social platforms use ML to determine when to alert users. Because humans respond more strongly to the threat of missing something than to the promise of gain, ML-optimised notification timing is designed to trigger this loss aversion — sending alerts when engagement has dropped, implying the user is "missing out." The psychological response pattern is encoded into the ML's scheduling decisions, extending and re-initiating sessions.
Write your own response to the following question. Aim for 5–6 minutes of focused writing. No notes.
"Discuss how cultural protocols or belief systems can be affected by the widespread deployment of ML and AI technologies. Refer to a specific community or framework."
Indigenous communities in Australia have increasingly asserted data sovereignty principles in response to the widespread collection and commercialisation of data about their people and practices. Facial recognition systems trained on images scraped from the internet, for example, may include images of Indigenous ceremonies or individuals taken without consent — data that has cultural and spiritual significance beyond its photographic content.
The CARE Principles for Indigenous Data Governance (Collective Benefit, Authority to Control, Responsibility, and Ethics) were developed specifically as a framework for how ML and AI development should engage with Indigenous data. These principles assert that Indigenous communities should control how data about them is used, not merely consent to collection.
This creates a direct tension with the dominant ML development paradigm, which treats large-scale public data collection as a legitimate training data source. When voice recognition systems underperform for Aboriginal English speakers, or when Indigenous cultural knowledge is used to train commercial AI without consent, the deployment of ML technology violates cultural protocols that predate the technology itself.
Engaging with these protocols — by seeking community consent, paying for data use, and building ML systems that serve affected communities — requires ML engineers to operate beyond purely technical criteria and engage with governance frameworks rooted in cultural and ethical responsibility.
How automation impacts worker safety — both reducing physical danger and introducing new hazards and psychological stress
Why employment disruption is not simply "job losses" — it's displacement of routine tasks, creation of new technical roles, and an unequal transition
How AI-powered accessibility tools benefit people with disability while poorly designed systems exclude them
The environmental paradox of ML — useful for climate solutions while also a significant and growing energy consumer
How variable reward schedules and dopamine psychology are deliberately engineered into ML recommendation systems
What acute stress response looks like in gig economy and content moderation contexts
How cultural protocols (CARE principles, Indigenous data sovereignty) create obligations for ML engineers beyond technical requirements
The distinction between dataset source bias and human bias, with specific case studies (Amazon, COMPAS, facial recognition)
How to write structured extended-response HSC answers on ML impact and ethics
DevOps/RPA/BPA, AI vs ML, types of learning, key algorithms and applications
OOP ML models, train/test split, linear regression, polynomial regression, logistic regression, neural networks, MLPipeline class
Worker safety, employment, disability, economy, environment, psychological responses, cultural protocols, bias case studies, HSC exam practice
Research task (one paragraph, 5–8 sentences)
Find an AI or ML system that caused a real-world harm, controversy, or unexpected impact. It can be anything — doesn't have to be from today's lesson. Something you find genuinely interesting is better than something you already know from class.
Your paragraph should cover:
Don't reuse the COMPAS, Amazon, or facial recognition cases from this lesson — find your own. Bring it to next session and we'll use it as exam practice material.