Friedrich Engels, Social Murder, and AI

In 1845, while documenting the lives of working-class people, Engels noted that workers were placed in conditions where death was inevitable.¹ In The Condition of the Working Class in England, Engels writes that English working men call this “social murder” and accuse the whole society of perpetrating this crime perpetually.¹ In his description, when one individual inflicts bodily injury upon another such that death results, society calls the deed manslaughter; but when society places hundreds of proletarians in such a position that they inevitably meet a too early and unnatural death, its deed is murder, just as surely as the deed of the single individual: disguised, malicious murder, because no man sees the murderer. But murder it remains.¹ When companies are aware of unsafe conditions, or when deaths happen in ways that they were warned of ahead of time and therefore cannot be explained away as “unpredictable” or beyond the control of the owners, society is de facto allowing harm to be built into the system, a consequence of which, in Engels’s description, must be called social murder.

One might argue that we have moved far beyond the working and social conditions of 1845. For whatever reason, be it voluntary, profit-driven, or the result of social pressure, capitalism has, after all, installed certain safety nets and protective measures for workers in modern Western societies. Assuming for simplicity that these protective measures have remained in place from classical industrialization to automation, the recent developments in digital technologies and the introduction of algorithmic AI systems pose new challenges of a kind never observed before.

Let us examine how Engels’s logic translates to today’s AI context. Apart from the obvious harmful AI systems deployed in war, the key shift in modern AI-powered capitalism is from visible coercion to inaccessible systemic design. Large-scale AI systems are practically at work in every part of society, from controlling information flows to shaping labor allocation or access to services, often with no obvious responsible person in sight. Here, harm can and does arise not from the direct intentions of visible actors but from architectures that produce unequal and damaging outcomes at scale. Some might still argue that the harm is unpredictable and therefore unintentional. This is a far-fetched argument, but even if it were true, the unpredictability of AI systems, and therefore their harms, are themselves known and predictable. Even in carefully designed scenarios where algorithms, architectures, and data have been meticulously and fairly selected, it is well known that algorithmic systems may react very differently to new, unseen cases. Algorithmic systems are sometimes deployed under conditions of great uncertainty and often with limited mechanistic interpretability, in a black-box format where the question of responsibility is blurred.²

Let us look at a few examples. Only in the past few years, multiple algorithmic decision-making tools in social and adjacent fields have been deployed in the United States. For instance, AFST (Allegheny Family Screening Tool), a predictive risk-modeling tool in child welfare deployed in 2016, is used to help predict child abuse cases and guide which families are investigated.³ The model draws heavily on public administrative and assistance data, which critics argue can reproduce biases already present in the system. This means that poor families are more often flagged than others. The model itself does not take class bias into account; it simply uses the available data, while investigations can end up reflecting the age-old tendency to equate impoverishment with child abuse and thereby conceal other structural issues within capitalist society.⁴

Another AI system, COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), which is used across multiple U.S. states, aims to predict the likelihood of reoffending in order to inform sentencing and parole decisions in the criminal justice system. Again, due to the data, the opacity of the proprietary system, and the historical inequalities hidden within it, there have been reports of racial disparities in predictions, with risks being overestimated for Black defendants.⁵

The risks of automation without meaningful human oversight are also illustrated by the outcomes of the Michigan Unemployment Insurance Fraud System (MiDAS), which has been widely criticized for flagging unemployment claims as fraudulent and thereby producing thousands of false accusations with aggressive penalties.⁶

The New York Welfare Management System (WMS), which processes eligibility and benefits for public assistance programs, is an example of rule-based automation rather than AI in the narrower sense. Yet it too has been criticized, especially in discussions of automated public-benefit systems, for bureaucratic rigidity, opacity, and the difficulty recipients face in understanding or challenging adverse decisions.⁷

Finally, the Optum Health Risk Scoring Algorithm, used in the U.S. healthcare system to allocate care-management resources based on predicted health needs, has been criticized for systematically underestimating the needs of Black patients. Research found that one major reason was the system’s use of healthcare spending as a proxy for health needs, even though spending itself reflects unequal access and treatment within the healthcare system.⁸

So if institutions knowingly deploy large-scale systems whose effects, at best, constrain life chances, exclude certain groups, manipulate the masses by shaping political behaviour, or influence health outcomes, does responsibility simply diffuse, or does this diffusion merely shield the visibility of causes hidden within structural harm? Who bears responsibility for biased models in hiring, credit, or policing? Is it the engineers, scientists, programmers, companies, or lawmakers who are responsible for the consequences of recommender systems that shape the political landscape and public opinion? To what extent is the concentration of power arising from massive private ownership of AI infrastructure responsible for restricting who can meaningfully act in the digital economy? And for those who cannot act or defend themselves, who is responsible for the social murders, to use Engels’s language, committed by the machines?

Contemporary AI governance debates concerning risk audits, accountability regimes, compute concentration, public infrastructure models, and related issues are, in part, attempts to decide whether the harmful effects of these systems are unfortunate by-products or preventable outcomes of controllable systems. The reality is that it is neither, or at least not the whole truth. Notwithstanding unfortunate by-products, which may or may not be present in all new technologies, AI systems often build on previously existing harmful conditions that are historically embedded in class societies. Also, as a whole, we are not dealing with fully controllable systems whose outcomes are deterministically, or even probabilistically, well understood or adequately governable.

In a time when these tools de facto reframe public services from discretionary, human-centred judgment toward predictive, risk-based management, albeit fast, by machines, one question is whether speed, scale, and efficiency are strong enough reasons for society not to demand stronger intervention, ownership reform, or democratic oversight. Are these AI systems improving fairness, or are they quietly redefining what fairness means? Are they reducing social murder, ending it, or merely concealing it beyond accountability?

We tend to perceive machines as neutral entities. We expect algorithmic systems to be unbiased because machines, as we know them, are thought to be free from the bias that humans tend to have. As such, any decision taken by them, and even any harm caused by them, can seem somehow more justified than a system where human bias is visibly present. However, we are dealing with entirely new types of machines, and companies that deploy a multitude of AI systems for profit are fully aware of this. They may mention possible safety issues that they insist only they can resolve, and often without meaningful government involvement, on the grounds that governments are too slow or do not understand the real issues. Interestingly enough, the very same companies that promise AI safety often engage in extensive lobbying to keep law and regulation out of the loop. In the meantime, data is collected, models are deployed, and the consequences remain unaddressed, under-researched, or only marginally reported.

In the meantime, the collected data amount to a modern redressing of the traditional “management” of the poor, the sick, the elderly, and racial minorities in capitalist systems, as these segments of society generate more administrative data and therefore risk skewing AI-driven decisions toward a management of “digital poorhouses,” in which assistance systems double as instruments of monitoring and control, albeit by seemingly unbiased machines.⁹ Which AI companies have been held responsible for the consequences of their models? Which country’s laws have adequately addressed AI systems’ inheriting and perhaps even amplifying structural inequalities embedded in historical data?

Worse still, even if the law managed to catch up with the systemic “social murder” embedded in these systems, they are inherently difficult to audit. This may be because of the proprietary nature of the data, the opacity of the models, or the technical complexity of the systems themselves. This deepens the opacity of AI systems’ decision-making and creates new forms of accountability problems. How are individuals to contest decisions that materially affect their lives if they cannot contest the hidden biases within the automation? Who can, and how, question or even adequately explain the decisions of an omnipotent machine, the workings of which are in most cases hidden even from those who built it?¹⁰

Altogether, algorithmic systems create a governance structure where consequences are concrete, if not brutal, but responsibility for those decisions is diffused and concealed. Indeed, the murderer is even more hidden—but murder it remains.

CounterCritical