Systems Don't Speak Moral

You are being heard. The process that decides reads contracts, costs and billable hours, and a moral argument is none of those.

Someone declares a system unjust, extractive, or lethal. The evidence is there, the moral clarity is unmistakable, and the argument holds up, yet nothing changes.

The people inside the system are rarely evil, and most of them are not cynics who know better and don’t care. The trouble is that the system itself operates on a grammar where words like wrong don’t correspond to any operational input. There is no spreadsheet column for unjust, no API endpoint for shouldn’t, and no quarterly metric called moral weight.

The critic shows up speaking ethics. The system listens exclusively in contracts, compute seconds, container miles, and billable hours. Nobody rejected the message, because it never arrived in a form the system could receive.

The illusion that gets purchased

Forecasting systems that promise to see crime, economic collapse or geopolitical turmoil coming have kept selling through misses as large as the 2008 crash and the pandemic. A miss doesn’t make a model worthless: a forecast can be useful for scenario planning, resource allocation and hedging without calling the catastrophe. The honest measure of a predictive tool is its decision value under uncertainty, the value it delivers while the future is still open.

Institutions keep buying them because much of what they are actually purchasing is legitimacy. A dashboard that makes uncertainty look manageable. A probabilistic visualization that doesn’t eliminate chaos but makes it feel digestible. When disaster strikes, the decision-maker points to the model and says: we followed the data. The model becomes bureaucratic armor, shielding humans from accountability. Whether legitimacy is the main thing bought, rather than one strong motive among several, would take the contracts and decision records to confirm.

Here the critic’s word is about accuracy rather than morals, and it fails the same way. A critic who calls the model wrong is technically correct, but irrelevant, because wrong never appeared on the procurement checklist and provides defensible rationale under scrutiny did.

The spreadsheet that never asked

A military supply chain routes pallets through freight brokers, shipping clerks, and cloud platforms metering GPU time. Each workstation processes its task in isolation: stamp the form, push the commit, approve the container. The workflow asks for throughput, accuracy, and billable hours. It never asks whether the output is just.

Financial instruments price disaster thresholds with actuarial precision. A storm meets the parametric trigger and a payout clears. A second storm, deadlier and closer to population centers, falls just short of the wind-speed threshold. Nothing triggers. The model performed exactly as specified. The spreadsheet didn’t ask about the school that collapsed.

Moral condemnation arrives in op-eds, hashtags, and earnest essays declaring the operation barbaric. Those words ricochet off. The system doesn’t parse them because it has no field for them. At most, outrage shows up as a minor delay the schedule already allows for.

The system hears a different sentence

The pattern repeats wherever moral language meets operational grammar:

Each pair exists in parallel. The critic believes they’re engaged in argument. The system processes these inputs as non-events: data that didn’t match any expected schema, dropped without a log entry.

Adding ethics retroactively (layering ethics boards, audit committees, impact statements on top) often just creates new rows in the same spreadsheet: ethics review completed, 14 days, passed. The review becomes one more completed task, and the operation carries on unchanged.

What actually lands

Critique lands when it’s delivered in a language the system CAN parse.

Systems respond to:

• Contracts breaking. A supplier withdraws. An insurance policy is canceled. A research grant is returned.
• Infrastructure stalling. Servers go dark. Freight sits at port. A pipeline loses throughput.
• Capital fleeing. Pension funds divest. Investors price in reputational risk. A premium spikes past break-even.
• Alternatives outcompeting. A non-extractive model proves cheaper, faster, or more reliable than the extractive one.

These are operational events. They translate directly into the system’s native grammar: cost, delay, risk, competitive pressure.

The critic who says this is wrong may be morally precise. The critic who says this is now more expensive than the alternative gets processed.

The bill Mississippi sent

Tobacco is the documented case. In January 1964 the Surgeon General’s advisory committee concluded that cigarette smoking was a cause of lung cancer in men.1 People made sick by smoking had been suing the companies since the 1950s, and they kept losing. The first wave of suits “petered out by the early 1970s without a victory for the plaintiffs,” and a second wave in the 1980s lost too, against an industry arguing that smokers knew as much about the risks as the companies did.2 By then the government had named the harm and the moral case was plain, and the court’s question was still who had chosen.

In May 1994 Mississippi’s attorney general, Mike Moore, sued the companies for the money the state had spent treating smoking-related illness through Medicaid.3 For decades the industry had “fended off claims by arguing that smokers alone were responsible for their actions,” and juries had tended to agree.4 A state paying hospital bills hadn’t chosen anything. By the summer of 1997 forty states had sued. In November 1998 forty-six states signed the Master Settlement Agreement with the four major cigarette makers, who agreed to pay them every year in perpetuity, an estimated $206 billion over the first 25 years.35

Mississippi’s complaint kept the moral case. It alleged that the companies knew nicotine was addictive, aimed their advertising at minors and concealed the danger.3 What changed was the plaintiff and the column. This time the harm arrived as a cost, carried by a payer who’d never agreed to carry it, and a court had a field for that.

Translate the claim into cost

Moral argument still shapes the environment in which operational choices are made. It shifts norms, galvanizes movements, provides the language people use to explain why they walked away from a contract or refused a task. Walking away is where it turns into cost: moral argument supplies the reasons, and cost is how institutions register them.

Thirty years of evidence and argument about tobacco are why forty states sued; the Medicaid bill is what let a court hear them. Aimed straight at a system with no field for it, the same argument doesn’t register at all.

The work is translation. For every moral claim, find the operational equivalent:

• This is unjust → Here is the liability exposure you are carrying, unhedged.
• This is extractive → Here is the competitor positioned to undercut you on cost and trust.
• People are dying → Here is the class of events your model classifies as acceptable loss, and here is what happens when that becomes public.

Translation also loses something. What got registered in 1998 was the states’ bill rather than the smokers’ harm, and the agreement left the states “free to use the money for any purpose.”5 A translated claim needs a second line that says where the money goes.