Few domains sit closer to the intersection of technological capability and ethical risk than public safety. Police forces, fire services, border agencies, and emergency responders are deploying AI across dozens of use cases — some demonstrably beneficial, others deeply controversial. The difference between the two often comes down to governance, not technology.
This guide covers the five main areas where AI for public safety is being deployed, the specific prohibitions and restrictions under the EU AI Act, and what organisations need to build internally before adopting these tools.
80%
of large law enforcement agencies in Europe and North America have piloted or deployed at least one AI system, yet fewer than 30% have a formal AI governance framework in place
Source : UNICRI Centre for AI and Robotics, 2025
Where AI is being used in public safety
1. Emergency response and dispatch
The least controversial and most immediately impactful application of AI in public safety is emergency response optimisation. AI systems analyse incoming emergency calls in real time — detecting urgency signals, classifying incident types, and recommending resource allocation before a human dispatcher has finished taking the call.
In the UK, several ambulance trusts use machine learning models to predict demand surges based on weather patterns, public events, historical call volumes, and even flu surveillance data. The result is faster response times and better resource positioning. Similar systems are deployed across Scandinavian emergency services, where AI-assisted dispatch has reduced average response times by 12-18% in pilot regions (European Emergency Number Association, 2025).
Fire services use AI for risk mapping — identifying buildings, neighbourhoods, and industrial sites with the highest fire probability based on structural data, inspection records, and environmental factors. New York City’s FDIC pioneered this approach, and versions now operate in London, Berlin, and Amsterdam.
These applications share a common trait: they augment human decision-making without replacing it, and they do not directly affect individual rights. That changes rapidly as we move into policing.
2. Predictive policing
Predictive policing uses historical crime data, geographic information, and statistical models to forecast where crimes are likely to occur or, more controversially, who is likely to commit them. The distinction between place-based and person-based prediction is critical.
Place-based prediction — identifying high-risk locations and time windows — is the less contentious form. It functions as a sophisticated version of what experienced officers do intuitively: directing patrols to areas where incidents are statistically more likely. Tools like PredPol (now Geolitica) and similar platforms are used by police forces across the US, UK, and Europe.
Person-based prediction — flagging individuals as potential offenders based on algorithmic scoring — is far more problematic. These systems inherit and amplify the biases embedded in historical policing data. If a neighbourhood has been over-policed for decades, the data will reflect that over-policing, not the true distribution of criminal activity. The result is a feedback loop that concentrates enforcement on already marginalised communities.
The EU AI Act classifies individual predictive policing — AI systems that assess the risk of a natural person committing a criminal offence solely based on profiling or personality traits — as a prohibited practice under Article 5. Agencies deploying person-based prediction within the EU face significant legal exposure. Place-based prediction remains permissible but falls under high-risk obligations.
Several major forces have abandoned predictive policing programmes entirely. The Los Angeles Police Department ended its PredPol contract in 2020 following audit findings of racial bias. The Dutch National Police discontinued its crime anticipation system after similar concerns. For organisations still considering these tools, a rigorous AI risk assessment is non-negotiable.
3. Facial recognition and biometric surveillance
Facial recognition technology (FRT) is perhaps the most publicly debated AI application in public safety. Law enforcement agencies use it for three broad purposes: identifying suspects from CCTV footage after an incident, real-time scanning of crowds at public events, and verifying identities at border crossings.
The accuracy of modern facial recognition systems has improved dramatically — top systems now achieve error rates below 0.2% on controlled benchmark datasets (NIST Face Recognition Vendor Test, 2025). But benchmark performance does not translate directly to real-world performance. Lighting conditions, camera angles, image quality, and demographic variations all affect accuracy. Multiple independent studies have confirmed higher error rates for women, darker-skinned individuals, and older adults.
The EU AI Act draws a sharp line. Real-time remote biometric identification in publicly accessible spaces for law enforcement is prohibited, with narrow exceptions for finding missing children, preventing imminent terrorist threats, and locating suspects of serious crimes — and even these exceptions require prior judicial authorisation. Post-incident identification (reviewing recorded footage) is classified as high-risk and subject to stringent requirements.
The UK has taken a more permissive approach. The Metropolitan Police and South Wales Police have conducted live facial recognition deployments at public events, though legal challenges continue. The UK’s evolving regulatory framework relies on existing laws — the Equality Act, Data Protection Act, and Human Rights Act — rather than AI-specific legislation, creating a patchwork that many legal scholars consider insufficient.
4. Surveillance and intelligence analysis
Beyond facial recognition, AI powers a broader surveillance ecosystem: automated number plate recognition, social media monitoring, communications metadata analysis, anomaly detection in sensor networks, and drone-based surveillance. Intelligence agencies and counter-terrorism units use AI to process volumes of data that would be impossible to analyse manually.
The value is real. AI-assisted intelligence analysis has contributed to disrupting planned attacks, identifying trafficking networks, and locating missing persons. But the same capabilities that make AI valuable for security make it dangerous for civil liberties when deployed without adequate oversight.
The critical governance question is not whether these tools are technically effective — most are — but whether the oversight mechanisms match the power of the technology. This requires robust AI governance frameworks, independent auditing, judicial oversight for intrusive capabilities, and meaningful transparency about what systems are deployed and how they are used.
35+
cities worldwide have enacted partial or complete bans on government use of facial recognition technology, including San Francisco, Brussels, and several Australian municipalities
Source : Access Now, 2025
5. AI in the justice system
AI is increasingly used downstream of policing: in bail and sentencing decisions, parole assessments, and case management. The most notorious example remains COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), the US risk assessment tool that ProPublica found in 2016 was twice as likely to falsely flag Black defendants as future criminals compared to white defendants.
Since then, the debate has matured but the fundamental tension remains. AI bias in the workplace is concerning; AI bias in the justice system can deprive people of their liberty. The EU AI Act classifies AI systems used in the administration of justice and democratic processes as high-risk, requiring conformity assessments, human oversight, and transparency obligations.
What the EU AI Act prohibits and restricts
For public safety organisations operating within or interacting with the EU, the regulatory framework is specific and consequential.
Prohibited practices (Article 5):
- Social scoring by public authorities
- Real-time remote biometric identification in public spaces (with narrow exceptions)
- AI systems that assess the risk of criminal offending based solely on profiling
- Emotion recognition in law enforcement and border management contexts
- Untargeted scraping of facial images from the internet or CCTV to build recognition databases
High-risk applications (Annex III):
- Biometric identification (non-real-time)
- Critical infrastructure management
- Law enforcement tools for individual risk assessment
- Migration, asylum, and border control management
- Administration of justice
Compliance is not optional. For a comprehensive overview of what these obligations entail, see our EU AI Act guide. Organisations deploying high-risk systems must implement documented risk management, data governance, human oversight, and transparency measures — and must ensure all staff involved have sufficient AI literacy under Article 4.
Building responsible AI capability in public safety
Start with governance, not technology
The pattern across failed public safety AI deployments is consistent: the technology was procured before the governance was established. Agencies need a clear AI governance framework before evaluating any vendor or system. This includes risk classification criteria, approval processes, bias testing protocols, incident response procedures, and accountability structures.
Train every level of the organisation
Frontline officers need to understand what AI tools can and cannot do, and when to override automated recommendations. Investigators need to understand the limitations of facial recognition and predictive analytics as evidence. Senior leaders need to understand their legal obligations under the EU AI Act and equivalent frameworks. An AI competency framework tailored to public safety roles is essential.
Engage the public
Public safety AI is unique because it directly affects citizens who have no choice in the matter — unlike a customer who can switch banks, a citizen cannot opt out of being policed. This creates an elevated obligation for transparency and consultation. Publish what systems you use, how they work, and how decisions can be challenged.
Audit continuously
AI systems in public safety should be subject to independent auditing — not just at procurement, but on an ongoing basis. Bias testing, accuracy monitoring, and impact assessments should be regular, documented, and accessible. The NIST AI Risk Management Framework provides a structured approach.
Whether your organisation is a metropolitan police force, a border agency, or a national security body, the regulatory direction is clear: AI in public safety will be among the most heavily regulated applications globally. Building governance capability now is not just risk management — it is operational readiness.
Preparing your public safety workforce
The challenge for public safety organisations is not access to AI technology — vendors are eager to sell it. The challenge is building the internal capability to evaluate, govern, and operate AI systems in a domain where the consequences of failure are measured in civil liberties, not quarterly earnings.
Brain provides AI training designed for public safety and government organisations — practical, role-based modules covering AI literacy, EU AI Act compliance, risk assessment, and trustworthy AI governance. Compliance documentation that meets Article 4 requirements and audit standards.
Related articles
AI for Defence: Applications, Ethics & Governance
Navigate AI in defence responsibly. Covers intelligence analysis, logistics, autonomous systems, cybersecurity, and EU AI Act military exemptions.
AI for Education: 5 Tools Reshaping Learning in 2026
Transform teaching and learning outcomes with AI. Covers personalised learning, automated assessment, plagiarism detection, and teacher productivity tools.
AI for Government: 5 Public Sector Use Cases (2026)
Improve citizen services and cut fraud with AI. Covers policy analysis, procurement automation, and workforce planning for public sector organisations.