Child Labour vs Algorithmic Bias: The Price of Progress

Series: Industrial Revolution & Data Revolution #05/05 | Read time: 35 min | Python (NumPy, Pandas, Matplotlib)

Author: Wina @ Code & Cogito

When Sarah Entered the Mine

1. Yorkshire, England.

Eight-year-old Sarah Gooder woke every morning at four o’clock and walked into the darkness of a coal mine.

Her job: hauling coal trucks.

A single loaded cart weighed roughly a hundred kilograms. Sarah dragged it by a chain fastened around her waist, crawling on hands and knees through tunnels too narrow for an adult to stand in.

Fourteen hours a day. No windows, no sunlight — only darkness and coal dust.

A parliamentary investigator asked her: “Are you frightened?”

Sarah answered:

“I am frightened. There are things in the dark. Sometimes I cry, but no one hears me.”

The investigator recorded: “Sarah Gooder, age 8, height three feet (91 centimetres).”

This was the dark side of the Industrial Revolution.

Technological progress generated enormous wealth. But who bore the cost?

Children.

1. Seattle.

Amazon’s AI hiring system was screening resumes.

The training data: ten years of hiring records. Ninety percent of those hired had been men.

The algorithm learned: male = good employee.

The result: a resume mentioning “women’s basketball team captain” was automatically penalised. “Women’s shelter volunteer” — penalised. Graduated from a “women’s college” — penalised.

A female engineer — MIT graduate, five years of industry experience — was automatically rejected.

The reason: the algorithm determined she was not a good fit.

This was the dark side of the Digital Revolution.

Technological progress generated efficiency. But who bore the cost?

The vulnerable.

Two stories separated by 176 years. The same pattern:

Technology optimises for efficiency. Fairness is not in the objective function.

In this article, I use Python to analyse nineteenth-century child labour data alongside twenty-first-century AI bias cases, quantify the mathematical structure of discrimination, trace the arc of reform, and confront a question that refuses to go away: Does technology make society fairer, or does it calcify injustice?

The Child Labour Inferno of the Nineteenth Century

The Industrial Revolution created unprecedented wealth.

But who paid the price?

The Numbers Speak: Child Labour Rates

# British child labour data (1800-1900)
child_labor_data = {
    'year': [1800, 1820, 1840, 1860, 1880, 1900],
    'child_workers_thousands': [130, 190, 280, 320, 260, 180],
    'child_labor_rate': [7.2, 8.6, 10.0, 9.1, 6.2, 3.6],  # %
    'average_age_start': [8, 7, 7, 8, 9, 11],
    'average_work_hours': [13, 14, 14, 12, 10, 9]
}

# Industry distribution at peak (1840)
industry_1840 = {
    'industry': ['Textiles', 'Coal Mining', 'Agriculture', 'Domestic Industry', 'Other'],
    'percent': [35.7, 17.9, 28.6, 10.7, 7.1],
    'daily_hours': [14, 14, 12, 16, 12],
    'mortality_rate': [12, 25, 8, 10, 10]  # per 1,000
}

# The "cost advantage" of child labour
adult_wage = 12   # shillings/week
child_wage = 3    # shillings/week
# -> Cost advantage: 75% (children cost 75% less)
# -> 280,000 child workers saved employers ~650,000 pounds annually

Full code on GitHub (free baseline analysis): View complete code ->

Visualising these numbers, several brutal facts emerge:

The cruel arithmetic of child labour:

Why did employers use children?

The answer: cost.

An adult worker earned twelve shillings a week. A child cost three. A saving of 75%. With 280,000 children in the workforce, employers saved roughly 650,000 pounds annually — equivalent to perhaps 65 million pounds in today’s money.

This was the price of “efficiency optimisation.”

As Jane Humphries documented in Childhood and Child Labour in Industrial England (2010), the economics were brutally straightforward. Children were not merely cheap — they were physically suited to the narrow mine shafts and cramped spaces beneath textile machinery where adult bodies could not fit. Their small size was not a disadvantage to be pitied. It was a competitive advantage to be exploited.

The Turning Point of 1842: The Ashley Commission Report

Why did child labour numbers begin to decline in the 1840s?

Because of a report that shocked the nation’s conscience.

The Ashley Commission Investigation (1842)

Anthony Ashley-Cooper (later the 7th Earl of Shaftesbury) led a parliamentary inquiry into conditions in the mines.

What the investigators found:

Case 1: Sarah Gooder, age 8

“I haul coal trucks. The chain is fastened around my waist, and I crawl on my hands and feet. It is very frightening in the dark. I work fourteen hours a day. I am very tired, but I cannot stop, or I shall be beaten.”

Case 2: Patience Kershaw, age 17

“I have worked in the pit since I was eight years old. I am seventeen now, but I look like forty. My back is bent and my legs are bent. I cough blood. The doctor says I shall not live to twenty-five.”

Case 3: Anonymous, boy, age 11

“I have seen many people die in the mine. Once, a tunnel collapsed and my friend was crushed. He was only nine.”

The statistics:
– 402 mines investigated
– Hundreds of child workers interviewed
– 138 fatalities documented (over a five-year period)
– Children’s accident rate was 3.2 times the adult rate

The result: The Mines Act of 1842

The legislation:
– Prohibited all children under 10 from underground work
– Prohibited all women and girls from underground work
– Established mine inspectors

For the first time, the law explicitly protected children in the workplace.

But how effective was enforcement?

Visualising child labour data before and after the Mines Act, we can see the law’s direct impact:

The law’s effect:

The 1842 Mines Act worked: coal mine child labour fell 56% within eighteen years. But the textile industry, lacking comparably strict legislation, saw only a 35% decline.

The problem: why was enforcement so difficult?

Four factors were intertwined. Inspectors were grossly understaffed — four inspectors for four hundred mines, a physical impossibility. Fines were trivial — five pounds per violation, when the savings from child labour far exceeded the penalty, so employers treated fines as an operating cost. Poor families themselves depended on their children’s income, creating a structural trap. And employers lobbied fiercely, claiming that banning child labour would destroy British industry.

It took until the 1880s for child labour to be substantially eliminated.

From the first Factory Act in 1833 to the effective disappearance of child labour: fifty years.

Clark Nardinelli, in Child Labor and the Industrial Revolution (1990), argued that the decline was driven not solely by legislation but by a combination of rising adult wages (which reduced families’ dependence on children’s earnings), compulsory education, and a gradual shift in social norms about childhood. The law was necessary, but not sufficient. Change required the entire ecosystem to move.

176 Years Later: The Bias of the Algorithm

In 2024, we confidently declare: “We would never repeat the mistakes of the nineteenth century.”

Would we not?

Let us examine the modern counterpart — algorithmic discrimination.

Case 1: Amazon’s AI Hiring System (2018)

Background:
– Amazon sought to automate resume screening with AI
– Training data: a decade of hiring records
– The problem: 90% of past hires were male

What the algorithm learned:
– “Male” features = positive signal
– “Female” features = negative signal

The result:
| Resume Content | AI Score Change |
|—————-|—————-|
| Women’s basketball team captain | -0.5 points |
| Women’s shelter volunteer | -0.3 points |
| Graduated from Smith College (women’s college) | -0.4 points |
| Resume contains the word “women’s” | -0.2 points |

One female engineer:
– MIT Master’s, GPA 3.9
– Five years at Google
– Ten published papers
– Automatically rejected by the AI system

Amazon ultimately scrapped the system.

But the question remains: how many companies are still running similar systems?

Case 2: COMPAS Criminal Risk Assessment (2016)

Background:
– US courts used an AI system called COMPAS to assess recidivism risk
– Score range: 1-10, higher means more “dangerous”

What ProPublica’s investigation discovered:

White defendants vs. Black defendants (with identical criminal records):
– White defendants: average risk score 4.2
– Black defendants: average risk score 6.8
– Disparity: 64%

The more damning comparison:
– White repeat offenders: risk score 5.3
– Black first-time offenders: risk score 6.1

A Black person with no prior record was rated more “dangerous” than a white person who had already reoffended.

ProPublica’s “Machine Bias” investigation (Angwin et al., 2016) became a landmark in algorithmic accountability journalism. It demonstrated that the system’s errors were not random — they were systematically distributed along racial lines. The algorithm was not merely imperfect. It was imperfect in a direction that reinforced existing inequalities.

Case 3: Facial Recognition Error Rates (2018-2024)

Joy Buolamwini’s groundbreaking “Gender Shades” study at the MIT Media Lab tested commercial facial recognition systems and found:

The error rate for Black women was 43 times higher than for white men.

Why?

The training data:
– 78% white faces
– 84% male
– Only 6% dark-skinned women

Buolamwini’s work, published with Timnit Gebru in the Proceedings of Machine Learning Research (2018), did not merely document a technical failure. It revealed a social one. The underrepresentation of dark-skinned women in training datasets was not an accident — it reflected whose faces the technology industry considered worth collecting, labelling, and optimising for.

Quantifying AI Bias with Python

# Discrimination index calculation
def discrimination_index(majority_val, minority_val):
    """Positive values = disadvantage for minority group"""
    return (minority_val - majority_val) / majority_val * 100

# COMPAS risk score discrimination index
compas_di = discrimination_index(4.2, 6.8)   # -> +61.9%
# Black defendants scored 62% higher than white defendants

# Facial recognition discrimination index
face_di = discrimination_index(0.8, 34.7)    # -> +4237.5%
# Black women's error rate is 4,238% higher than white men's

# Training data imbalance -> error rate imbalance
# White males: 65% of training data, 0.8% error rate
# Black females: 6% of training data, 34.7% error rate
# -> Strong negative correlation between data representation and error rate

Visualising the bias data from all three cases together, the pattern becomes unmistakable:

The mathematical nature of AI bias:

AI is not “neutral.”

AI learns from historical data.

If the historical data contains bias, AI will amplify it.

Why?

Because AI optimises for prediction accuracy, not fairness.

Cathy O’Neil made this argument with devastating clarity in Weapons of Math Destruction (2016): mathematical models are not objective arbiters of truth. They are opinions embedded in code. And when those opinions are trained on data that reflects centuries of structural inequality, the models do not correct the inequality. They automate it. They scale it. They give it the false authority of mathematical precision.

Child Labour vs AI Bias: The Same Pattern

Let us set the nineteenth-century child labour crisis beside the twenty-first-century AI bias crisis.

Plotting both on radar charts and comparative timelines, the structural similarities are striking:

The parallels are uncomfortable:

The pattern has repeated.

How Do We Avoid Repeating History?

The child labour crisis took fifty years to resolve.

How long will AI bias take?

Approach 1: Algorithmic Auditing

Just as the 1842 Mines Act created mine inspectors, we need “algorithm inspectors.”

In practice:
– Mandatory disclosure of training data demographics
– Regular testing of error rates across demographic groups
– Public bias reports

Case in point: New York City Local Law 144 (2023)
– Requires AI hiring systems to publish bias audit reports
– Must test for differential impact across gender and race
– Violations carry fines of $500-1,500 per day

This is a start, but it is modest. As O’Neil has argued, the real challenge is not writing the law — it is building the infrastructure of enforcement. Four inspectors for four hundred mines was not enough in 1842. A handful of auditors for thousands of AI systems will not be enough in 2024.

Approach 2: Fairness Constraints

Embed “fairness” as an explicit objective in AI training.

Traditional AI optimises solely for prediction accuracy, ignoring disparities across groups. With fairness constraints, the system is required to keep error rate gaps below a specified threshold.

Simulation results are encouraging:

The effect of fairness constraints:

Traditional AI:
– Male accuracy 92%, female accuracy 85% (7% gap)
– Male false positive rate 5%, female 15% (10% gap)

Fairness-constrained AI:
– Male accuracy 90%, female accuracy 88% (2% gap)
– Male false positive rate 8%, female 10% (2% gap)

The trade-off: a slight reduction in majority-group accuracy in exchange for substantially greater overall fairness.

Approach 3: Diversified Training Data

The root cause demands a root solution: balance the training data.

The facial recognition case:
– Original data: 78% white, 6% dark-skinned women
– Rebalanced data: every group at least 20%

The effect:
– White male error rate: 0.8% -> 1.2% (slight increase)
– Black female error rate: 34.7% -> 4.5% (dramatic decrease)

Approach 4: Legislative Regulation

The EU AI Act (2024):

This is the most ambitious attempt yet to regulate artificial intelligence. It classifies AI systems by risk level:

• Unacceptable risk: Banned outright (e.g., social credit scoring systems)
• High risk: Strict regulation (e.g., hiring, credit scoring, criminal justice)
• Limited risk: Transparency requirements
• Minimal risk: No special requirements

High-risk AI systems must:
– Disclose training data characteristics
– Undergo bias testing
– Provide a human review mechanism
– Violations carry fines of up to 35 million euros or 7% of annual global revenue

The EU AI Act is, in effect, the twenty-first century’s Factory Act — an attempt to impose minimum standards of decency on a system that, left to market forces alone, will optimise for efficiency at the expense of the vulnerable. Whether it will be enforced more effectively than the 1842 Mines Act remains to be seen.

Conclusion: Does History Repeat?

Nineteenth-century child labour and twenty-first-century AI bias exhibit the same structural pattern:

Technological progress -> Efficiency optimisation -> Sacrifice of the vulnerable -> Social backlash -> Legislative intervention

The child labour crisis took fifty years to resolve (1833-1880).

How long will AI bias take?

The good news: we are learning faster.

It took forty-seven years from the first Factory Act to the effective elimination of child labour. It took only eight years from the emergence of AI bias as a public concern (2016) to comprehensive EU legislation (2024).

The bad news: the problem is harder to see.

Child labour was visible — eight-year-old Sarah in a coal mine, anyone who walked into the shaft would be horrified. AI bias is invisible — the algorithm operates inside a black box, and rejected applicants may never know they were discriminated against.

The essential question: Does technology make society fairer, or does it entrench unfairness?

The answer depends on us.

In the nineteenth century, society chose to legislate against child labour.
In the twenty-first century, what will we choose?

Inaction means AI continues to optimise for efficiency, bias becomes entrenched, marginalised communities are pushed further to the margins, and social tensions deepen. Active intervention means algorithmic auditing, fairness constraints, diversified data, and legislative regulation.

History tells us: the cost of inaction is always higher in the end.

In 1842, the British Parliament read Sarah Gooder’s testimony and chose to act.

In 2024, we have the data on AI bias before us. What will we do?

The answer is in our hands.

Afterword: Five Lessons from the Mine to the Algorithm

Having written the final article in this series, I find myself returning to the same image.

Yorkshire, 1842. A girl barely three feet tall, dragging a hundred-kilogram coal truck by a chain around her waist, crawling through darkness. When the parliamentary investigator asked if she was frightened, she said she cried sometimes, but no one heard.

That image stays with me not only because of its cruelty, but because in writing these five articles, I kept encountering the same structure — technology creates possibility, business models lower barriers, rapid adoption reshapes society, power concentrates in the hands of the few, and the cost falls on those with the least voice.

Article One: Watt’s steam engine and AWS cloud computing. One sold leases, the other sells on-demand compute — the commercial logic is identical. We calculated that cloud prices fall 5.6 times faster than steam engine costs did, but monopoly concentration is also accelerating. The HHI index does not lie — ownership of infrastructure is the distribution of power.

Article Two: Taylor’s stopwatch and Uber’s algorithm. The form of control has shifted from “the visible foreman” to “the invisible app,” but the substance has not changed. Workers have lost control over the rhythm of their own labour — only this time, they do not even know who is in control.

Article Three: Marx’s theory of alienation. I initially assumed this was an obsolete nineteenth-century concept. But when I applied its mathematical structure to the gig economy, the fit was unsettling. Separation of workers from their product, from their process, from their species-being — Uber drivers, food delivery riders, data labellers, every one of them maps onto Marx’s framework with uncomfortable precision.

Article Four: Data archaeology. Using Python to reconstruct the economic footprint of the Industrial Revolution, I understood for the first time what “long waves” truly mean. Kondratieff cycles are not abstract economic theory — they are the concrete trajectory of how technological revolutions reshape entire societies over fifty to sixty years. We are standing in the middle of the fifth wave — the expansion phase of the Digital Revolution, not yet at the peak.

Article Five — the one you are reading now. Sarah Gooder and the MIT-educated engineer rejected by Amazon’s algorithm are separated by 176 years, but their stories share the same mathematical structure: the objective function optimises for efficiency, and “fairness” is not a variable.

Across five articles, what struck me most was not any single historical fact or data point. It was the stubbornness of the pattern. The technology changes — from steam to cloud, from stopwatch to algorithm, from coal to data — but humanity’s response to technology has barely changed at all. First excitement, then deepening inequality, then social tension, then institutional adjustment. Each cycle takes decades. Each cycle exacts a disproportionate cost from those least able to bear it.

But this is not a pessimistic conclusion.

Because each time, society eventually found some kind of equilibrium. Child labour was abolished. Trade unions were recognised. Labour laws were passed. Not because those in power suddenly became compassionate, but because the contradictions accumulated until they could no longer be ignored, and change became unavoidable.

The question is: this time, can we move faster?

We have something the nineteenth century did not — the lessons of history, the tools of data analysis, and a global platform for dialogue. The EU AI Act was passed eight years after the problem surfaced; the first Factory Act took decades of campaigning. That is progress.

But we also face challenges the nineteenth century did not — algorithmic bias hides inside black boxes, unlike Sarah’s coal mine, which anyone could see with their own eyes. Digital-age injustice is more covert, and therefore harder to fight.

Writing this series, my original motivation was curiosity: how similar are the Industrial Revolution and the Digital Revolution, really? The answer: more similar than I imagined. Not as surface-level analogy, but as structural repetition. The same S-shaped diffusion curves, the same HHI concentration trajectories, the same arc of power from dispersal to concentration to challenge.

History does not repeat, but it rhymes.

If these five articles accomplish one thing, I hope it is this: the next time you read a headline proclaiming that “AI improves efficiency by 50%,” you pause and ask one more question: Efficiency has improved — but who bears the cost?

In 1842, no one heard Sarah crying in the darkness. Not until the Ashley Commission’s report made the entire nation listen.

In 2024, the people rejected by algorithms — are their voices being heard?

See the past clearly through data, and we can plan the future rationally.

That was the starting point of this series. And it is the ending.

Deep Dive: Complete Analysis Pack

This article shared the structural comparison between child labour data and AI bias — from industry distribution to discrimination indices, from Mines Act effectiveness to fairness constraint simulations. The complete analysis pack goes further:

• Child labour rate time series analysis: Cross-industry stacked area charts + GDP correlation dual-axis analysis, revealing hidden connections between child labour rates and economic cycles
• Algorithmic bias detection models: A unified quantification framework for Amazon hiring bias, COMPAS racial proxies, and facial recognition error rates
• Vulnerable group impact heatmaps: Approval rates across 8 demographic groups x 5 domains + annualised economic loss quantification
• Fairness evaluation framework: Demographic Parity / Equal Opportunity / Predictive Parity three-metric radar charts with adjustable parameters to observe the trade-offs between different fairness definitions
• ~400 lines of teaching-grade Python code with detailed English comments, 12 advanced visualisations (PNG 300dpi)

Get the Article 05 Deep Dive Pack ->

Series Retrospective: Five Lessons from the Industrial and Data Revolutions

Looking back across all five articles, a recurring pattern emerges:

1. Technological breakthrough creates new possibilities (steam engine / cloud computing)
2. Business model innovation lowers the barrier to entry (leasing / SaaS)
3. Rapid adoption reshapes social structures (factory system / platform economy)
4. Power concentration produces oligopoly (coal barons / tech giants)
5. Social contradiction triggers movements for change (labour movement / ?)

This pattern applies far beyond the Industrial and Digital Revolutions. In the forthcoming “Finance, Bubbles & Crises” series, we will see the same logic play out in capital markets — from the South Sea Bubble to the subprime crisis, from tulip mania to cryptocurrency, the mathematical structure of bubbles is remarkably consistent.

History does not repeat, but it rhymes. Our task is to use data to read the rhyme scheme.

References

• Parliamentary Papers (1842). Report on the Employment of Children in Mines.
• Buolamwini, J. & Gebru, T. (2018). “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.” Proceedings of Machine Learning Research, 81.
• Angwin, J. et al. (2016). “Machine Bias.” ProPublica.
• O’Neil, C. (2016). Weapons of Math Destruction. Crown Publishing.
• Humphries, J. (2010). Childhood and Child Labour in Industrial England. Cambridge University Press.
• Nardinelli, C. (1990). Child Labor and the Industrial Revolution. Indiana University Press.
• European Commission (2021). Proposal for a Regulation on Artificial Intelligence (AI Act).