Appendix E: Argument Maps

Structured Maps of the Major Contested Debates in AI Ethics

How to Use This Appendix

Each argument map addresses a genuinely contested question in AI ethics — one where reasonable, informed people reach different conclusions. The maps are not designed to tell you what to think. They are designed to help you think more clearly by:

Stating the question precisely
Presenting the strongest arguments on each side (steelmanned, not strawmanned)
Identifying where the disagreement is empirical (a dispute about facts) vs. normative (a dispute about values)
Noting what evidence or argument would, in principle, resolve the disagreement

The debates are presented in order of increasing philosophical abstraction — from regulatory design questions to foundational questions about existential risk. All are live debates in 2025.

Debate 1: Innovation vs. Precaution — Should High-Risk AI Be Banned Until Proven Safe?

The Question

When a new AI application — predictive policing, automated hiring, autonomous weapons, medical diagnosis — is deployed in a high-stakes context, should regulators require proof of safety before deployment ("precautionary approach") or permit deployment subject to monitoring and corrective action ("permissive approach")?

Arguments for the Precautionary Approach

Argument 1 — Asymmetric harm: The harms of premature deployment (false arrests, denied loans, medical errors) are real and immediate; the harms of delayed deployment are speculative efficiency losses. When harm distributions are asymmetric, precaution is rational.

Argument 2 — Market failure: Companies do not bear the full costs of the harms their AI systems cause. This externality means they systematically underprovide safety relative to the social optimum. Regulation corrects this market failure by requiring safety demonstration before deployment.

Argument 3 — Irreversibility: Some harms caused by AI deployment are irreversible — a person wrongly incarcerated, a community displaced by automated surveillance. Precaution is more valuable when harms cannot be undone.

Argument 4 — Power asymmetry: Those who bear the costs of premature deployment (typically marginalized communities) are different from those who capture the benefits (typically technology companies and their customers). Precaution protects the less powerful.

Argument 5 — Historical precedent: We require proof of safety before deploying new drugs, aircraft, and nuclear reactors. The "innovation vs. precaution" framing implies that AI is different; proponents of precaution argue it is not.

Arguments for the Permissive Approach

Argument 1 — Opportunity costs: Delayed deployment of beneficial AI has real costs: lives not saved by earlier medical AI, crimes not prevented by better prediction tools, inefficiencies not corrected. These costs are less visible than deployment harms but not less real.

Argument 2 — Proof of safety is impossible: No complex system can be proven safe in advance. The question is whether the expected benefits exceed the expected harms. Requiring impossible proof is a covert ban.

Argument 3 — Comparative baseline: The alternative to AI decision-making is human decision-making, which is also biased, inconsistent, and opaque. The relevant question is whether AI is safer than the status quo, not whether AI is safe in absolute terms.

Argument 4 — Monitoring is effective: Modern AI systems generate data that enables real-time monitoring and correction. A regime of deploy-then-monitor is more adaptive than pre-deployment prohibition.

Argument 5 — Regulatory capture: Precautionary regulation tends to entrench incumbents by raising compliance costs too high for new entrants. This concentrates AI development in large companies, potentially reducing innovation and diversity of approach.

Key Empirical Disagreements

How common are false positives in high-risk AI deployments, and what are their costs?
What is the counterfactual? How good are the human decisions AI replaces?
How effective is post-deployment monitoring in practice?
What are the actual opportunity costs of delayed deployment?

Key Value Disagreements

Who should bear the burden of proof — those who want to deploy, or those who want to restrict?
How should harms to identifiable individuals be weighted against diffuse benefits to populations?
Should precaution apply symmetrically to both action (deployment) and inaction (non-deployment)?

What Would Resolve This Debate?

Large-scale comparative studies of AI vs. human performance in specific high-stakes decisions, combined with data on the actual costs of regulatory delay, would narrow the empirical dispute. The normative question of how to weigh certain individual harms against uncertain aggregate benefits cannot be resolved empirically.

Debate 2: Fairness of Algorithmic vs. Human Decision-Making

The Question

Is algorithmic decision-making fairer than human decision-making? Or does it simply replace one form of bias with another that is more systematic and less correctable?

Arguments That Algorithms Are Fairer

Argument 1 — Consistency: Human decisions vary based on irrelevant factors — time of day, the decision-maker's mood, weather. Algorithms apply the same criteria uniformly to all applicants, eliminating this "noise."

Argument 2 — Transparency: An algorithm's criteria can (in principle) be documented, audited, and challenged. A human decision-maker's reasoning is invisible. Algorithmic transparency enables accountability that human decisions do not.

Argument 3 — Scalability of oversight: It is impractical to audit every human decision-maker. A single audit of an algorithm covers all decisions made by that algorithm.

Argument 4 — Eliminating explicit bias: Humans often discriminate based on visible characteristics (race, gender, disability) in ways that algorithms trained on relevant features may not. An algorithm that uses only predictive features cannot intentionally discriminate.

Argument 5 — Empirical track record in some domains: Studies in some domains (recidivism risk, credit risk) have found that simple statistical models outperform clinical judgment on accuracy and reduce expert variability.

Arguments That Algorithms Are Not Fairer

Argument 1 — Bias in training data: Algorithms trained on historical data inherit and often amplify historical biases. A system trained on historically biased hiring decisions will perpetuate those decisions at greater scale.

Argument 2 — Opacity in practice: Despite theoretical auditability, in practice algorithms are often trade secrets, their training data is proprietary, and affected individuals have no practical ability to challenge their scores. Theoretical transparency without legal access is not transparency.

Argument 3 — Proxy discrimination: Algorithms that do not use protected characteristics directly may use proxies (zip code, names, network connections) that are highly correlated with protected characteristics. Absence of explicit protected characteristics does not ensure absence of discrimination.

Argument 4 — Scale of harm: When a human discriminates, the harm is local. When an algorithm discriminates, the harm is applied to every person processed by the system — potentially millions of people. Algorithmic bias is industrial-scale bias.

Argument 5 — Feedback loops: Algorithmic decisions change the world in ways that affect future training data. Predictive policing sends more police to high-prediction neighborhoods, generating more arrests in those neighborhoods, which confirms the predictions. Human decisions do not create these systematic feedback loops at scale.

Key Empirical Disagreements

How large is decision-making "noise" in human judgments in specific domains, and how large is the resulting harm?
What is the actual accessibility of algorithmic criteria to affected parties in practice?
In which domains have algorithms demonstrably outperformed human judges on both accuracy and fairness?

Key Value Disagreements

Is consistency (same treatment) the right conception of fairness, or does fairness require attending to different circumstances?
How should we weigh the costs of individual discriminatory decisions against the costs of systematic algorithmic discrimination?
Should the ability to challenge a decision be treated as an independent value, separate from the accuracy of the decision?

What Would Resolve This Debate?

Domain-specific comparisons of algorithmic and human decision-making on both accuracy and bias metrics, with attention to whether the comparison is to actual human behavior or to idealized human behavior, would address the empirical dispute. The normative questions are genuinely difficult.

Debate 3: Technical vs. Structural Fix — Can Fairness Metrics Resolve Algorithmic Discrimination?

The Question

Can the problem of algorithmic discrimination be resolved by better technical tools — fairness constraints, bias audits, debiasing algorithms — or does it require structural change in how data is collected, who builds AI systems, and what social conditions AI systems operate within?

Arguments for the Technical Fix

Argument 1 — Measurability: Fairness metrics make discrimination measurable in ways that subjective assessments do not. Measurement is a precondition for improvement. The technical approach provides tools for accountability.

Argument 2 — Tractability: Structural social change is slow, uncertain, and politically contested. Fairness-aware machine learning can be deployed now, with measurable impact on specific systems. Perfect should not be the enemy of good.

Argument 3 — Scale advantage: Technical solutions scale in ways that structural solutions do not. A fairness constraint applied to a widely deployed algorithm improves fairness for every person processed by that algorithm simultaneously.

Argument 4 — Organizational leverage: Technical teams can implement fairness metrics within existing organizational structures. Structural change requires political coalition-building outside the organization. Technical tools are available within the current system.

Arguments for the Structural Critique

Argument 1 — Garbage in, garbage out: Fairness metrics applied to biased training data, biased features, and biased labels cannot eliminate underlying discrimination. The technical tools assume the data is a neutral representation of the world; the critique is that it is not.

Argument 2 — Metric choice as politics: Choosing which fairness metric to optimize is itself a political decision about who bears the cost of error. Technical tools don't resolve this decision — they can only implement whatever decision is made. The political choice precedes the technical solution.

Argument 3 — Whack-a-mole: Satisfying a fairness metric on a benchmark dataset does not guarantee fair behavior in deployment. Models are optimized to pass the tests they are given. This produces Goodhart's Law problems: when the measure becomes a target, it ceases to be a good measure.

Argument 4 — Root cause: Algorithmic discrimination is a symptom of structural inequality, not its cause. Algorithms that discriminate in hiring do so because the labor market is structured to advantage some and disadvantage others. Fixing the algorithm doesn't change the labor market.

Argument 5 — Who is at the table: Fairness-aware ML is developed overwhelmingly by researchers from privileged backgrounds at elite institutions. The communities most harmed by algorithmic discrimination have minimal representation in the technical community that defines "fairness." The structural critique includes a critique of the social structure of AI research.

Key Empirical Disagreements

Do fairness constraints applied in development actually reduce disparate impact in deployment?
What is the size of residual discrimination in AI systems after technical debiasing?
What is the evidence that structural change (e.g., more representative training data, more diverse development teams) reduces algorithmic discrimination?

Key Value Disagreements

Should we accept incremental technical improvement while working toward structural change, or does incremental improvement legitimate unjust systems?
Is fairness within a given AI system the right unit of analysis, or should we focus on whether the AI system as a whole (including its social context) promotes justice?

What Would Resolve This Debate?

Longitudinal data on whether technically debiased systems show sustained reductions in real-world disparate impact would address the empirical core. The normative question of whether technical fixes are appropriate interim measures or problematic legitimations of unjust systems resists empirical resolution.

Debate 4: Should AI Developers Be Held Strictly Liable for AI Harms?

The Question

When an AI system causes harm — a false arrest from facial recognition, a denied loan from a biased model, a medical error from an AI diagnosis tool — should the developer of the AI be held strictly liable (liable regardless of fault or care taken) or only negligently liable (liable only if they failed to exercise reasonable care)?

Arguments for Strict Liability

Argument 1 — Internalization of costs: Developers are currently able to profit from AI systems while externalizing the costs of harms to users, communities, and society. Strict liability forces developers to internalize these costs, improving incentives for safety.

Argument 2 — Asymmetric information: Developers know more about the risks of their systems than users or regulators. Strict liability creates the correct incentives even when regulators cannot assess risk independently.

Argument 3 — Product liability precedent: Products liability law holds manufacturers strictly liable for defective products without requiring proof of negligence. AI systems are products; the same logic applies.

Argument 4 — Practical access to justice: Proving negligence in AI cases requires expert testimony about how the system works, what care was reasonable, and what errors were made. This is prohibitively expensive for individual plaintiffs. Strict liability lowers the barrier to recovery.

Arguments Against Strict Liability

Argument 1 — Innovation disincentive: Strict liability for harms that cannot be entirely eliminated will cause AI developers to avoid high-risk applications entirely — including beneficial ones like medical diagnosis and autonomous vehicles. This produces underinvestment in socially valuable AI.

Argument 2 — Proportionality: If a developer took all reasonable precautions and harm still occurred, strict liability treats a responsible developer the same as a negligent one. This is both unjust and removes the incentive to take precautions.

Argument 3 — Causal complexity: AI harms often involve multiple parties — developer, deployer, user, training data providers. Strict liability creates allocation problems when cause is distributed across actors.

Argument 4 — Insurance availability: Strict liability for AI requires actuarially sound insurance markets. If AI risks are too novel for insurers to price, strict liability may make important AI applications uninsurable and therefore impossible.

Key Empirical Disagreements

How elastic is AI development to expected liability costs? Would strict liability substantially reduce beneficial AI investment?
What are the actual rates of compensable harm from current AI systems?
How effectively does the negligence standard deter AI developer carelessness?

Key Value Disagreements

Should the law prioritize compensation for victims or incentives for developers?
Is the analogy to products liability correct, or do AI systems have distinctive features that warrant a different framework?
How should liability be allocated across the value chain (developer, deployer, user)?

Debate 5: Is There a Meaningful Right to Explanation for AI Decisions?

The Question

When an AI system makes a decision affecting a person — denying a loan, flagging a résumé, assessing recidivism risk — does that person have a meaningful right to an explanation of why?

The Goodman-Wachter Debate

The key academic dispute is between Bryce Goodman and Seth Flaxman (2017), who argue that GDPR Article 22 creates a right to explanation, and Sandra Wachter, Brent Mittelstadt, and Luciano Floridi (2017), who argue no such right exists in GDPR and that even if it did, the existing XAI methods cannot fulfill it.

Arguments That a Right to Explanation Is Meaningful and Achievable

Explanation is necessary for effective remediation: if you cannot understand why you were denied, you cannot correct the situation.
Post-hoc explanation tools (LIME, SHAP) can provide actionable information about feature influence.
Companies already provide explanation in other contexts (adverse action notices in credit); AI need not be different.
Even imperfect explanations improve accountability over no explanation.

Arguments That a Right to Explanation Is Either Illusory or Impossible

Wachter et al.'s legal argument: GDPR Article 22 provides a right to human review, not to an explanation. The "right to explanation" is a misreading of EU law.
Technical impossibility argument: For complex neural networks, there may be no explanation that is simultaneously faithful to the model, understandable to the human, and actionable — these three requirements cannot all be satisfied simultaneously.
Gaming argument: If people know what factors explain decisions, they will game those factors rather than address the underlying issue.
False certainty argument: Providing an explanation implies the model is more interpretable than it is, creating false confidence in systems that are opaque even to their developers.

Key Empirical Disagreements

Do existing XAI methods (LIME, SHAP) provide faithful explanations, or do they produce plausible-looking but inaccurate rationalizations?
Do adverse action notices (the credit industry's existing explanation mechanism) actually help people improve their situations?
Do explanations enable gaming?

Key Value Disagreements

Is explanation valuable instrumentally (it helps people improve their situations) or intrinsically (people have a right to know why they are treated as they are, independent of whether it helps)?
Should the law require explanations that are technically accurate but incomprehensible, or practically useful but technically simplified?

Debate 6: Should Facial Recognition in Public Spaces Be Banned?

The Question

Should the use of facial recognition technology in publicly accessible spaces — by police, by governments, or by private actors — be prohibited?

Arguments for Prohibition

Documented accuracy disparities: NIST FRVT documents substantial error rate disparities by race and gender. These disparities have already caused wrongful arrests.
Chilling effect on association: Mass facial recognition in public spaces suppresses political activity, protest, and religious observation by creating a permanent record of who was where.
Normalization of surveillance: Even accurate facial recognition enables social control and concentration of power in ways incompatible with democratic values.
No consent is possible: There is no mechanism for individuals to consent to or opt out of facial recognition in genuinely public spaces.
Absence of demonstrated necessity: There is limited evidence that facial recognition improves public safety outcomes at levels that justify its civil liberties costs.

Arguments Against Prohibition

Security benefits: Facial recognition assists in locating missing persons, identifying terrorist suspects, and solving violent crimes that would otherwise go unsolved.
The algorithm is not the policy: The question is not whether facial recognition is used but how — with what oversight, what accuracy thresholds, what confirmation requirements, what transparency obligations. Banning the technology forecloses better-regulated uses.
Accuracy is improving: NIST documents significant accuracy improvements across the FRVT testing periods. A technology that is unacceptably inaccurate today may be adequately accurate with appropriate oversight in five years.
Alternatives are worse: Manual identification through photos or video may be less accurate and less transparent than algorithmic identification.
Democratic legitimacy: Questions about surveillance should be decided through democratic processes, not preemptively resolved by prohibitions that eliminate public choice.

Key Empirical Disagreements

At what accuracy threshold, if any, does facial recognition become acceptable for law enforcement use?
What is the actual crime-solving value of facial recognition, and is it achievable through less privacy-invasive means?
Do chilling effects on political activity measurably occur in places where facial recognition is deployed?

Key Value Disagreements

How should privacy and security be traded off? Who gets to make that trade?
Is the chill on political association an acceptable cost of a crime-solving tool, or an independently unacceptable harm?
Should accuracy requirements be absolute (the technology must be equally accurate across demographic groups before deployment) or relative (it must be more accurate than the alternatives)?

Debate 7: Can AI Ethics Principles Documents Achieve Genuine Change?

The Question

The proliferation of AI ethics principles documents — published by companies, governments, and civil society organizations — is the dominant form of AI ethics governance. Can voluntary principles achieve genuine improvements in AI practice, or does the principles movement primarily serve to legitimate otherwise unchanged behavior?

Arguments That Principles Can Work

Standard-setting effect: Widely adopted principles establish expectations that can be enforced through soft mechanisms — reputational pressure, investor scrutiny, employee resistance — even without legal force.
Coordination function: Principles create shared vocabulary and frameworks that enable productive discussion across organizations and sectors.
Legalization pathway: Voluntary principles often precede binding regulation and establish the normative baseline that regulations codify. The OECD AI Principles directly informed the EU AI Act.
Internal leverage: Published principles give internal AI ethics advocates leverage to push back against business decisions that violate them. Commitment creates accountability.
Iteration and improvement: Principles can be updated as understanding develops, faster than law can respond.

Arguments for the Ethics Washing Critique

No enforcement: Principles without enforcement mechanisms are not binding. Companies can publish principles they do not implement without consequence.
Conflict of interest in design: AI ethics principles are typically designed by the same organizations that benefit from minimal AI regulation. This shapes what principles are included and how they are framed.
Window dressing: Published research has documented the gap between AI ethics principles and actual organizational practice. Meredith Whittaker and the AI Now Institute have catalogued cases where companies published principles while lobbying against equivalent regulatory requirements.
Legitimacy effect: Principles documents may reduce pressure for binding regulation by creating the impression that the industry is self-governing. If so, they produce a net negative governance outcome.
Missing accountability: The most important elements of accountability — independent auditing, transparency of training data and model behavior, redress for harmed individuals — are consistently absent from voluntary principles.

Key Empirical Disagreements

Has the proliferation of AI ethics principles been accompanied by measurable improvements in AI system fairness or safety?
Is there evidence that companies with published principles perform differently from companies without them?
Have voluntary principles delayed binding regulation, or accelerated it?

Key Value Disagreements

Is the relevant comparison for voluntary principles a world with binding regulation or a world with no governance at all?
How much weight should be given to legitimacy concerns (that principles washing makes regulation harder) vs. immediate impact concerns (that some governance is better than none)?

Debate 8: Does AI Pose an Existential Risk That Should Dominate AI Ethics Discourse?

The Question

Some AI researchers and philosophers argue that advanced AI poses risks to human existence or autonomy at civilizational scale — and that these risks should be the dominant focus of AI ethics. Others argue that near-term, concrete harms to real people should dominate the agenda.

Arguments for the Long-Termist/Existential Risk Focus

Magnitude argument: Even a small probability of existential harm, multiplied by the magnitude of harm (the loss of humanity's entire future), produces expected harm that dominates any finite set of near-term harms.
Lead time argument: Preparing for existential risk requires decades of research. If we wait until existential risk is imminent, it is too late to prepare.
Neglectedness: Near-term AI harms receive substantial attention from researchers, journalists, and regulators. Existential risks are neglected. The marginal value of attention is higher where the discourse is thinner.
Technical tractability: Alignment research — ensuring AI systems pursue intended goals — addresses the root cause of both near-term and long-term AI harms.

Arguments for the Near-Term Harm Focus

Speculative vs. concrete: Existential risk from AI depends on empirical claims — about the likelihood of artificial general intelligence, about the difficulty of alignment, about the speed of capability development — that are deeply contested. Near-term harms are documented and measurable.
Opportunity cost: Attention to speculative existential risk displaces attention from documented harms to actual people today — wrongful arrests, denied benefits, discriminatory lending, surveillance of dissidents.
Political economy: The existential risk framing is promoted primarily by organizations funded by AI technology billionaires, who also benefit from reduced near-term regulation. Skepticism about the political function of this framing is warranted.
Who bears the costs: Near-term AI harms fall disproportionately on already-marginalized communities. The communities focused on in long-termist discourse (future humans, potential AIs) have no current political voice. Prioritizing their interests over those of present, marginalized communities raises justice concerns.
Prediction is hard: The history of AI contains decades of overconfident predictions about the timeline to transformative AI capability. Organizational planning based on speculative long-term predictions has poor track record.

Key Empirical Disagreements

What is the probability of artificial general intelligence capable of posing existential risk within 50 years?
Does technical alignment research provide genuine risk reduction, or is it solutions to hypothetical future problems?
Is attention to existential risk actually displacing attention to near-term harms, or are these agendas complementary?

Key Value Disagreements

How should present harms to identifiable individuals be weighted against speculative future harms to hypothetical individuals?
Is utilitarian aggregation (large probability × large harm = large expected harm) the right framework for evaluating risk?
Whose interests — present, future, human, potentially non-human — should be represented in AI ethics discourse?

What Would Resolve This Debate?

Unlike most empirical disputes, this debate cannot be resolved by currently available evidence because the key claims are about future capabilities and risks. It is partly a philosophical dispute about methodology (how to reason under radical uncertainty) and partly a political dispute about whose interests the AI ethics field should serve.

These argument maps are tools for clearer thinking, not conclusions. The best AI ethics practice requires holding multiple perspectives simultaneously: taking near-term harms seriously as documented facts while maintaining appropriate concern for longer-term risks; advocating for technical improvements while insisting on structural accountability; supporting voluntary governance initiatives while demanding binding regulation. The ability to hold these tensions is the hallmark of sophisticated AI ethics reasoning.