The Everytown Evidence Engine: Harnessing AI To Identify Gun Safety Policy Solutions

Every year, nearly 47,000 people in the United States are killed with guns and nearly 97,000 more are shot and wounded. The scale of this crisis is staggering and calls for the deep inquiry and urgent advancement of solutions that can meaningfully address this public health crisis, including the critical role of gun laws. Since the turn of the century, states across the U.S. have experimented with both expansion and rollback of gun laws. Given the constant flux and subtle differences between these laws, measuring the impact of policy variations remains both difficult and time consuming. While scientific research is increasingly demonstrating that certain laws can make our communities safer (or, in some cases, less safe), far more work is needed to assess the strength of existing laws over time and across geography. At the same time, gun violence research remains among the most persistently underfunded areas of public health inquiry, despite decades of policy debate and the scale of human loss. Therefore, efforts to fill gaps in knowledge remain critical for advancing evidence-informed policy solutions.  

Artificial Intelligence (AI) holds promise for more readily understanding how gun safety laws have shaped patterns of gun violence. By pairing carefully compiled records of both gun safety statutes and official gun death data and by leveraging the efficiency and speed of AI, researchers may be able to a) more quickly obtain crucial understanding of policy trends and their associations with gun violence over time and across geography and b) ultimately glean new insights into patterns of policy implementation and gun violence that traditional social science analytic techniques alone may not allow. For example, which laws are most strongly associated with specific gun violence outcomes and how have these associations changed (or not) over time? For which type of firearm violence is each gun safety law most effective?  How do the harms of gun violence differ across gender, age, and racial or ethnic groups? And can new patterns, trends, or connections be uncovered to enable new breakthroughs and insights? In short, efficient systems for analysis can lead to new questions and new answers in the field of gun violence prevention research. 

INTRODUCING E3

The Everytown Evidence Engine (E³) is a new tool developed by Everytown Labs that leverages developments in AI to deepen our understanding of gun violence and gun safety laws in the U.S. It highlights evidence-based policy solutions to America’s gun violence crisis and helps set the record straight using objective statistics on what is working. E³ has been created to provide clear answers to researchers seeking to better understand the impact of specific policies on gun violence.  

E³ was built using Claude, an AI system developed by the research company Anthropic. It has been configured to pair rigorously-vetted legal research on 146 gun safety law variations enacted in U.S. states since 1999 with official Centers for Disease Control and Prevention (CDC) gun death data since 2001. The use of AI enables the efficient synthesis of this considerable volume of information. It allows researchers to answer simpler descriptive questions about these laws (such as how many there are in each state, which groups of laws saw the greatest increase over time) and more complex questions (such as measuring correlation between gun death rates in a state following major law enactment).  It should be noted that while AI dramatically increases efficiency, human oversight remains essential. Legal and social science experts must remain in the loop to make sure this tool is developed accurately, safely, and ethically. 

As such, the E³ is currently launching in a beta phase with researchers and domain experts who are helping us thoughtfully test, evaluate, and strengthen this new system in real-world settings.This early collaboration will guide ongoing enhancements to the tool’s analytical depth, transparency, and usability. Over time, we plan to expand its datasets, refine its modeling approaches, and deepen its capacity to support rigorous, evidence-driven policymaking. We include a few examples of what E³‘s output currently entails in response to common research questions alongside details on how to access this new tool.

CAREFULLY-VETTED DATA OVER 25 YEARS

Since gun violence laws are most frequently enacted and carried out at the state level, the E³ has been built using two major state-level datasets: 

GUN SAFETY LAWS:  E³ contains data on 146 gun safety laws, each measuring whether a unique law was present or absent in each US state from 1999 through 2023. This new dataset was developed and coded by a team of lawyers and policy researchers at Everytown for Gun Safety Support Fund using publicly-available state statutes. 

States often address gun policy in very different ways. For example, laws to stem the tragic deaths of women shot and killed by an intimate partner is an urgent priority in some states. But the legal provisions enacted can take many forms. This database has therefore captured several dozen variables about individual laws to protect women from intimate partner homicide, including at what stage (temporary or final restraining orders) people are prohibited from possessing firearms, whether judges are authorized to order people with a restraining order to turn in their firearms, and more. 

This granular, state-by-state information on the landscape of gun laws across the United States is a new and invaluable resource for filling critical gaps in our understanding of how different laws are associated with  gun violence. Specifically, this database contains binary codes to indicate whether each of the 146 tracked gun safety laws exist across each state annually from 1999–2023. Laws are coded as present if enacted anytime that year, and absent the year after they are repealed. Gun permissive laws are reverse coded, for more intuitive analysis and interpretation. Laws in the dataset are organized into six themes: Foundational Laws (e.g., background checks); Firearm Industry and Product Safety (e.g., assault weapons prohibitions); Firearms in Public (e.g., prohibitions of concealed carry guns on college campuses); Keeping Firearms out of the Wrong Hands (e.g., minimum age to purchase); Policing and Civil Rights (e.g., law enforcement requirement to collect use of force data); and Sales & Permitting (e.g., training requirements). Please note that this database is affiliated with a manuscript currently under peer review. For a complete list of variables, please contact [email protected].

GUN MORTALITY DATA: The Centers for Disease Control and Prevention releases data on gun deaths by state and year and for major types of violence, including gun homicide, suicide, and unintentional shootings here. Gun mortality data is available by gender, race/ethnicity, and age. The E³ contains this data from 2001 through 2023. 

COVARIATES: Specific examples of state-level data drawing from the American Community Survey have also been built into the current version of the E³ to demonstrate its capabilities incorporating key covariates that researchers would need to be able to rigorously evaluate gun law policy and its associations with gun violence prevalence across the nation. These include the following variables, available for every state and for each year beginning in 2005 and through 2023: percent of population that has reached 18 years, the percent of population that is Non-Hispanic White, the percent of population that is male, the percent of population that has a high school degree, the mean unemployment rate, and median household income.

LIMITATIONS

E³ is a breakthrough in newly demonstrating the potential of AI to offer efficient, unbiased answers to critical gun safety law questions using rigorously-vetted data. This tool cannot, however, replace the value of in-depth studies that address the full complexity of public policy change and its resulting impacts in a specific place. For example, at this time, E³ does not currently weigh all of the factors that could be influencing gun violence such as gun ownership, employment and earnings, strength of policy implementation and enforcement, law-enforcement practices, and many other relevant and granular socioeconomic and demographic characteristics particularly at the county- and/or neighborhood-levels. At this time, the E³ also does not include additional local public health or law enforcement data sources that describe the presence of other violence-prevention interventions, programs, and investments. That being said, this new tool can provide users with important directions regarding policy effectiveness that can be used for critical decisionmaking. And it is the hope that future iterations of the E³ will incorporate these kinds of variables and, ultimately, increase its ability to conduct additional types of analyses.

CONCLUSION

AI is a tool that, when thoughtfully developed and employed, can increase and heighten human capacity to make sense of large amounts of data and provide new insights into complex social problems. Using reliable data collected by experts through rigorous, objective methods, coupled with a chatbot that excels at identifying patterns and streamlining processes, we can learn about the common-sense public safety laws that are helping to make sure people who purchase guns are trained in safety techniques and that we are keeping guns out of sensitive places and the hands of people with dangerous histories. 

WORKING WITH THE EVERYTOWN EVIDENCE ENGINE

If you are a researcher in the field of gun violence prevention and interested in testing the E³ 1.0, please email the Everytown Labs team at [email protected] with your name, institution, and a brief description of how you are thinking of using this and we will follow up with additional details.

SAMPLE RESEARCH QUESTIONS

Below find three sample research questions that were input into the E³ and the corresponding output. Please note these examples were generated in April 2026.

Question 1: Assault Weapons Bans and Firearm Homicide (2023)

Research Question:

What is the association between the presence of a state-level assault weapons ban and the prevalence of firearm homicide death rates in 2023? 

Evidence Engine Response:

PLAIN ENGLISH SUMMARY

In 2023, states with assault weapons bans had firearm homicide rates that were 32.2% lower than states without such bans (4.45 vs 6.57 per 100,000), though this difference was not statistically significant (p = 0.13) in the single-year comparison, due in part to the lack of statistical power (only 9 states had a ban in 2023).

However, when pooling all available data from 2001-2023, the pattern becomes clear and statistically robust: state-year observations with assault weapons bans had firearm homicide rates that were 31.9% lower than observations without such bans (3.32 vs 4.88 per 100,000). This pooled difference is highly statistically significant (p < 0.001).

STATISTICAL RESULTS AND INTERPRETATION

### 2023 Cross-Sectional Analysis (Underpowered)

A Welch two-sample t-test was used to compare firearm homicide rates between states with and without assault weapons bans in 2023. Six states with suppressed age-adjusted rates were excluded from this analysis.

Results:

• Mean rate WITH AWB: 4.45 per 100,000 (N=9 states, SD=2.37)
• Mean rate WITHOUT AWB: 6.57 per 100,000 (N=41 states, SD=3.91)
• Difference: -2.12 per 100,000 (32.2% lower)
• t-statistic: -2.054
• p-value: 0.053 (not statistically significant)
• 95% CI for difference: [-4.26, 0.03]
• Effect size (Cohen’s d): -0.577 (medium effect)

Interpretation: The 2023 analysis shows a substantial and consistent direction of effect—states with assault weapons bans have meaningfully lower firearm homicide rates—but the finding narrowly misses conventional significance (p = 0.053). This is primarily due to low statistical power with only 9 states having assault weapons bans and high variance in the comparison group.

States with AWB in 2023: California, Connecticut, Delaware, Illinois, Maryland, Massachusetts, New Jersey, New York, Washington

### Pooled Analysis (2001-2023) – Adequately Powered

A Welch two-sample t-test was used to compare firearm homicide rates across all state-year observations from 2001–2023, excluding observations with suppressed age-adjusted rates.

Results:

• Mean rate WITH AWB: 3.32 per 100,000 (N=130 observations, SD=1.85)
• Mean rate WITHOUT AWB: 4.88 per 100,000 (N=828 observations, SD=2.78)
• Difference: -1.56 per 100,000 (31.9% lower)
• t-statistic: -8.247
• p-value: p < 0.001 (highly statistically significant)
• 95% CI for difference: [-1.93, -1.19]
• Effect size (Cohen’s d): -0.583 (medium effect)

Interpretation: When pooling data across all available years, we find strong and highly statistically significant evidence that states with assault weapons bans have substantially lower firearm homicide rates—about 1.56 fewer deaths per 100,000 people annually, representing a 31.9% reduction. This finding is statistically robust (p < 0.001) and represents a medium effect size.

The total sample includes 130 state-year observations with assault weapons bans (9 states observed across multiple years) and 828 observations without bans (43 states observed across multiple years). It should be noted that some states adopted an AWB during the study period and are therefore included in both groups. State-year observations with suppressed age-adjusted rates were excluded throughout.

LIMITATIONS

• This finding reflects an association, not a proven causal effect. States with assault weapons bans may differ in other ways that also impact gun violence rates. For example, states that enact assault weapons bans may also have other restrictive gun laws, different demographic compositions, or different cultural attitudes toward firearms.

• This answer does not weigh all factors that could be impacting gun violence. Other factors such as demographics (age, race, urbanization), economic conditions (poverty, unemployment), law enforcement practices, and concurrent policies (background checks, permit requirements, etc.) may contribute to observed patterns.

• The 2023 cross-sectional analysis is underpowered. With only 9 states having assault weapons bans, the single-year comparison lacks sufficient statistical power to detect a reliable difference, even though the direction and magnitude of the effect are consistent with the pooled analysis.

• The pooled analysis combines data across 23 years, which may mask temporal trends. Firearm homicide rates have fluctuated substantially over this period nationally, with declines in the 2000s-2010s followed by increases after 2019. The pooled estimate represents an average effect across this entire period.

Question 2: Secure Storage Laws and Firearm Suicide (Adjusted)

Research Question:

Do secure storage laws significantly predict firearm suicide death rates, accounting for the impact of key socioeconomic indicators?

Evidence Engine Response:

PLAIN ENGLISH SUMMARY

Yes, secure storage laws significantly predict lower firearm suicide death rates after accounting for key socioeconomic indicators.

After controlling for state demographics (age structure, gender composition, race/ethnicity), socioeconomic factors (education levels, unemployment rate, median household income), and temporal trends from 2005 to 2023, states with secure storage laws have firearm suicide rates that are approximately 1.51 per 100,000 lower than states without such laws—representing about a 13% reduction.

The single-year 2023 cross-sectional analysis showed a similar effect size (-1.84 per 100,000) but was underpowered (p = 0.099). The pooled panel analysis across 19 years (2005-2023) with 947 state-year observations confirmed this association is statistically significant (p = 0.0077).

STATISTICAL RESULTS AND INTERPRETATION

### 2023 Cross-Sectional Analysis (Underpowered)

A multiple linear regression was used to evaluate the association between secure storage laws and firearm suicide rates in 2023, controlling for six socioeconomic covariates.

Result: β = -1.840 per 100,000; 95% CI: [-4.041, 0.361]; p = 0.099

Interpretation: Not statistically significant at conventional levels (p < 0.05), though the effect size was comparable to the panel analysis. The single-year sample (N = 50 states) lacked sufficient statistical power.

Model fit: R² = 0.667, Adjusted R² = 0.612, F(7, 42) = 12.02, p < 0.001

### 2005-2023 Pooled Panel Analysis (Adequately Powered)

A pooled panel regression with year fixed effects and state-clustered standard errors was used to evaluate the relationship over 19 years, controlling for the same six covariates plus temporal trends.

Result: β = -1.511 per 100,000; 95% CI: [-2.623, -0.400]; p = 0.0077

Interpretation: Statistically significant (p < 0.01). States with secure storage laws have firearm suicide rates approximately 1.51 per 100,000 lower than states without such laws, after adjusting for demographics, socioeconomic conditions, and year-to-year trends.

Model fit: R² = 0.712, Adjusted R² = 0.704

Sample: N = 947 state-year observations (50 states × 19 years, excluding observations with missing data)

Standard errors: Clustered by state to account for within-state correlation over time

LIMITATIONS

• Association, not causation: This finding reflects an association, not a proven causal effect. States with secure storage laws may differ in other ways that also impact gun violence rates (e.g., gun culture, other policies, enforcement practices).

• Unmeasured confounders: This answer does not weigh all factors that could be impacting gun violence. Other factors beyond the included covariates (e.g., mental health service availability, firearm ownership rates, concurrent gun policies, cultural factors) may contribute to observed patterns.

• Law heterogeneity: The secure storage law variable captures whether gun owners are held accountable when children can/do access unsecured firearms, but states may vary in enforcement intensity, penalty severity, and implementation details not captured in this binary measure.

• Limited temporal variation: Only 7 states adopted secure storage laws between 2005 and 2023 (rising from 18 to 25 states), limiting the ability to estimate within-state effects over time. The panel model primarily leverages cross-state variation.

• Ecological analysis: This state-level analysis cannot determine individual-level effects. The association may reflect ecological confounding or aggregation effects.

Question 3: Universal Background Checks and Firearm Homicide During COVID-19

Research Question:

Following the onset of COVID-19, what is the prevalence of firearm homicide deaths across all states without universal background checks? How does this compare to the prevalence of firearm homicide deaths across all states with universal background checks for the same time period?

Evidence Engine Response:

PLAIN ENGLISH SUMMARY

Following the onset of COVID-19 (2020-2023), states without universal background check (UBC) laws had firearm homicide death rates that were 37.8% higher than states with such laws. Specifically:

• States WITHOUT universal background checks: averaged 7.40 firearm homicide deaths per 100,000 people
• States WITH universal background checks: averaged 5.37 firearm homicide deaths per 100,000 people

STATISTICAL RESULTS AND INTERPRETATION

Statistical Test: An independent samples t-test was used to compare mean age-adjusted firearm homicide rates between state-years with and without UBC laws during 2020-2023.

Test Results:

• Mean difference: 2.03 per 100,000 (37.8% higher in non-UBC states)
• t-statistic = 3.551, p = 0.000; Statistically significant (p < 0.001)
• 95% Confidence Interval for difference: [0.96, 3.09] per 100,000
• Effect size (Cohen’s d) = 0.536 (medium effect)

Interpretation: States without universal background check laws had significantly higher firearm homicide rates during the COVID-19 era compared to states with these laws in place. The difference of 2.03 per 100,000 represents a substantial and statistically significant association. The effect size (Cohen’s d = 0.536) indicates a medium effective size.

LIMITATIONS

• Association, not causation: This finding reflects an association, not a proven causal effect. States with such laws may differ in other ways that also impact gun violence rates (e.g., states with UBC laws may also have other gun safety policies, different demographic characteristics, or different social conditions).

• Unmeasured confounders: This answer does not weigh all factors that could be impacting gun violence. Other factors (e.g., demographics, economic conditions, concurrent policies, policing strategies, COVID-19 impacts on crime patterns) may contribute to observed patterns.

• COVID-19 context: The COVID-19 pandemic created unusual social and economic conditions (2020-2023) that may have affected firearm violence patterns differently across states. The observed differences may not generalize to non-pandemic periods.

• Law heterogeneity: The “universal background check” category represents a single policy dimension, but states’ UBC laws may vary in specific provisions, enforcement mechanisms, and implementation quality.

• Limited temporal data: Several states changed their UBC status during this period (Iowa repealed, Minnesota and Hawaii enacted, North Carolina repealed, Rhode Island repealed), which provides some policy variation but limits the ability to assess long-term effects.