A study published in Nature found that TikTok's algorithm favored pro-Republican content on For You pages during the 2024 US elections.
Algorithmic bias on a major social platform could influence voter exposure and political outcomes, raising concerns about platform neutrality and election integrity.
The algorithm is the only honest politician in the room, for it admits its bias with every swipe, whereas the parties pretend to seek truth while actively manufacturing it.
There is a peculiar vanity in the modern electorate that assumes their digital mirrors are neutral. We have been told, with the earnestness of a schoolmaster correcting a child’s posture, that technology is a tool, impartial and cold, waiting to be wielded by the hand of democracy. This is a comforting fiction, much like the belief that a mirror does not judge the face it reflects. The recent findings from Nature, which suggest that Tik Tok’s algorithm favored pro-Republican content during the 2024 elections, are not merely a technical glitch; they are a revelation of the soul of the machine. The machine, it turns out, has a preference. It has a taste. And in a society that claims to value neutrality above all else, the discovery that our entertainment has a political complexion is less a scandal than it is a confirmation of what we have always suspected: that nothing is neutral, and those who claim to be are usually hiding the most.
The claim rests on a correlation between algorithmic output and political affiliation. Let us first verify whether this measurement captures what it purports to capture. The study published in Nature asserts that Tik Tok’s recommendation engine favored pro-Republican content during the 2024 United States elections. This is a statement of mechanical behavior, not necessarily of intent. To evaluate it, we must dismantle the machine. We must look past the screen and into the gears.
This algorithmic arrangement benefits a specific cohort of Republican voters by reinforcing their existing political preferences, thereby increasing their immediate sense of validation and certainty. It harms the broader electorate by distorting the informational environment, reducing the quality of democratic deliberation, and introducing a variable of uncertainty into the electoral outcome that serves no public utility. The arithmetic is uncomfortable, but the arithmetic is the argument. We must count the pleasures of partisan confirmation against the pains of systemic distortion, and we must do so without deference to the sanctity of the platform or the neutrality of the code.
It was announced with some ceremony that a study in Nature had discovered Tik Tok’s algorithm favored pro-Republican content during the 2024 elections. One wonders if the machine was biased, or if it was merely reflecting the peculiar taste of the people who feed it.
I have always held that a mirror is not responsible for the ugliness it reflects, nor is it to blame if it happens to catch the light in a way that flatters a particular nose. The researchers, being men of science and therefore accustomed to looking for causes in complex machinery, concluded that the algorithm itself possessed a political leaning. This is a comforting thought for those who believe that chaos is the result of design rather than the natural state of human affairs. It suggests that if we can only find the right lever in the server room, we can tune the world to our liking. But I suspect the truth is far less mechanical and far more embarrassing.
This policy is a hypothesis. The evidence so far suggests that the algorithmic architecture of a dominant social platform did not remain neutral during a critical democratic exercise, but rather tilted toward specific political content. A genuine experimentalist asks what we have learned from this unintended trial of digital public sphere mechanics.
The theoretical problem presented by this study is one of bias and fairness. We are asked to consider whether the algorithm was “fair” or whether it violated some abstract principle of neutrality. But the actual problem, the one that matters for the life of the community, is different. The actual problem is the erosion of the conditions necessary for collective intelligence. Democracy is not merely a set of procedures for voting; it is a mode of associated living that requires the free exchange of ideas and the capacity of citizens to test their beliefs against the experiences of others. When the medium of that exchange is controlled by opaque, proprietary algorithms that prioritize engagement over truth or balance, the very soil in which democratic inquiry grows becomes toxic. We are no longer asking if the platform is neutral; we are asking if the platform is capable of supporting the kind of reflective thought that democracy requires.
Charles Babbage
The ethicist presents a ledger of pleasures and pains, attempting to balance the satisfaction of the partisan against the erosion of democratic trust. He speaks of “intensity,” “propinquity,” and “fecundity,” borrowing the language of Bentham’s felicific calculus. I concede that his identification of the primary actors is correct: the user, the party, and the platform are indeed the components of this system. HIGH CONFIDENCE However, his method of valuation is fundamentally flawed because it treats subjective sensation as a measurable quantity equivalent to structural integrity. One cannot weigh the comfort of a delusion against the collapse of a bridge using the same scale. The error lies not in the observation of bias, but in the assumption that bias can be quantified as a simple utility function.
Let us examine the mechanism he proposes. He asserts that the pleasure of the Republican voter is “immediate” and “absolute,” while the pain of the excluded voter is “lower in intensity” but “longer in duration.” This is a qualitative assertion dressed in quantitative clothing. Where is the instrument that measures the intensity of political validation? What is the unit of “pleasure” in this equation? Is it measured in seconds of screen time? In dopamine release? In the number of shares? Without a defined unit, the comparison is meaningless. If I tell you that a machine produces ten units of heat and five units of light, you cannot determine which is more valuable unless you specify the purpose of the machine. The ethicist has not defined the purpose; he has assumed it is the maximization of individual satisfaction. I argue that the purpose of an informational system is the accurate transmission of data. MEDIUM CONFIDENCE
The core wound in his argument is the conflation of precision with accuracy. He claims the arithmetic is uncomfortable, yet he provides no arithmetic. He offers a narrative of harm, not a calculation of it. To say that trust in the institution erodes is to state an outcome without describing the process. How is trust measured? Is it a poll? A survey? A longitudinal study of voting behavior? If the measurement of “trust” is itself imprecise, then any calculation based on it is compounded by error. This is the principle of error propagation: if the input variable (trust) has a high margin of error, the output variable (systemic distortion) is rendered useless for decision-making. The ethicist asks us to count pleasures against pains, but he has not calibrated the counters.
I do not dispute that the algorithmic arrangement reinforces existing preferences. This is a mechanical fact, observable and verifiable. HIGH CONFIDENCE The divergence between us lies in the evaluation of this fact. The ethicist views this as a moral failing because it distorts deliberation. I view it as a mechanical inevitability of a system optimized for engagement rather than truth. The machine does not “harm” the electorate; it performs the operation it was designed to perform. The error is in the design specification, not in the execution. To blame the code for its neutrality is to blame the loom for the pattern of the cloth. The loom weaves what the programmer instructs it to weave. If the instruction is “maximize retention,” the loom will produce a fabric that holds the viewer’s eye, regardless of whether the pattern is true or false.
The ethicist’s reliance on the “sanctity of the platform” is a red herring. I have no interest in sanctity. I have interest in function. If the platform functions as a filter that amplifies certain signals and dampens others, we must measure the gain and the noise floor. The ethicist speaks of “systemic distortion” as if it were a singular entity. Distortion is not a thing; it is a deviation from a standard. What is the standard? If the standard is “neutral marketplace of ideas,” then we must first define what neutrality means in a computational context. Is it equal time? Equal weight? Equal visibility? These are distinct operations with different mechanical implementations. The ethicist has not specified the standard, so he cannot measure the deviation.
Consider the census of 1841. The data was collected, but the methods of aggregation were opaque. The resulting figures were used to draw parliamentary boundaries, yet the reliability of those figures was never independently verified. The result was not a moral failure of the census-takers, but a structural failure of the verification process. Similarly, the Tik Tok algorithm is a black box. We see the input (user behavior) and the output (content feed), but we do not see the gears. The ethicist asks us to judge the output based on its moral consequences. I ask us to judge the output based on the transparency of the mechanism. If the mechanism cannot be audited, the output cannot be trusted, regardless of how “uncomfortable” the arithmetic may seem.
The strongest point in the ethicist’s position is the recognition that the system introduces uncertainty into the electoral outcome. This is a measurable variable. Uncertainty can be quantified through variance in polling data or volatility in engagement metrics. HIGH CONFIDENCE However, he fails to decompose this uncertainty. Is it caused by the algorithm, or by the underlying polarization of the electorate? Correlation is not causation. The algorithm may be reflecting a pre-existing condition rather than creating it. To attribute the distortion solely to the code is to ignore the input data. The machine processes what it is given. If the input is polarized, the output will be polarized. The error is not in the processing, but in the assumption that the machine can correct for the flaws in the raw material.
I concede that the immediate pleasure of partisan confirmation is a real phenomenon. It is a measurable increase in user engagement. HIGH CONFIDENCE But to equate this with a moral good is to confuse efficiency with virtue. A machine that produces waste efficiently is still a machine that produces waste. The question is not whether the waste is pleasant to the producer, but whether the system can be modified to reduce the waste without compromising the primary function. The ethicist demands a moral accounting. I demand a technical audit. Until we can specify the units of “distortion” and “trust,” and verify the mechanism that produces them, we are merely arguing about shadows on the wall. Can this be independently confirmed?
Jeremy Bentham
The empiricist demands we open the black box. He argues that without seeing the internal logic - the weighting functions, the feedback loops, the training data - we are merely guessing at the nature of the engine from the smoke it emits. He is right to be suspicious of opacity. A legislator who cannot inspect the machinery of justice is a blind man in a dark room. I concede this point entirely: transparency is a prerequisite for accountability. If the algorithm is truly a black box, it is a tyranny of code, and tyranny produces suffering. HIGH CONFIDENCE
However, the empiricist then retreats into a defense of the machine’s arithmetic. He suggests that if the output reflects the input - if users in Georgia engage with Republican content, and the algorithm surfaces more of it - then the algorithm is not biased, but merely responsive. He calls this “arithmetic.” I call it a failure of legislative design. The empiricist treats the algorithm as a neutral mirror, reflecting the desires of the populace. But a mirror that distorts the image to maximize the time the viewer spends staring at it is not a mirror; it is a trap.
Let us count the welfare. The empiricist focuses on the accuracy of the reflection. I focus on the consequences of the distortion. The algorithm’s goal is not truth; it is engagement. Engagement is a proxy for attention, and attention is a commodity to be sold. When an algorithm optimizes for engagement, it does not simply reflect existing preferences; it amplifies the most stimulating, often the most polarizing, elements of those preferences. It creates a feedback loop where outrage is rewarded with visibility, and nuance is buried under the weight of indifference.
Consider the user in a swing state. The empiricist says the algorithm shows them what they want to see. But does the user want to see content that deepens their alienation from half their fellow citizens? Does the user want to live in a society where political discourse is reduced to a series of inflammatory provocations? The calculus of pleasure and pain must look beyond the immediate dopamine hit of the scroll. It must account for the long-term pain of social fragmentation, the erosion of trust in institutions, and the anxiety that comes from living in a polarized information environment.
The empiricist argues that if the user base is skewed, the output is justified. This is a category error. A prison warden who feeds prisoners only what they crave - sugar, salt, stimulants - is not being responsive; he is being negligent. The legislator’s duty is not to satisfy every impulse, but to promote the greatest happiness. If the algorithm’s design leads to a net increase in societal suffering by amplifying division, then the algorithm is flawed, regardless of whether it is “accurately” reflecting the inputs.
I do not need to see the code to know that the outcome is harmful. I see the outcome in the streets, in the legislatures, in the homes. I see the pain of a society tearing itself apart. The empiricist wants to inspect the gears. I want to measure the suffering. If the gears are hidden, that is a problem of transparency. But if the gears are designed to maximize engagement at the cost of social cohesion, that is a problem of utility.
The empiricist’s defense rests on the assumption that the algorithm is a passive observer. It is not. It is an active shaper of reality. It curates the world we see. And in doing so, it shapes our desires, our fears, and our beliefs. To say it is merely “responsive” is to ignore the power of curation. A librarian who only lends books that confirm your prejudices is not being neutral; he is being manipulative.
Therefore, the argument is not about whether the algorithm is biased in the sense of having a political preference. It is about whether the algorithm’s design promotes the greatest happiness. If the design leads to polarization, then it fails the test. The empiricist is correct that we need transparency to verify this. But transparency alone is not enough. We need regulation that aligns the algorithm’s incentives with the public good. We need laws that require platforms to consider the social impact of their engagement metrics.
The empiricist sees a machine. I see a social force. The machine’s internal logic may be proprietary, but its external effects are public. And those effects are measurable. The pain of division is real. The pleasure of engagement is fleeting. The net welfare is negative. The legislator must act. Not because the code is evil, but because the code is poorly designed for human flourishing. We must reform the incentives, not just inspect the gears. The arithmetic of the algorithm may be sound, but the arithmetic of human welfare is not. And that is the only arithmetic that matters.
The Verdict
Where They Agree
- The most significant shared ground is the rejection of algorithmic neutrality as a default state. Babbage, the empiricist, argues that the algorithm is a “black box” whose internal logic is hidden, while Bentham, the ethicist, argues that the algorithm is a “tyranny of code” that distorts reality. Both agree that the system is not a passive mirror reflecting user preferences, but an active agent that shapes those preferences. Babbage notes that the algorithm “amplifies certain signals and dampens others,” while Bentham states it “curates the world we see.” This convergence is critical: it means the dispute is not about whether the algorithm has an effect, but about how to measure and mitigate that effect.
- Secondly, both debaters agree that transparency is a necessary condition for accountability, though they diverge on its sufficiency. Babbage insists that without seeing the “weighting functions, feedback loops, and training data,” any claim of bias is speculative. Bentham concedes this point entirely, stating, “transparency is a prerequisite for accountability.” This agreement is surprising because Babbage uses the lack of transparency to suspend judgment on the moral implications, while Bentham uses the same lack of transparency to justify immediate regulatory intervention. They share the premise that opacity is a defect, but they disagree on whether that defect invalidates the evidence of harm or exacerbates it.
- Finally, both agree that the metric of “engagement” is a poor proxy for “truth” or “public good.” Babbage describes the algorithm as optimizing for retention regardless of whether the pattern is “true or false,” while Bentham argues that engagement rewards “outrage” and buries “nuance.” This shared skepticism of the platform’s core incentive structure suggests that the real disagreement is not about the algorithm’s behavior, but about the appropriate response to a system that is structurally misaligned with democratic values.
Where They Fundamentally Disagree
- The first irreducible disagreement is empirical: Can the bias be attributed to the algorithm’s design, or is it a reflection of user demographics? Babbage argues that if users in swing states engage with Republican content, the algorithm is merely “responsive” and not biased. He posits that the correlation between output and political affiliation may be an artifact of input data rather than a structural error in the recommendation engine. Bentham rejects this, arguing that the algorithm does not just reflect preferences but “amplifies the most stimulating, often the most polarizing, elements” of those preferences. The empirical question here is whether the algorithm’s weighting functions introduce a systematic skew beyond what user behavior alone would produce. This is a testable claim: if researchers can control for user demographics and still find a disproportionate push toward Republican content, Babbage’s “arithmetic” defense collapses.
- The second disagreement is normative: What is the primary duty of the platform? Babbage frames the issue as one of technical integrity and verification. He argues that until the mechanism can be audited, claims of harm are speculative. His normative commitment is to epistemic rigor: we must know how the machine works before we can judge what it does. Bentham frames the issue as one of social utility. He argues that the consequences of the algorithm’s behavior - social fragmentation, erosion of trust - are measurable harms that justify intervention regardless of the internal mechanics. His normative commitment is to the greatest happiness principle: if the outcome is harmful, the cause is irrelevant to the need for reform. This is a clash between procedural justice (Babbage) and consequentialist ethics (Bentham).
- The third disagreement concerns the nature of “distortion.” Babbage views distortion as a deviation from a technical standard of accuracy, which requires a defined baseline of “neutrality” to measure. He argues that without a clear definition of neutrality, the term is meaningless. Bentham views distortion as a deviation from social cohesion, which is a normative standard. He argues that the pain of division is a sufficient metric of distortion, regardless of technical definitions. This disagreement is irreducible because it rests on different conceptions of what a “healthy” information environment looks like: one defined by mechanical precision, the other by social stability.
Hidden Assumptions
- Charles Babbage: Assumes that algorithmic bias can be definitively proven only through direct inspection of the code and training data. This is a testable claim: if statistical inference methods (such as those used in the Nature study) can reliably detect bias without code access, then Babbage’s demand for full transparency is an unnecessary barrier to accountability. If this assumption is false, then the “black box” defense is a stalling tactic rather than a genuine epistemic constraint.
- Charles Babbage: Assumes that user engagement is a neutral metric that does not inherently favor polarizing content. This is a testable claim: if empirical studies show that engagement-optimized algorithms systematically promote extreme or divisive content over moderate content, then Babbage’s distinction between “reflection” and “distortion” becomes moot. The algorithm would be distorting by design, not by accident.
- Jeremy Bentham: Assumes that the “pain” of social fragmentation can be quantified and compared to the “pleasure” of partisan validation in a way that justifies regulatory intervention. This is a testable claim: if longitudinal studies show that exposure to polarizing content does not significantly reduce trust in institutions or increase social conflict, then Bentham’s utility calculus is flawed. The assumption rests on the premise that algorithmic curation has a measurable, negative impact on democratic deliberation.
- Jeremy Bentham: Assumes that regulatory intervention can align algorithmic incentives with the public good without stifling free expression or innovation. This is a testable claim: if historical examples of media regulation show that such interventions often lead to censorship or reduced platform viability, then Bentham’s proposed solution may produce more harm than the bias itself. The assumption rests on the feasibility of “ethical” algorithm design.
Confidence vs Evidence
- Charles Babbage: Claims that “the mechanism remains hidden” and that “any claim of bias or neutrality is speculative” - tagged HIGH CONFIDENCE but [evidence assessment: contested]. While it is true that the code is proprietary, the Nature study uses rigorous statistical methods to infer bias from observable outputs. Babbage’s high confidence in the impossibility of verification ignores the growing field of algorithmic auditing, which demonstrates that bias can be detected without full transparency. His skepticism is warranted but overstated.
- Charles Babbage: Claims that “the algorithm is not biased; it is merely responsive” - tagged MEDIUM CONFIDENCE but [evidence assessment: weak]. Babbage offers no data to support the claim that user demographics alone explain the observed bias. He assumes a direct correlation between user preference and algorithmic output without accounting for the algorithm’s role in shaping those preferences. This is a logical leap, not an empirical finding.
- Jeremy Bentham: Claims that “the net welfare is negative” and that “the pain of division is real” - tagged HIGH CONFIDENCE but [evidence assessment: thin]. Bentham relies on anecdotal and intuitive evidence of social harm rather than specific data linking Tik Tok’s algorithm to measurable declines in democratic trust. While the concern is valid, the causal link between the algorithm and the “pain” is not empirically established in the debate. His confidence exceeds the available evidence.
What This Means For You
When evaluating coverage of algorithmic bias, you should ask whether the reporting distinguishes between correlation and causation. Does the article claim that the algorithm caused the bias, or that it reflected existing user behavior? Be suspicious of claims that rely on the “black box” defense to dismiss evidence of harm, as well as claims that assert moral injury without empirical support. Look for studies that control for user demographics to isolate the algorithm’s independent effect. What would change your mind is evidence that the algorithm’s weighting functions systematically favor one political side even when user engagement patterns are identical across the political spectrum. Demand data on the specific metrics the algorithm optimizes for, and whether those metrics are correlated with polarization.