Science Papers at 'War': 21.4 Million Abstracts Reveal Why Scientists Increasingly 'Fight'

A Strange Linguistic Phenomenon

Open the abstract of any AI paper and you'll likely read sentences like:

• "We attack the problem…"
• "Our method beats the baseline…"
• "This finding conquers a long-standing challenge…"

A University of Pennsylvania interdisciplinary team analyzed 21.4 million scientific paper abstracts (2010-2025, from the OpenAlex and PubMed databases) and found that "war vocabulary" in scientific writing has risen 48% (OpenAlex) and 32% (PubMed) over 15 years — with the growth accelerating sharply after 2019.

More surprisingly: using war framing doesn't make papers more persuasive — it reduces readers' trust in the research.

Key Findings

• Scale: Militaristic language in abstracts rose 48% (OpenAlex, all disciplines) and 32% (PubMed, biomedical) from 2010-2025. Cross-database correlation: r = 0.96, p < 10⁻⁸ — a real cross-disciplinary trend, not a database artifact.
• Timing: Growth accelerated sharply after 2019; the COVID era and the post-2022 large language model era further amplified it.
• Disciplines: Social sciences show the highest *levels* of militaristic language; engineering and computer science show the *fastest* growth.
• Geography: Abstracts by Global South authors grew fastest in militarization.
• Language background: The militarization gap between non-native and native English speakers *narrowed* during the COVID and LLM eras — non-native authors "caught up," likely by imitating high-frequency papers in AI/biomedical fields.

Link to Real-World Wars

Comparing against Uppsala Conflict Data Program (UCDP) data:

• Country-level conflict intensity correlates with militarized language in that country's authors' abstracts: r = 0.77-0.84
• Global annual conflict intensity correlates with abstract militarization.

This isn't just scientists using war metaphors — real-world warfare is seeping into scientific language.

Causal Experiment: War Framing Reduces Credibility

A within-subject experiment tested causality:

• 801 participants, each reading pairs of abstracts (32,040 trials total)
• Identical content, one version war-framed ("attack the problem"), one neutral ("address the problem")

| Measure | Effect | p-value | |---------|--------|---------| | Credibility | -0.18 Likert units (d_z = -0.28) | < 10⁻²⁰ | | Funding willingness | d_z = -0.12 | significant | | Policy support | d_z = -0.08 | significant | | Urgency | slight upward trend (d_z) | marginal |

War-framed abstracts were judged less credible, less fundable, and less policy-worthy; the only minor "benefit" was a trend-level increase in perceived urgency. It's like writing "I defeated all my competitors on the battlefield" on a résumé — readers don't think you're better, they think you're boastful or aggressive.

Why Scientists Increasingly 'Fight'

1. COVID's linguistic shock: "fighting the pandemic" dominated discourse in 2020-2022, contaminating scientific vocabulary ("frontline workers," "battlefield of the pandemic"). 2. LLM-era competition narrative: "arms race," "beating baselines," "attacking benchmarks" are now standard AI writing — matching the fastest-growing disciplines (CS, engineering). 3. Global South linguistic convergence: non-native authors learn by imulating high-frequency (heavily militarized) AI/biomedical papers. 4. Seepage from real wars: scientists don't write in a vacuum; conflict-saturated news and discourse influence word choice unconsciously — experiment participants attributed their judgments to "the evidence itself," not the framing.

Methodological Highlights

• Large-scale observation + causal experiment: 21.4M abstracts show *what* happens; the 801-person experiment shows *why*.
• Cross-database validation: r = 0.96 agreement between OpenAlex and PubMed rules out database bias.
• Real-world conflict linkage: connecting scientific language to UCDP data elevates the work from linguistics to social psychology.

Honest Assessment

Strengths: massive scale (21.4M abstracts, 16 years, two databases); robust within-subject design (32,040 trials); interdisciplinary team (bioengineering + communication + network science); small but highly stable effects.

Limitations:

• The definition of "militaristic language" may be too broad — words like "target protein" or "attack the problem" have long had neutral scientific uses, possibly inflating the effect.
• The 801 experiment participants may not represent typical readers (scientists vs. general public).
• The study doesn't distinguish deliberate vs. unconscious militarization — crucial for interventions.
• Only four outcomes were tested; war framing might have positive effects elsewhere (memorability, virality) that weren't measured.

Implications for Science Writing

The practical advice: use fewer war framings — not for political correctness, but because they genuinely lower the odds your paper is trusted by readers, reviewers, and funders.

The deeper point: how we write is shaped by our era. The COVID era, the LLM era, the era of geopolitical conflict — these aren't neutral backdrops; they permeate our vocabulary, our thinking, and our scientific judgments. Scientists often claim objectivity, but this study of 21.4 million abstracts shows: no scientist is truly neutral — we are all products of our times, down to our abstracts.

---

Paper: https://arxiv.org/abs/2606.23462 HTML version: https://arxiv.org/html/2606.23462v1 Authors: Sovesh Mohapatra, David Lydon-Staley, Dani S. Bassett (University of Pennsylvania) Code: no official code repository provided