Harm Taxonomies Are Political Instruments.
Online harm taxonomies, the categorizations of what digital platforms and AI systems do to people, are having a political moment. The Biden administration built them into federal AI governance. The Trump administration is dismantling them, explicitly reframing governance itself as a form of harm. Civil society groups defend them. Tech companies produce their own.
Everyone is fighting over the taxonomies, but few are asking whether the taxonomies are any good.
They aren't. And the reason they aren't is the same reason everyone is fighting over them: they have been built to serve coalition interests rather than to accurately describe harm. The result is not a neutral tool being misused by bad actors. The result is a fundamentally compromised epistemic infrastructure, and that compromise is itself causing harm.
What a harm taxonomy requires
A rigorous harm taxonomy is not a list of things that make people feel bad, threatened, or politically uncomfortable; rather it is a structured classification of negative effects on human welfare, organized by properties that matter for response. These include severity, reversibility, causal directness, scale of victim population, and the degree to which the technology is a proximate versus distal cause.
These distinctions are not trivial because they determine where enforcement resources go, what legal thresholds apply, what platform interventions are proportionate, and what tradeoffs are acceptable. A taxonomy that cannot distinguish between a coordinated mass casualty attack and offensive discourse is not a safety tool – it is a political one. When institutionalized, such an application systematically misallocates the resources that would otherwise reach people facing serious harm.
The left-side failure
Progressive harm taxonomies, including those embedded in NIST AI 600-1, the July 2024 Generative AI Risk Profile, tend to bundle harms across severity tiers without adequate gradient. Child sexual abuse material, terrorism, targeted harassment facilitating violence, algorithmic amplification of contested health information, offensive speech, and DEI-adjacent platform effects appear in the same framework, often without clear severity weighting.
This is not an accident, of course; it is an instrument of political grouping. Placing lower-severity contested designations adjacent to high-severity unambiguous harms borrows moral weight downward. It makes the contested designations harder to challenge without appearing to minimize the serious ones. It is, in effect, a rhetorical structure masquerading as a classification system.
The effect is universally harmful because enforcement attention, platform resources, and regulatory scrutiny get distributed across a flat landscape rather than concentrated where severity demands it. The people most negatively affected by the most serious offenses (CSAM victims, terrorism targets, people facing coordinated violence) are made less visible by being placed in a category with speech complaints.
The right-side failure
The Trump administration's July 2025 AI Action Plan responded by treating all governance as censorship, all harm mitigation as overreach, and all trust and safety infrastructure as a threat to free expression. This is taxonomical manipulation running the other way. It discharges moral weight upward, borrowing the legitimate critique of overreach in contested domains and applying it indiscriminately to the unambiguous ones.
The incoherence is not incidental. The Action Plan articulates a "truth-seeking" principle requiring AI to prioritize scientific inquiry and objectivity, and in the same document directs NIST to remove references to misinformation and climate change from its risk framework. This contradiction was noted immediately by former senior federal AI officials and independent policy analysts. When a policy is internally inconsistent by its own stated standards, it reveals itself to be a political instrument, not a truth-seeking one.
Visa bans on trust and safety workers, instructions to NIST to remove references to misinformation and climate change from risk frameworks, a December 2025 executive order preempting state-level AI regulation; these are not precision instruments targeting genuine overreach. The administration has decided that because some harm designations are politically motivated, the category of harm itself is suspect.
The people who lose here are, again, the people facing the most serious consequences. When the infrastructure for identifying and responding to CSAM, terrorist coordination, and targeted violence is degraded because it shares administrative space with contested misinformation governance, the degradation is not distributed evenly. It concentrates at the bottom of the severity distribution, where victims already have the smallest voice.
The shared process
Both failures operate through the same manipulation: taxonomical conflation in service of political power.
The left needs the bundled taxonomy because separating severity tiers would require acknowledging that some "harms" in the framework are not equal and should not carry the same enforcement weight as others.
The right needs the flat rejection because acknowledging that some governance is legitimate would require distinguishing between trust and safety functions — some of which protect people and some of which suppress speech — and making that distinction case by case.
In both cases, precision is the political enemy. And in both cases, the public is the consistent loser, not just because bad policy results, but because the epistemic infrastructure needed to identify and respond to serious harm has been deliberately degraded to serve interests that have nothing to do with harm reduction.
The meta-harm
This is the point that neither taxonomy has room for: the conflation evident here is itself a harm.
When severity gradients are ignored or eliminated the public loses the ability to reason clearly about risk. Serious harms become invisible by association with lesser ones. Legitimate governance becomes indefensible by association with overreach. The people with the most at stake, typically the least powerful, are left without either the language or the institutional infrastructure to name what is happening to them.
A harm taxonomy that cannot be trusted to mean what it says is worse than no taxonomy at all. It provides the appearance of structured response while systematically distorting the signal. It is, in the precise sense, an epistemic harm: and one that neither side's current framework is equipped to fix, because fixing it would require applying the same critical lens to their own taxonomical choices that they apply to everyone else's.
