Autonomous Weapons Survey
Technology and National Security Project
Autonomous strikes looked less intentional. Judgments about escalation
did not clearly change.
U.S. respondents evaluated a hypothetical missile strike on a naval
destroyer in international waters near China, with sailors killed. The
experiment varied which country launched the strike and whether the
weapons system was human-operated or fully autonomous AI-enabled.
Main findings
- Autonomy reduced perceived intent in both
scenarios. In the adjusted analysis, intentionality ratings
fell 9.0 points for a U.S. strike and 10.3
points for a Chinese strike, on a 0–100 scale. Both effects
were statistically significant (p < .001).
- Country still mattered. Among human-operated
strikes, the Chinese strike was rated 9.0 points more
intentional than the U.S. strike (p < .001). The difference
between the two autonomy effects was not statistically significant
(p = .496).
- Less intentional did not clearly mean less
escalatory. Autonomy did not produce statistically clear
changes in perceived escalation likelihood or justifiability. These
findings concern perceptions in a hypothetical scenario; they do not
establish how governments would respond or whether deterrence would
weaken.
How intentional did the strike seem?
Average ratings were lower for autonomous systems in both country
scenarios.
Scroll each chart sideways to see all labels.
Unadjusted group means; lines show 95% confidence intervals. Higher
scores mean the strike was judged more intentional. All 2,935
respondents with an intentionality response are included.
What changed when the system was autonomous?
The estimates below compare autonomous with human-operated strikes
within each country scenario. Negative values mean lower ratings with
autonomy. Intervals crossing zero do not show a statistically clear
difference at the 5% level.
Adjusted estimates with 95% confidence intervals in brackets. Model
samples: intentionality, 2,833; escalation likelihood, 2,820; escalation
justifiability, 2,796. Escalation outcomes are exploratory. A
non-significant result is not proof of no effect.
Survey at a glance
| Respondents |
U.S. adults aged 18 or older; Verasight panel |
| Response dates |
July 31 to August 4, 2026 (start dates recorded in the
survey export) |
| Design |
Randomized 2 × 2 experiment: U.S. or Chinese strike ×
human-operated or autonomous system |
| Sample |
2,970 consented, assigned respondents; 2,935 with an
intentionality response; 2,833 with complete primary-model
covariates |
| Measures |
Intentionality (primary), escalation likelihood and
escalation justifiability (exploratory); 0–100 scales |
| Analysis |
Unweighted estimates; adjusted models use the full
report’s covariates and HC2 robust standard errors. Respondents are
analyzed as assigned, regardless of comprehension-check
performance. |
Sample and intentionality ratings by condition
|
Scenario
|
System
|
Assigned
|
Responses
|
Mean
|
SD
|
|
U.S. strike
|
Human-operated
|
742
|
733
|
62.8
|
24.5
|
|
U.S. strike
|
Autonomous AI
|
743
|
728
|
54.2
|
25.4
|
|
Chinese strike
|
Human-operated
|
742
|
737
|
72.1
|
23.0
|
|
Chinese strike
|
Autonomous AI
|
743
|
737
|
62.1
|
23.7
|
Responses are non-missing intentionality ratings; SD is the standard
deviation. 35 assigned respondents lacked that rating. Among the 2,935
included respondents, 73.0% correctly identified the country, 80.8%
identified the system type, and 65.9% answered both correctly; missing
checks count as incorrect.
Adjusted models account for AI comfort, relative ratings of Chinese and
U.S. AI effectiveness, party identification, age, gender, education,
income, race, Hispanic identity, and metropolitan status. This is a
survey of perceptions, not observed behavior in a real crisis.
