Keywords

Introduction: The ALIGNER Methodology

The world is changing at an unprecedented rate, and artificial intelligence (AI) is at the forefront of this change. While providing numerous benefits, many have raised concerns over the impact AI has or will have on matters such as security. The EU-funded ALIGNERFootnote 1 project aims to unite European actors who have concerns about AI, law enforcement, and policing to jointly identify and discuss how to enhance Europe’s security whereby AI strengthens law enforcement agencies (LEAs) while providing benefits to the public.

To achieve this goal, ALIGNER has established expert groups from policing and law enforcement, civil society, policymaking, research, and industry, who work together in regular workshops to identify opportunities and risks arising from law enforcement use of emerging AI technologies, which capability enhancement needs are associated with the increased use of AI (both by LEAs and criminals), and which ethical, legal, and technical/operational impacts the deployment of AI technologies by LEAs might have.

To provide a basis for the work of ALIGNER’s expert groups, a sound methodological approach is necessary for identification of promising (emerging) AI technologies for LEAs and their potential implications in the technological, ethical, legal, and organisational dimensions at LEAs. To the knowledge of the authors, such a comprehensive, integrated assessment approach for AI technologies specific for law enforcement is currently missing. ALIGNER fills this gap by pursuing a collaborative technology watch process, followed by three assessments, as shown in Fig. 18.1. In particular, the assessment phase begins with the AI Technology Impact Assessment (section “AI Technology Impact Assessment”); if this leads to significant scores, the Risk Assessment follows (section “Risk Assessment”); if this leads to significant scores, the assessment phase ends with the Fundamental Rights Impact Assessment (section “Ethical and Legal Assessment”). These assessment approaches are described in detail in the following sections.

Fig. 18.1
A process diagram of the technology watch and assessment starts with needs, scope setting, and brainstorming. It proceeds to a prioritization workshop, scanning activities, narrative activity, workshop planning, assessment workshop, scenario cards, assessment, and ends with investments.

The ALIGNER technology watch and assessment process. In grey: Technology watch process; in white: the assessment process

Full size image

AI Technology Impact Assessment

The first step in assessing law enforcement AI technologies is conducting the AI Technology Impact Assessment. By doing this evaluation, LEAs can identify which AI technologies they are most interested in using and prioritise the ones to consider for research or investment.

The ALIGNER AI Technology Impact Assessment and its associated criteria have been inspired by an existing method for assessing technologies [1], which already combined an assessment of added value and feasibility to recommend technology investments. However, the ALIGNER AI Technology Impact Assessment further expands this approach: as both added value and feasibility are assessed on discrete and non-linear ordinal scales rating 1–4, the investment recommendations will result from discrete combinations of these assessments. This allows more detailed recommendations, reflecting different levels of potential investments.

The AI Technology Impact Assessment consists of three steps: (i) assessing the added value; (ii) assessing the feasibility of an AI technology; and (iii) recommending AI technology investments [2].

The goal of the added value evaluation is to estimate to what degree an AI capability will improve the LEA capacity. It is assessed based on six criteria: need; comparative effectiveness; beneficial side effects; flexibility; future proofness; and future additional potential. Each criterion is assessed on a scale from 1 to 4 by each LEA stakeholder participating in the assessment. The stakeholders’ ratings of each criterion are then considered to determine an overall score of added value. LEAs may decide how to combine each individual’s rating to one overall and agreed score.

The goal of the feasibility assessment is to estimate how likely it is that the AI technology will work as intended and/or advertised. It is assessed based on eight criteria: blocking factors; comparative costs; reliability of cost estimate; type of investment; technical feasibility; usability; credibility; and acquisition and development incentives. Similar to the added value assessment, each criterion is assessed on a scale from 1 to 4 by each stakeholder participating in the assessment and thereafter combined to one agreed and overall score of feasibility.

Example, Workshop Session ALIGNER

The assessment conducted during a workshop in the project ALIGNER can serve as an example for how to determine an overall score. First, each stakeholder assesses each criterion individually by holding up 1 to 4 fingers. If the discrepancy of the highest and lowest score in the group is more than one step, those participants put forward arguments to motivate their stance. Then, the voting process is repeated. The average score is used for added value and feasibility. However, if significant differences still exist, the participants are encouraged to document each side’s strongest argument.

The overall score of added value and feasibility are used to recommend AI technology investment decisions. Figure 18.2 shows how the rating for feasibility combined with the rating for added value gives an output L1–6.

Fig. 18.2
A table depicts references for A I technology investment with added value versus feasibility. The overall scores of added value and feasibility are used to recommend A I technology investment decisions.

Reference table for recommended AI technology investment [2]

Full size image

As shown in Fig. 18.3, the L-number provides guidance in whether to prioritise the research or development of an AI technology or not. The higher the L-number, the more prioritised the research or investment should be.

Fig. 18.3
A table depicts A I technology investment decisions for T R L 1-6 and T R L 7-9. The L-number guides whether or not to prioritize the research or development of an A I technology. A higher L-number indicates a higher priority for research or investment.

Legend for recommended AI technology investment decision [2]

Full size image

Risk Assessment

Although AI technologies provide many benefits for LEAs and society in general, there is an urge to evaluate the associated risks. Therefore, AI technologies that have achieved a significant score in the AI Technology Impact Assessment will then go through the Risk Assessment.

The ALIGNER Risk Assessment instrument is still under interdisciplinary development between the technology, ethics, and legal experts of the project, but its main idea is shortly described below.

The Risk Assessment will pursue two main objectives. First, it will identify the risks or unintentional unwanted impacts, posed by AI in the context of law enforcement. Additionally, it will support AI-based technologies improving the EU’s resilience against emerging, ‘classical’ and ‘new’ AI-supported threats. Consequently, the capacities of LEAs at national and at EU level will be reinforced.

By using a combined methodological approach, applying research literature, AI policies, and AI regulations, the objective is to integrate the latest technical research and best practices with European law. To do so, the first step is conducting a review of the risks associated with development of AI technologies, while the second step consists of a review of the existing instruments for AI risk assessment and technical and organisational measures for risk mitigation. This will result in the ALIGNER Risk Assessment instrument that will be carried out during future project workshops.

Ethical and Legal Assessment

If both the AI Technology Impact Assessment and the Risk Assessment have obtained significant scores, the AI technology is subject to a final ethical and legal assessment. This is performed using the ALIGNER Fundamental Rights Impact Assessment (AFRIA) templates [3].

The AFRIA is a tool addressed to LEAs aiming to deploy AI technologies for law enforcement purposes within the European Union. The AFRIA is a reflective exercise, seeking to enhance the ethical and legal governance systems of LEAs. Hence, the AFRIA has two main functions. First, it helps LEAs identify and mitigate the impact of the deployment of a certain AI technology on ethical principles and those fundamental rights of individuals most likely to be impacted. Second, it is a suitable instrument for LEAs to explain and record their decision-making processes.

LEAs should perform a single AFRIA for a single AI technology, deployed for a single law enforcement purpose or connected law enforcement purposes.

An AFRIA should be performed by LEAs prior to the deployment of the AI technology, to inform the decision-making process on the if, when, why, and how of the deployment. However, performing an AFRIA should be considered as an iterative process: the AFRIA needs to be recorded, reviewed, and updated throughout the whole lifecycle of the AI technology, to reflect eventual changes in the functioning of the technology and/or its circumstances of deployment.

To perform an AFRIA, LEAs should establish a diverse and multidisciplinary team. This should include members of the organisation with legal, operational, and technical expertise.

The AFRIA consists of two different, but connected, templates: the Fundamental Rights Impact Assessment (section “Fundamental Rights Impact Assessment Template”) and the AI System Governance (section “AI System Governance Template”).

Fundamental Rights Impact Assessment Template

The Fundamental Rights Impact Assessment template helps LEAs identify and assess the impact that the AI technology may have on the fundamental rights of individuals.

The template focuses on those fundamental rights most likely to be impacted by law enforcement AI [4], i.e., presumption of innocence and right to an effective remedy and to a fair trial; right to equality and non-discrimination; freedom of expression and information; and right to privacy and data protection. Accordingly, the template is divided into four sections and, in each one of them, a group of fundamental rights is used as benchmark for the following assessment.

As shown in Fig. 18.4, each section of the template is divided into three columns. The first column lists some ‘challenges’, namely some possible characteristics embedded in the assessed AI technology that may have a negative impact on the considered fundamental right. In the second ‘evaluation’ column, LEAs need to precise whether and, if so, to what degree the listed challenges relate to the assessed AI technology. In the last ‘estimated impact level’ column, LEAs need to estimate the level of negative effect the deployment of the AI technology may have on the considered fundamental right. To do so, LEAs should use an impact matrix based on two factors: the severity of the prejudice and the number of affected individuals.

Fig. 18.4
A table of three columns. A. Challenges of A I technology that may impact fundamental rights. B. Evaluation, requires L E As to specify if and to what degree challenges apply. D. Estimated impact level, assesses the negative effect on fundamental rights.

Section of Fundamental Rights Impact Assessment template

Full size image

AI System Governance Template

The AI System Governance template helps LEAs identify, explain, and record possible measures to mitigate the negative impact that the deployment of the AI technology would have on AI ethics principles and fundamental rights of individuals.

The template relies on the seven key requirements that a trustworthy AI technology should fulfil, as identified by the high-level expert group on artificial intelligence [5]. The requirements are human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being; and accountability. Accordingly, the template is divided into seven sections and, in each one of them, a key requirement for trustworthy AI is used as benchmark for grouping the minimum standards an AI technology should achieve.

As shown in Fig. 18.5, each section of the template is divided into seven (groups of) columns. In the first ‘components’ column, the building blocks of the considered key requirements are listed. The second column reports some ‘minimum standards’, namely characteristics an AI technology should embed or possible governance procedures the organisation should implement for the deployment of the technology to be considered trustworthy. The third group of columns, ‘initial impact estimate’, connects the minimum standard with previously estimated challenges and impact levels. In the fourth ‘additional mitigation measures implemented’ column, LEAs need to precise how the minimum standard is implemented and how it mitigates the initial impact estimate. In the fifth group of column, ‘final assessment’, LEAs need to estimate the final level of negative effect the deployment of the AI technology may have on fundamental rights and list possible further actions to improve the implementation of the minimum standard. In the last two columns, LEAs need to specify the ‘responsible department’ and the ‘timeline’ for the implementation of the mitigation measures.

Fig. 18.5
A table presents a template section divided into seven columns, components, minimum standards, initial impact estimate, additional mitigation measures implemented, final assessment, responsible department, and timeline. These aid L E As in evaluating and improving A I technology impact.

Section of AI System Governance template

Full size image

Conclusion

Both the AI Technology Impact Assessment method and the Fundamental Rights Impact Assessment have already been successfully validated with stakeholders from ALIGNER’s expert groups, including practitioners from law enforcement, industry professionals, civil society representatives, and researchers. Similar validation exercises will be held for the Risk Assessment method, once fully operational.

The complete integrated assessment approach will be used to assess around 20 emerging AI technologies around different future scenarios that cover both the criminal misuse of AI as well as the use of AI by LEAs for the benefit of society. These scenarios include, amongst others: (i) identifying and countering disinformation and social manipulation; (ii) AI-supported cybercrime against individuals and organisations; (iii) the use of AI-supported vehicles, robots, and drones; and (iv) future organisation of LEAs with deeply integrated AI processes.