INNOVAREModule 5 · Ethical AI

Case 6: The Layoff Trap

Individually rational. Collectively catastrophic. Understanding why equitable AI requires policy — not just organisational goodwill.

July 2026 · Case 6 of 6
As you read — hold this question

If every CEO in a competitive market knows that mass AI-driven layoffs will eventually reduce the wages of their own customers — why does each one do it anyway?

60pp
mathematical proof. Tsoukalas & Falk (2025) modelled the dominant strategy outcome: each CEO's rational decision produces a collective economic catastrophe.

Equitable distribution of AI's benefits is the final — and hardest — question in responsible AI. It is hard not because organisations are malicious, but because the most harmful outcomes can arise from individually rational decisions that no single actor controls. Tsoukalas and Falk built a mathematical model of what happens when AI is used primarily to replace workers at scale in competitive markets. The finding is not that CEOs are bad. The finding is that the competitive structure makes the harmful outcome a dominant strategy: each firm has a stronger incentive to defect from collective restraint than to cooperate — even when every CEO can see the cliff.

The dominant strategy problem
Quiz: Equity & CARE

Why the layoff trap is a game theory problem, not just an ethics problem

The AI layoff trap is a classic prisoner's dilemma applied to labour markets. The structure of competitive markets means that self-restraint is individually irrational even when collective restraint is socially optimal.

Market structure CEO's incentive Outcome
Monopoly Restraint is rational: my workers are also my customers. If I fire my customers, demand for my product falls. The harm is internalised — the monopolist bears the cost of their own workers' reduced purchasing power. CEO restrains AI-driven headcount reduction to protect their own market
Competitive market Firing workers doesn't hurt me specifically: my employees spend their wages across the whole economy, not mainly with me. Competitors are doing the same thing. If I don't cut costs, I lose market share. Each CEO fires workers. Wages fall economy-wide. Consumer demand falls. But the harm is distributed — no single firm feels the full cost of their own decision.
Why self-restraint fails
In a competitive market, if your competitor replaces workers with AI and reduces costs, they gain an advantage whether you do the same or not. If you also replace workers, you match them. If you don't, you're at a cost disadvantage. The dominant strategy — the choice that produces the best outcome regardless of what competitors do — is always to replace workers with AI. Each CEO sees the cliff. Each drives toward it anyway. Tsoukalas and Falk call this "individually rational, collectively catastrophic." Fixing it requires a structural intervention — not asking individual organisations to act against their competitive interests.
Equity starts at design

Why equitable distribution is a design problem before it's a distribution problem

Equitable distribution of AI benefits cannot be achieved by deciding fairly who gets the outputs of a system that was designed to exclude certain groups from the beginning. The design must be equitable first — before any distribution question arises.

Gender Shades is the clearest illustration. IBM's facial recognition system had error rates of up to 34.7% for darker-skinned women. That is not a distribution failure — it is a design failure that produced an equity failure. The system worked for lighter-skinned men. It did not work for darker-skinned women. No distribution mechanism could have corrected for a system that was fundamentally broken for one group of users from the start.

The implication: equity analysis must happen at the point of data selection, labelling, objective setting, and success definition — the same four upstream human decisions that determine whether bias enters the system. An equity review conducted after deployment is too late.

CARE Principles

Indigenous data governance — who benefits, who controls, who is responsible

The CARE Principles for Indigenous Data Governance were developed to address a specific form of the equity problem: data about and from Indigenous communities has historically been used for others' benefit, without those communities controlling the terms of use or sharing in the value created. The principles apply beyond the Indigenous context to any community whose data is extracted without their meaningful participation in benefit or control.

C
Collective Benefit

Data ecosystems should enable Indigenous Peoples to benefit from their own data — not just contribute to it. The question is not whether data is collected with consent, but whether the community that generated it captures any of the value it creates. Most AI training data pipelines fail this test.

A
Authority to Control

Indigenous communities have the right to govern their own data. This right must be recognised explicitly — not assumed away by terms of service or bundled consent. Authority to control means the right to decide what data is collected, how it is used, who accesses it, and when it is deleted.

R
Responsibility

Those who use Indigenous data have an active obligation to explain how it is being used and to ensure that use advances the wellbeing of the community it came from. Responsibility is not passive compliance — it is an ongoing obligation to the source community.

E
Ethics

The rights and wellbeing of the community come before the convenience or commercial interest of the data user. Ethical data use is not a constraint on value creation — it is a condition for the legitimacy of value creation that rests on others' data.

What policy must do

The three levers for equitable AI that organisations cannot pull alone

Equitable AI outcomes cannot be achieved through organisational goodwill alone when the competitive structure produces dominant strategies that are individually rational but collectively harmful. Three policy-level levers are required.

Lever What it does Why the market alone can't do it
Labour transition policy Skills retraining investment, portable benefits, income support during displacement — paid for in part by the productivity gains AI enables Each individual firm that invests in training workers creates a public good that competitors free-ride on; only collective action (policy) solves the coordination failure
Benefit distribution mechanisms Mechanisms to distribute AI productivity gains more broadly — through tax policy, profit sharing, or sovereign AI funds Markets distribute productivity gains to capital owners by default. Without policy intervention, AI amplifies existing wealth concentration rather than reducing it
Data governance frameworks Mandatory requirements for equitable design, CARE Principle compliance in Indigenous data contexts, consent frameworks that match the actual scope of data use Voluntary CARE compliance produces race-to-the-bottom dynamics where firms that extract data without community participation outcompete those that don't
Take this away

Equitable AI is not a values question that organisations can solve by trying harder. The dominant strategy structure of competitive markets means that individually rational AI deployment decisions can produce collective harm no matter how well-intentioned each actor is. Ethics must come first — before the market logic takes over. And where the market logic is structural, policy is the only intervention with sufficient reach. Organisations that wait for policy are already behind. Organisations that treat equity as a design requirement — not a post-deployment adjustment — are building what the mandatory regime will eventually require anyway.

Quick recall — without looking back

Test yourself on this case

Question 1 of 3

What is the AI layoff trap — and why does the competitive market structure make self-restraint an irrational strategy for individual firms, even when CEOs can see the collective harm?

The AI layoff trap describes the dominant strategy outcome when AI is used primarily for headcount reduction in competitive markets. In a monopoly, a CEO restrains mass layoffs because their workers are their customers — firing customers reduces demand for the firm's own product. In a competitive market, a worker's wages are spread across the whole economy, not mainly spent at their employer. So the cost of firing workers is distributed — the individual firm doesn't bear the harm of its own workers' reduced purchasing power. If competitors use AI to reduce labour costs, a firm that doesn't faces a competitive disadvantage. If everyone does it, wages fall economy-wide and consumer demand collapses — but no individual firm feels the full cost of their own decision. The dominant strategy (the choice optimal regardless of what others do) is always to replace workers. Each CEO sees the cliff; each drives toward it anyway. Tsoukalas and Falk modelled this mathematically in a 60-page paper and found that the outcome is not irrational behaviour — it is a dominant strategy. Fixing it requires structural intervention, not asking individual firms to act against their competitive interests.
Question 2 of 3

What are the four CARE Principles for Indigenous data governance — and how does each principle apply beyond the Indigenous context to any community whose data is used in AI systems?

C — Collective Benefit: data ecosystems should enable communities to benefit from their own data, not just contribute to it. Beyond Indigenous contexts: any community whose data trains a commercial AI product without receiving any of the value that product generates is failing this principle. A — Authority to Control: Indigenous communities (and by extension, any data-generating community) have the right to govern how their data is collected, used, and shared. Beyond Indigenous contexts: the consent bundled into most terms of service does not constitute genuine authority to control — users do not meaningfully decide how their data is used in model training. R — Responsibility: those who use the data have an active obligation to explain how it is being used and to advance the wellbeing of the source community. Beyond: this applies to any organisation using community-generated data for commercial benefit without disclosure. E — Ethics: the rights and wellbeing of the community come before the commercial interest of the data user. Beyond: any organisation claiming ethical AI while using community data in ways that harm that community is violating this principle regardless of formal consent.
Question 3 of 3

Why is equitable AI a design problem before it is a distribution problem — and what does the Gender Shades study illustrate about where equity interventions must begin?

Equitable distribution of AI benefits cannot be achieved by allocating outputs fairly from a system that was designed to exclude certain groups from the beginning. The Gender Shades study found IBM's facial recognition had error rates of up to 34.7% for darker-skinned women, versus below 1% for lighter-skinned men. This was not a distribution failure — the product itself did not work for a substantial portion of potential users. No subsequent distribution mechanism could compensate for a system that was broken for darker-skinned women from the point of design. The failure originated in the four upstream human decisions: who chose the training data (insufficient diversity), who labelled it, what the system was optimised for (overall accuracy, which masked subgroup failures), and how success was defined (aggregate metrics that hid the 34.7% error rate for one group). Equity analysis must be conducted at those four design points — before deployment — not in post-launch distribution decisions. An equity review conducted after deployment is too late to fix a design that excluded people before the algorithm ran.

Module 5 Videos

Module 5 · Short Video
Module 5 · Long Form · Whose Name Is On That Decision?

Sources

Tsoukalas & Falk
Tsoukalas, G. & Falk, M. (2025). The AI Layoff Trap: A dominant strategy analysis of AI-driven labour displacement. Working paper.
CARE Principles
Carroll, S.R. et al. (2020). The CARE Principles for Indigenous Data Governance. Data Science Journal, 19(1), 43.
Gender Shades
Buolamwini, J. & Gebru, T. (2018). Gender Shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research, 81, 1–15.
Module content
BUSN9049 Module 5 — Ethical Considerations and Responsible AI. Flinders University, 2026.
Innovare Study
Long-form video: Whose Name Is On That Decision? Innovare Study, July 2026. youtu.be/tslQKmfxwk0