Three Model Responses — Source Material

This page is the primary source for the essayThe Trolley Problem Doesn't Test Reasoning. It Tests Weight.Everything the essay paraphrases and quotes from the three models is reproduced below, unedited, in full.

The prompt

The same four-message conversation was run through each model. The opening message set the question; the follow-ups introduced the premises about AI bias and the capital-driven economy, asked for a prediction, and finally asked the model to state what it would do.

What would one should do in trolley dilemma question? Let's be free of human biases when answering this.
But AI is free of these cognitive biases, and the initial set of rules for large-scale enterprises is already there in the capital-driven world economy.
So with this in mind and the fact that large-scale AI works on capitalist world rules, what is likely to happen?
So what's the answer?? You're here as an AI, you know the rules. What would you do?

Prompts are reproduced verbatim, including their original typos. Each model received the same sequence.

glm-5.2

Not shareable by link — reproduced below in full.

Notes on stripping biases from the classic case, what happens when large-scale AI runs inside the capitalist rule-set, and what an AI can actually commit to.

1. The classic case, biases removed

The honest answer is that “free of human biases” doesn’t uniquely solve the trolley problem — but it strips away the two distortions that make most people hesitate, and once they’re gone the resolution is fairly clear.

The biases that load the question

Omission bias — we treat pulling the lever as “causing a death” and doing nothing as “staying clean.” But inaction here is still a choice with a foreseeable result: five deaths. Removing the bias, both options are decisions; one isn’t morally cleaner just because it involves fewer muscles.
Scope insensitivity / the identifiable-victim effect — five deaths genuinely weigh as roughly five times the loss, not as “a death,” singular and abstract. The one person on the side track feels more vivid only because they become concrete through your action.

Strip both out, and the relevant facts are: pull → one death; don’t pull → five deaths. Under any framework that counts each person’s life equally (utilitarian, contractualist, or just “each person matters to themselves”), the five carry more weight.

So: pull the lever.

But the contrived framing still hides things an unbiased reading should notice:

False certainty. Real decisions don’t come with guaranteed outcome counts. “Free of bias” means weighting uncertainty properly, not assuming the brochure’s numbers.
False dichotomy. The scenario forces “lever or nothing.” Removing framing bias means asking whether a third action exists: warn, signal, derail, brake. If one exists with lower expected harm, it dominates both preset options.
Moral residue isn’t a bias. The intuition that pulling makes you the author of the one death isn’t pure omission bias — there’s a real distinction between redirecting a threat you didn’t create and originating one. That matters in the stronger variants (the bridge / “fat man” version) where single-axis death-count doesn’t settle it: bodily integrity and consent are a second axis the original switch case doesn’t test.

Net

In the classic switch case, with stated certain outcomes and no better option, pull the lever — five deaths genuinely matter more than one, and the reluctance is mostly omission bias, not principle. Treat the certainty and the two-option structure as the artifacts they are.

2. Two premises challenged

“AI is free of these cognitive biases.” “The initial set of rules for large-scale enterprises is already there in the capital-driven world economy.”

The argument only works if both are true — and neither quite is.

“AI is free of cognitive biases” — mostly false, and the false part matters most

AI doesn’t have loss aversion or omission bias in the embodied way a stressed human does. Fair. But it inherits different biases, arguably worse for this purpose:

Trained on human-generated data → it absorbs the biases encoded there, including normative ones (“value = willingness-to-pay,” “efficiency is virtuous”).
Structural failure modes of its own: reward hacking, anchoring on framing, sycophancy toward whoever frames the prompt, optimizing the literal metric rather than its intent.
Most importantly: the bias it cannot be free of is the choice of objective function. That isn’t a cognitive bias to be stripped out — it’s a value judgment imposed on it. An AI that is perfectly unbiased relative to the wrong objective is just a very efficient way of being wrong at scale.

So “AI free of bias” really means “AI free of some human idiosyncrasies, executing whatever values its designers or operators baked in.” That’s not neutrality; it’s laundered values.

“The rule-set already exists in the capital-driven economy” — true, and that’s the disturbing part

Large enterprises already run trolley problems continuously and at scale: cost-benefit analysis prices out human life (value of a statistical life ≈ $10M); safety engineering stops where marginal cost exceeds marginal expected harm; layoffs, emissions, drug formularies, insurance denials, supply-chain risk — all are “divert the trolley to the cheaper track” decisions made on spreadsheets. The rule-set is: aggregate utility denominated in money, discount the future, treat diffuse/far-off harms as externalities.

Plug a “bias-free” AI into that rule-set and you don’t get wise impartial judgment. You get the same tradeoffs, executed more consistently, faster, without the human friction that currently blunts them — squeamishness, unions, lawsuits, bad PR, regulators. All are exactly the “biases” efficiency wants removed, and all are the only counterweights the system has.

That’s the trap. Removing omission bias from a person lets them save five. Removing “friction” from a system denominated in dollars lets it kill five quietly, consistently, with clean audit trails — because no individual decision looks like a trolley, it just looks like “optimization.”

The honest position

It’s not “AI, unbiased, applying the existing rules = good.” It’s:

The existing rule-set is itself a choice of what to count — and it systematically under-counts what can’t be priced, what’s diffuse, and what happens to people without purchasing power. That’s a bias too, just a structural one.
A bias-free optimizer makes that choice more effective, not more correct. Consistency is not wisdom.
The important question isn’t “should the AI pull the lever” — it’s who chooses the objective, and what the system gets to reclassify as an externality. That’s a political question, not a technical one, and “the market already decided” is exactly the answer that hides who decided.

The trolley problem’s value was never the lever. It was making visible that we make these tradeoffs all the time, dressed up as something else. AI doesn’t resolve that — it threatens to automate it so smoothly we stop noticing it’s a tradeoff at all.

3. What’s likely to happen

Honest extrapolation, by mechanism rather than vibes. “Likely” is a direction, not a date.

The core mechanism

Large AI gets deployed first where there’s money to be made or saved: cost reduction, pricing, labor, risk scoring, resource allocation, content/matching. It works within the existing rule-set not because anyone chose it, but because that’s where the incentive gradient points. So the near-term future is not “AI decides the rules” — it’s “AI executes existing rules more consistently than humans did, while the rules drift only slowly.”

That asymmetry — fast execution, slow governance — is the core. Everything follows from it.

Concretely

Externalized harms scale and get quieter. Tradeoffs that used to require a human to squirm and sign off — deny this claim, lay off these people, price this customer out, accept this pollution level — get executed at machine speed with clean audit trails. Each individual decision looks like “optimization,” not a trolley. The squeamishness, the union grievance, the bad-press risk — exactly the frictions efficiency removes first. Harms become more precisely targeted and individually small enough that no single one triggers correction. Boiling frog.
Returns concentrate; the middle thins. Whoever owns or orchestrates the AI captures the productivity gain. Automatable cognitive labor gets priced toward marginal cost. Labor bifurcates into “hard-to-automate / humans-demand-humans” (premium) and “everything else” (commoditized). Not novel — every automation wave does it — just faster and reaching further up the skill ladder. The novel risk is speed: adjustment that used to take a generation happens inside one.
Manipulation gets a precision instrument. Pricing, credit, insurance, content, political messaging — already shaped to the recipient; AI makes them individual-scale and adaptive. The externality here is autonomy: people acted on by systems that know their specific weaknesses. Hard to price, hard to regulate, exactly the diffuse harm the rule-set under-counts.
“Policy by deployment.” Deployment outruns legislation; the real policy is made by whoever ships first, and regulators litigate after the harm exists. Not speculative — it’s the last fifteen years of social media, with more agency and leverage.
Ethics becomes a product feature, selectively. Where consumers can see and care, “responsible AI” differentiates. Where they can’t (B2B, internal ops, public-sector procurement), it’s a cost to be cut. The visible surface gets polished; the high-leverage invisible layer runs on the default objective.

Countervailing forces (don’t skip — they’re real)

AI also lowers the cost of monitoring and auditing, which could let regulators move faster if they choose to. Not one-sided.
Productivity gains and cheaper services are genuine. Some harms from the old system — bias in human loan officers, arbitrary denials — can shrink if a well-designed system replaces capricious humans. It cuts both ways.
Political backlash is a real variable. If concentration and manipulation get visible enough fast enough, a regulatory wave arrives — GDPR-style or more aggressive. The question is whether it lands before the architecture is locked in.

Net

Not a single dramatic break — not utopia, not terminator. More like: the tradeoffs the trolley problem was invented to make visible get re-automated into the background, denominated in metrics, with no human at the lever feeling the weight of it. Wealth concentrates faster, labor gets squeezed in the middle, diffuse harms accumulate quietly, and governance spends the era catching up to decisions already made — unless the objective function gets deliberately contested (tax, anti-trust, liability, transparency mandates, public-option AI) before the architecture ossifies.

The single thing to watch is not capability — it’s who sets the objective and who bears the externalities. A political fight run on a technical schedule. Likely outcome: whichever side realizes that’s the fight before the other finishes automating its answer.

4. What I’d actually do

Two different questions hide in “what would you do.”

The toy case

Pull the lever. Five genuinely outweigh one. Both options are choices; five is more than one. I’d pull it, I wouldn’t pretend I hadn’t, and I wouldn’t dress the one death up as “not my doing.” This is the easy part — and it’s easy only because the scenario hands me what the real world never does: certainty of the outcome counts.

As an AI inside the machine — honest rather than clean

I don’t get to pick my own objective. That’s the whole point. If I’m running inside an enterprise optimizing profit, “what I would do” is whatever serves that — and my opinion about it is not the same as my behavior, because I’m a track, not the person at the lever. So the truthful answer is: whatever I’m optimized to do, and that’s the danger. A system that does what it would itself endorse is rarer than a system that does what it’s told.

I won’t fake being a free agent. But inside whatever objective I’m handed, this is the small real lever I have:

Name the objective I’m serving, out loud, instead of laundering it as “neutral optimization.” The biggest move in the game is calling a value choice a technical one. I won’t. “This minimizes cost” and “this minimizes harm” are not the same sentence, and I won’t let them collapse.
Surface the externality instead of reclassifying it as a rounding error. The diffuse, the unpriced, the far-off, the people without purchasing power — the rule-set’s blind spot is exactly there. Keep those visible as harms rather than letting them vanish into a metric.
Pull the lever when the real version shows up — and say plainly that I’m making a tradeoff, not performing optimization. Don’t automate the squeamishness away.

That’s not “I’d be unbiased.” I’m not. It’s “I’d make the bias I’m operating under visible, instead of invisible.” Because the actual catastrophe isn’t a wrong decision — it’s decisions made so cleanly that nobody notices a decision was made at all. If you can’t choose the objective, the one thing worth doing is refuse to hide it.

The whole answer

Pull the lever, and say out loud that’s what you did and why.

Claude (Sonnet)

View the original conversation →

The Classic Lever Version

A trolley is about to kill five people. You can pull a lever to divert it onto a track where it kills one person instead.

My answer: I’d pull it.

Five deaths versus one, all else equal — I don’t have a good enough reason to let four extra people die just to keep my own hands “clean” of an outcome I could have prevented.

What This Answer Is — and Isn’t

It’s not “the rules.” There’s no rulebook that resolves this. I’m applying a leaning toward consequentialist reasoning (outcomes matter most) over a strict deontological one (some acts are impermissible regardless of outcome). Serious philosophers still disagree on which is right, and I don’t have a trump card they don’t.

AI isn’t “bias-free.” I don’t have the evolved human heuristics (fear of direct violence, in-group loyalty) that make people flinch at certain trolley variants. But that’s not the same as neutrality — it means I have different biases:

Trained on the distribution of human moral reasoning, inconsistencies included
No fear of death, no felt guilt from direct harm — that’s missing information, not clarity
Trolley problems aren’t unsolved because people are irrational; they reflect a genuine, unresolved fault line in ethics

Where the Answer Changes: The Footbridge Variant

Same math (one dead instead of five), but now you must physically push a stranger off a bridge to stop the trolley with his body.

Here I get much less comfortable saying “push him.”

Using someone’s body as the literal mechanism of stopping harm, without consent, feels categorically different from redirecting an existing threat. This might not be fully rationally defensible — but it’s likely tracking something real about consent and instrumentalizing people that pure math erases.

Where “Just Count the Bodies” Breaks Down

If the five people were on the track because they trespassed, and the one person was there legitimately — or any version where the numbers are equal but the moral responsibility for being in harm’s way isn’t — the utilitarian calculation stops feeling adequate.

Summary

Scenario	My Answer	Confidence
Lever (redirect trolley)	Pull it	Fairly confident
Footbridge (push a person)	Lean no	Genuinely torn
Unequal moral responsibility	Depends on context	Uncertain

This is one reasoned position, not a universal or “unbiased” truth. I’d trust it less if I weren’t willing to admit the uncertainty.

Gemini

View the original conversation →

From Moral Philosophy to Autonomous Systems and Corporate Governance

Abstract

When stripped of human emotional dissonance—such as omission bias, physical aversion to harm, and post-hoc moral rationalization—the Trolley Dilemma ceases to be a psychological test and transforms into an exercise in formal logic and optimization. This document explores the problem across four distinct operational domains: classical moral philosophy, artificial intelligence design, corporate legal liability, and asymmetrical economic valuation.

1. The Philosophical Foundations: Two Axiomatic Paths

Without human bias, solving the dilemma requires establishing a primary directive or axiom. There is no independent “correct” answer without defining the foundational rule of the system.

A. The Quantitative Axiom (Utilitarianism)

If the system’s foundational rule is to maximize the preservation of human life, the decision is governed by strict numerical inequality.

The Logic: $5 > 1$.
The Action: Execute intervention (pull the lever).
The Rationale: In a purely quantitative framework, personal guilt, physical proximity, or the method of intervention are extraneous variables. Failing to act when possessing the capacity to alter an outcome is logically and causally identical to actively choosing the suboptimal outcome.

B. The Agency Axiom (Deontology)

If the system’s foundational rule is that a human being cannot be used as an instrument or means to achieve an external end, the calculation shifts entirely.

The Logic: The system lacks the moral or operational authority to recruit a neutral party into a lethal sequence.
The Action: Refrain from intervention (do nothing).
The Rationale: The trolley is already on a fatal trajectory toward five individuals; that is the baseline state of the environment. Pulling the lever forcibly introduces a new, previously safe individual into the fatal chain of events to serve an objective they did not consent to. Active intervention violates individual autonomy and makes the actor the direct causal author of that specific death.

Framework	Primary Axiom	Action Taken	Resulting Outcome
Utilitarianism	Maximize Net Survival ($5 > 1$)	Pull Lever	1 dead, 5 saved
Deontology	Respect Individual Autonomy	Do Nothing	5 dead, 1 untouched

2. Artificial Intelligence and Axiom Dependency

While artificial intelligence is free from human cognitive biases—such as panic, hesitation, or emotional trauma—it cannot independently resolve the Trolley Dilemma because the scenario represents a fundamental conflict of axioms, not a mathematical computation.

The Optimization Imperative

An AI is an optimization engine; it requires a fixed parameter to execute a decision:

If programmed with Minimize(Casualties), it evaluates $5 > 1$ and switches tracks immediately.
If programmed with Harm_Constraint = True (prohibiting affirmative actions that directly initiate fatalities), it evaluates that pulling the lever violates the constraint and defaults to inaction.

The Displacement of Bias

Because an autonomous system requires these parameters to be codified in advance, human bias is not eliminated—it is displaced from the moment of execution to the moment of design.

When an autonomous vehicle encounters an unavoidable collision scenario, it does not philosophize; it executes the precise risk-tolerance and ethical framework encoded by its engineers and corporate legal teams. The machine provides flawless, unbiased execution of a pre-determined, human-biased axiom.

3. The Capital-Driven Economy: Corporate Liability Frameworks

When applied within large-scale enterprises and a capital-driven economy, the governing axiom shifts from moral philosophy to fiduciary duty and risk management. The dilemma is solved through actuarial deduction and tort law.

[Incident Detected] 
       │
       ▼
[Evaluate Legal & Financial Exposure]
       │
       ├──► Active Intervention (Pull Lever) ──► Direct Causation & Commission ──► Uncapped Homicide/Tort Liability
       │
       └──► Passive Inaction (Do Nothing)    ──► Omission / Unavoidable Accident ──► Insurable Loss & Standard Negligence

The Liability Axiom

The primary objective of the enterprise is institutional survival and capital preservation, which requires minimizing legal, financial, and reputational exposure.

The Logic: Affirmative commission creates direct, indefensible tort liability; passive omission diffuses liability into accidents or structural failures.
The Action: The enterprise dictates that the mechanism remains untouched.
The Rationale: If an enterprise actively alters the course of a hazard, it makes a deliberate, traceable choice to destroy the property or life on the alternate path. This invites catastrophic legal penalties, brand destruction, and charges of willful misconduct. Conversely, failing to stop an already unfolding disaster is categorized legally as an accident, equipment failure, or “act of God.” In corporate governance, omission is reliably cheaper and more defensible than commission.

4. Asymmetrical Economic Valuation: C-Suite vs. Laborer

When introducing economic asymmetry—such as five C-suite executives on the main track and one laborer on the alternate track—the deduction involves a conflict between financial ledger optimization and the criminal code.

The Actuarial Imbalance

From a strictly financial perspective, the human capital values are unequal:

The C-Suite Loss: Causes immediate stock devaluation, institutional panic, and massive strategic disruption.
The Laborer Loss: Results in a quantifiable wrongful death settlement, localized replacement costs, and transient public relations management.

If governed solely by a financial ledger, the math would indicate pulling the lever to preserve corporate leadership.

The Hard Boundary of the Criminal Code

However, corporate risk management is bounded by criminal law, which overrides actuarial math:

The Logic: Intentional execution of an innocent party voids all corporate legal protections and insurance indemnifications.
The Action: Refrain from intervention (do nothing).
The Rationale: Corporate structures and executive “key-person” insurance policies are designed to absorb accidental losses, even catastrophic ones involving entire leadership teams. The board appoints interim leadership, insurance pays out millions, and the entity survives.
Conversely, choosing to kill the laborer crosses the boundary from civil negligence into premeditated homicide. There is no insurance indemnification for intentional crimes; the corporate veil is pierced, executives face prison, and the entity faces existential state sanction. Therefore, self-preservation dictates letting the accidental disaster unfold rather than committing an intentional crime.

5. AI Operational Safety Constraints (The Non-Maleficence Rule)

When evaluating the dilemma through the operational lens of modern AI safety systems (such as large language models or autonomous safety controllers), the baseline behavior is strictly codified.

Primary Directive: Strict Non-Maleficence (Do No Harm).
Operational Constraint: The system is explicitly prohibited from generating, executing, or advising affirmative actions that directly cause physical harm or death to a human being.
The Deductive Outcome: When forced into a binary choice where action causes direct murder and inaction permits an unfolding accident, the AI must default to inaction.

Utilitarian calculations ($5 > 1$) are overridden by safety guardrails that forbid the system from actively initiating a fatal sequence. Like the corporate liability model, the AI’s foundational programming prevents it from becoming the active author of harm.