The Difficulty of Coding Terrorism

On Sept. 25, we released a report with the Center for Strategic and International Studies (CSIS) analyzing terrorism trends in the United States. Among other findings, the report concluded that the number of left-wing terrorism incidents rose in the first half of 2025 (albeit from a low baseline) while right-wing terrorism fell considerably during that same period. The report generated considerable attention, picked up by the White House, members of Congress, and popular political influencers. Some critics of this report and other terrorism analysts pointed out the challenges of coding terrorism incidents. More specifically, they noted that small numbers of cases can be overinterpreted as evidence of large trends, that definitions and coding rules are difficult to design and apply consistently across ideologies, and that these choices can create a misleading sense of equivalence between different types of threats. They are not wrong. Indeed, we confronted all of these issues and more in our efforts to track and analyze terrorism in the United States.

Defining terrorism is itself a long-running and unresolved debate. Scholars, policymakers, and major data projects—from the Global Terrorism Database to Europol and the U.S. State Department—use different criteria depending on their analytical purposes and policy goals. Indeed, even seemingly obvious categories such as “left-wing” or “right-wing” are contested. This piece is not about resolving those definitional disagreements. Instead, it focuses on the practical challenges that arise after a definition has been chosen, when coders must determine whether a given incident meets that threshold and how to analyze it among other incidents.

We identified 11 recurring difficulties in coding and analyzing domestic terrorism, concluding that this field is marked by a high degree of subjectivity. Yet these challenges do not diminish the importance of collecting and analyzing terrorism data. Data-backed analysis is essential for identifying real trends amid political noise, allocating resources effectively, and helping policymakers and the public understand the scale and nature of violent threats. Data plays a crucial role in grounding debates in evidence rather than speculation. Because of the importance of terrorism data, it is imperative that consumers of this research understand the difficulties involved when conducting analysis.

These 11 challenges can be grouped into three categories. The first set of challenges involves inclusion questions about which incidents do and do not count as terrorism; the second set involves coding challenges for acts that are clearly terrorism but do not fit neatly into coding categories; and the third set involves analysis questions that might lead different analysts to draw different conclusions even when they agree on the underlying data. In the end, terrorism researchers must recognize these ambiguities even as they use data to inform policy.

Inclusion Challenges

Mental Illness and Incoherence

A distinct coding challenge lies in distinguishing between political extremism and psychological disturbance. Many violent actors display elements of both. For example, lone actors responsible for some of the Islamic State’s biggest attacks in the West had signs of mental illness.

Coders face an epistemic dilemma: Should acts accompanied by both ideological rhetoric and strong indicators of mental illness be coded as terrorism? Psychological incoherence undermines one of terrorism’s definitional pillars—intentionality. That is, when a perpetrator’s actions are heavily shaped by mental illness, it becomes difficult to tell whether the individual is deliberately using violence as a tool to advance a political cause and send a message to a broader audience or lashing out for personal reasons. However, excluding such incidents risks missing how extremist narratives can channel personal instability into political violence. Radicalization and pathology are not mutually exclusive. Coding decisions depend on whether the analyst privileges stated intent or psychological credibility, a choice that cannot be resolved by data alone. 

Where Does Homeland Begin?

Definitional confusion also extends to geography. Most datasets distinguish “domestic” from “international” terrorism, but these categories blur under globalization. Diasporas, online radicalization, and transnational politics make it increasingly unclear where “domestic” begins. Many “homegrown” terrorists are influenced by foreign ideologies and groups, while actors based abroad may strike within U.S. borders.

Such a question is especially vital for understanding attempts to target civil aviation. Thanks to improved airline security, such attacks are relatively rare. This category, however, includes highly consequential attacks and plots such as the Pan Am 103 bombing (which killed 190 Americans). It also includes plots by al-Qaeda and its affiliates to bomb U.S.-bound airplanes, including the massive 2006 plot to bomb nine transatlantic flights simultaneously (which is why travelers can no longer bring unregulated amounts of liquid aboard flights) and the near-miss “underwear bomber’s” attempt to blow up a flight over Detroit in 2009. Different datasets might legitimately code these cases as international or domestic depending on whether they prioritize the origin and destination of the flights, the identity of the attackers, the location where the attack occurred, or where the plot was disrupted. Given their scale, these incidents have a significant effect on how dangerous people think the terrorism threat is, which is often measured by actual and intended death tolls.

Property Damage Only?

Another difficulty concerns whether violence targeting property rather than people qualifies as terrorism. Certain definitions, such as the FBI’s, leave room for the inclusion of property damage. Groups such as the Earth Liberation Front or the Animal Liberation Front committed hundreds of attacks against logging operations, construction sites, and laboratories that did not harm people. These acts clearly used violence toward political ends, yet many databases (including the CSIS database) exclude them because their violence is typically explicitly not intended to harm people. This has become a flashpoint of some criticism from the political right, particularly regarding the exclusion of attacks on buildings, infrastructure, and commercial property during the 2020 Black Lives Matter protests, as well as attacks against Tesla vehicles and facilities in early 2025.

Our own methodology draws two relevant lines here. First, we generally exclude riots, which are typically spontaneous, driven by collective unrest, anger, and opportunistic criminal activity. We do, however, include incidents where perpetrators may have used the cover of riots to attempt or carry out premeditated attacks, such as some elements of the assault on the U.S. Capitol on Jan. 6, 2021, that involved threats to human lives as well as property damage.

Second, attacks using nonlethal weapons are generally excluded from our terrorism definition, except for incidents clearly intended to kill or maim people and cases of significant property destruction in contexts where a political issue’s history gave the attack the character of an implicit lethal threat. Most of these latter cases involved arson attacks on religious institutions, abortion-related targets, and political offices.

These coding choices reflect an implicit bias toward physical casualties as the measure of terroristic “seriousness.” Yet, as some leading scholars emphasize, fear—not fatalities—defines terrorism. Arson at a synagogue, vandalism of abortion clinics, or destruction of critical infrastructure may serve the same coercive function as bombings. Coders must decide whether to prioritize human harm or symbolic intimidation. Narrow definitions undercount ideological violence; broad ones risk overextending terrorism into political vandalism. 

Nonpolitical Attacks on Political Targets

Another conceptual problem arises with apparently purposeless or opportunistic attacks. Many recent acts of violence combine nihilism, notoriety-seeking, and inconsistent or confused ideological gestures. The 2024 attempt to assassinate then-presidential candidate Donald Trump in Butler, Pennsylvania, is a prime example. Despite targeting one of the most prominent political figures in the country, evidence suggests the shooter did not hold a clear political grievance and reportedly searched for campaign events for both Trump and then-President Biden. In such cases, offenders may display only a thin patina of politics, leaving coders to decide whether an act is political violence or simply violence surrounded by political imagery. 

Assessing intent is increasingly inferential, especially as a growing share of offenders in the United States act without organizational guidance. For these cases, coders must reconstruct bad actors’ motives from manifestos, online posts, personal grievances, or target choice. These sources are often contradictory, sparse, or intentionally misleading. For example, the 23-year-old shooter who killed two children and injured 19 others in August at a Minneapolis Catholic school and church left a deliberately jumbled trail of memes, symbols, and phrases to invite confusion and partisan finger-pointing as a kind of ideological trolling. By contrast, more sophisticated extremist groups have founding documents, propaganda articulating their ambitions, and leadership statements that provide a more stable guide to motive. Traditional coding systems—built for hierarchical organizations such as the Provisional Irish Republican Army (PIRA) or al-Qaeda—struggle to interpret atomized acts where personal and political motives are harder to separate.

The Intended Audience

A further challenge involves determining whether an attacker sought to generate a broad psychological impact. As one PIRA terrorist said, “you don’t bloody well kill people for the sake of killing them.” Terrorism is conventionally understood as violence designed not only to harm victims but also to send a message to a wider political community. Yet distinguishing between violence aimed at a specific individual and violence intended to resonate more broadly is rarely straightforward. This problem is especially difficult because the question is not whether the violence caused a broad psychological impact, but whether that result was the attacker’s goal: some attacks unintentionally cause a broader impact, while others intended to do so fizzle out.

Our methodology distinguishes between terrorism intended to intimidate and cause fear among certain groups and hate crimes that were unlikely intended to reach a wider audience. In practice, this requires assessing whether a reasonable person could expect news of the violence to extend beyond the immediate victims to a broader political or social group and shape the actions of a broader community. As with other dimensions of intent, these assessments introduce unavoidable interpretive uncertainty into terrorism coding. 

Coding Challenges

“Salad Bar” Ideologies

Modern attackers often draw selectively from what counterterrorism officials have described as a “salad bar” of ideological references—borrowing fragments from multiple traditions without internal coherence. For example, the perpetrators of the 2019 El Paso and 2022 Buffalo terrorist attacks blended the traditionally right-wing ideology of white supremacy with the traditionally left-wing concern of environmentalism by advancing the eco-fascist view that immigrants and racial minorities were responsible for overpopulation and environmental destruction.

These hybrid ideologies complicate coding because they upend traditional ideological categories. Databases (including our own) typically require identifying a dominant ideological “type” (religious, right-wing, left-wing, nationalist, and so on). For both the El Paso and Buffalo attacks, we determined the primary ideology was right-wing based on our definitions. Nonetheless, contemporary attackers increasingly self-assemble their worldviews from online cultures and conspiracy theories. The salad-bar nature of modern extremism thus blurs the line between clear ideological categories. 

Shifting Israel-Palestine Violence

The Israel-Palestine conflict also creates definitional issues. CSIS has traditionally coded pro-Palestine or pro-Israel violence as “ethnonationalist,” considering it to be an attempt to advance a cause of a particular people. For example, in 1994, Rashid Baz opened fire on a van of Orthodox Jewish students on the Brooklyn Bridge, killing one, as revenge for an earlier attack against Muslims in the West Bank by Israeli settlers.

However, much of the recent anti-Israel sentiment can, at least in part, be described as closer to a left-wing perspective, with those critical of Israel seeing the Palestinian struggle as part of a broader anti-imperialist campaign on behalf of the oppressed. Some violence, however, is linked to more traditional antisemitism, which is typically (but not always) the domain of the extreme right. 

Single-Issue Terrorism

Another category problem involves “single-issue” terrorism: violence over narrow causes such as abortion, animal rights, or taxation. These cases often meet definitions of terrorism because they involve premeditated political violence intended to have a broader psychological effect, but they raise challenges when it comes to classifying ideology. Their highly specific motives may not map easily onto broader ideological families, yet treating each cause as its own category risks an unwieldy level of fragmentation. Our dataset addresses this by tracking these narrower motivations in a secondary ideological variable, which preserves important nuance. For example, our definition of right-wing terrorism (a primary ideological category) includes attacks motivated by ideas of racial or ethnic supremacy; opposition to government authority on the basis that it is tyrannical and illegitimate; misogyny, including “involuntary celibates” (incels); hatred based on sexuality or gender identity; belief in the QAnon conspiracy theory; opposition to abortion; or partisan extremism where violence is justified against political opponents and parties perceived as advancing left-wing agendas. 

While aggregating incidents into broader ideological buckets can reveal important broader trends, it can also obscure important distinctions among narrow single-issue motives and risk overstating the coherence of otherwise unrelated causes.

Analysis Challenges

Uneven Plot Detection

The empirical record of terrorist plot detection is uneven. Law enforcement and intelligence agencies devote disproportionate attention to some ideological categories, especially Islamist threats after 9/11. As a result, jihadist plots are more frequently detected, disrupted, and reported than plots motivated by other ideologies. The resulting selection bias is most pronounced when the early-detected plots might not have even progressed to attacks absent government intervention. This bias filters directly into open-source datasets like that of CSIS, which rely on public reports that themselves mirror political and policing priorities.

Consequently, the apparent frequency of certain types of terrorism plots often reflect surveillance intensity rather than objective prevalence. For much of U.S. history, white supremacist violence was not recorded as terrorism. If it had been, large parts of U.S. history would be known for the prevalence of terrorism. At times, frequency reflects an increase in attacks, but in other cases, it means attention and perceptions changed. 

Casualties Versus Political Impact

Measuring the threat of terrorism is also difficult, in part because different groups attack different target sets, and it is difficult to compare across them. For example, the assassination of Charlie Kirk resulted in a single fatality, yet its political repercussions have been immediate and far-reaching. By contrast, jihadist and white supremacist attacks in the United States in recent years have killed dozens of people in single incidents, generating profound social and local trauma but little political response and only limited long-term attention outside the affected communities. Comparing these forms of impact—symbolic, political, psychological, and lethal—reveals how difficult it is to develop a single metric of “severity” across ideological categories. 

Low Numbers

Because the United States experiences relatively few terrorism incidents each year, assessing trends is inherently difficult. When dealing with low numbers, the inclusion or exclusion of just a few borderline cases can shift apparent patterns. This makes coding decisions especially consequential. Choices about whether to classify an incident as terrorism or how to assign its ideology can alter trendlines that are based on small samples. For example, in January 2025, 24-year-old Riley Jane English approached police on the National Mall in Washington, D.C., and turned herself in while carrying two Molotov cocktails, a folding knife, and a lighter, saying she intended to kill senior officials in the Trump administration. Reasonable people can disagree over whether her voluntary surrender and apparent intent qualify the act as a terrorist plot. In analysis where different interpretations of a small number of events can produce noticeably different pictures of the threat, transparency about methodology and consistency in coding are essential for credible analysis.

Managing Ambiguity

While these challenges do not render terrorism coding impossible or meaningless, scholars should be clear about definitional criteria, coding procedures, and sources. Triangulating between datasets, acknowledging uncertainty, and systematically reviewing contested cases can improve reliability.

The goal is not to eliminate ambiguity but to manage it. Coders and analysts must recognize that they are not merely recording facts but participating in meaning-making processes that shape policy, perception, and history. Methodological transparency, rather than false precision, should be the field’s guiding principle.

As states and researchers continue to produce terrorism databases, threat assessments, and counterterrorism strategies, they must confront the uncomfortable truth that violence can be ideological without coherence, political without organization, and terrifying without intent. The challenge is not to erase this randomness but to understand it—to study how modern societies classify, rationalize, and respond to acts that defy easy categorization.