Support our work: become a Friend of Statewatch from as little as £1/€1 per month.
In this section
This annex compiles information from official studies and report on potential uses of AI technology for criminal justice, policing and migration purposes. There is no indication that all of these uses of AI technology may be employed. However, some of those included in the reports are incredible invasive and raise significant legal and ethical questions. They are listed here to give an indication of the potential scope for applying AI technologies in these policy areas.
Studies by Deloitte[1] and eu-LISA and Eurojust[2] have proposed a host of AI techniques that could be deployed for criminal justice purposes.
Legal/criminal procedures and investigations use a lot of what is known as ‘unstructured data’. Unstructured data is normally stored in its original form in ‘data lakes’ - a type of massive data storage. It is not organised or standardised. Types of unstructured data include audiovisual, geospatial and text data. Unstructured data can be collected quickly and stored in a variety of ways. However, processing unstructured data requires specific tools and expert data science knowledge.
Automated document processing comes in two main forms. The first is a technology known as “computer vision.” This allows a computer to ‘see’ and interpret images or visual information. When processing documents, the computer can ‘read’ text from images or scanned documents. This can then be used to convert read-only documents such as PDFs or scanned papers into text that can be edited and searched on a computer.
The second form of automated document processing is natural language processing (NLP). This technology lets computers ‘understand’ language, and produce it in a form that appears human. NLP allows computers to analyse the contents of a document and sort it accordingly into a user-friendly archive. It is also useful for extracting key information from (e.g. names, dates, addresses etc.) or providing summaries of documents.
These technologies allow to swiftly process large amounts of standardised documents.
Automated document processing can then allow more AI tools to be used, such as:
However, these tools must be used with extreme caution, especially in legal investigations, as the information extracted may be used as evidence. There is always a risk of error. Meaningful human oversight and decision-making is crucial.
This technology can translate material from one language to another, or even to/from multiple different languages. It can also assist in communication between parties that may speak different languages. This is particularly useful when analysing evidence that contains specialist terms, or is in a less common language.
To ensure that the translation system works well, it needs to be developed as a specialised system. This requires major resources and may be very expensive. Machine translated documents cannot be used as evidence.
However, automated translation is useful at the investigation stage in making evidence more accessible to the whole team. Machine translation provides valuable insight into which parts of the document are most important to have professionally translated (by a human) and sworn in as evidence. It also saves time and costs on sworn translations.
Summarisation technology can help condense large volumes of textual information, but cannot match human ability to interpret text. It can make information more accessible for further in-depth analysis by a human. Automated text summarisation has effectively been used in academic environments. Within the field of justice specifically, automated summaries produced by these systems may be incorrect.
In any case, any material made by automated systems will need to be approved by humans before being used as evidence. Users of these technologies cannot rely completely on the content produced by automated systems.
Natural language processing is currently mainly used in investigating ‘white collar crime’. These are types of crime that are essentially financially motivated and “non-violent” – for example, excise duty fraud or insider trading. Natural language processing techniques can help process digitised text that is often used as evidence in white collar crime cases: invoices, emails, contracts or shipping documentation, for example.
Natural language processing can also further develop communication tools between EU agencies to find more accurate links between cases. For example, the legislation regarding cooperation between Eurojust, Europol and the European Public Prosecutor’s Office (EPPO) introduces ‘hit/no hit’ search systems between the agencies. Natural language processing can strengthen this hit/no-hit function.
As well as being used for investigation and evidence analysis, named entity recognition and classification (NERC) can be used to protect documents/identities by anonymising or pseudonymising data (removing names or creating fake names).
Legal research provides information relevant to a case. It normally involves reviewing statutes or case law. Natural language processing allows legal professionals to search for relevant information, such as relevant statutes, related legislation, case law, or doctrinal opinion. For cases that take place across different nation-states and use multiple languages, tailored solutions may be required. There are some initiatives seeking to install AI tools within legal analysis in Europe.
Images recorded by CCTV cameras are often low quality, and therefore are of limited use by computer vision systems in criminal investigations. However, when the image is higher quality, computer vision systems are a valuable tool for identifying people of interest in recorded or live-streamed media. Specialised video/image enhancing algorithms have been developed to tackle the issue of image quality.
Machine learning algorithms are generally only considered robust when used in narrowly defined tasks. They are also only as good as the data they are trained on. The basic technology used in the analysis of images/video is exactly the same as the technology used for facial recognition by border control authorities or law enforcement. Therefore, biometric identification algorithms can be relatively easily used in forensic image/video analysis systems.
Biometric anonymisation techniques can be used to anonymise images or video containing biometric identifiers (i.e. human faces). This can be used to protect the identities of victims, witnesses etc. This technology can also be used to remove identifiers such as car license plates. Anonymisation techniques can also be used on audio evidence (such as telephone conversations/recordings) to conceal/distort voices. This can again preserve the privacy and protect the personal data of victims/witnesses.
The following potential uses of AI were included in a Deloitte report for the European Commission, based on surveys of and interviews with EU and member state officials.[3]
An AI chatbot (or virtual assistant) could be used to support individuals requesting a visa during the online application process. Potential uses:
The chatbot could be enhanced with multiple languages/translation ability to answer questions in the user’s native languages.
AI can be used to classify and sort visa applicants based on an initial AI assessment. This approach is similar to the process of triaging in healthcare where an AI can process scan images and doctors’ medical notes to allocate the appropriate next steps for a patient.
Classification categories could be defined based on a risk level or specific indicators. A more flexible/dynamic system could also group applications based on similarities it observes in the data, such as similar occupation, rather than rigidly defined categories.
As part of the risk assessment, the system could perform automatic searches within national and central databases (Schengen Information System, Visa Information System etc.) to check if there are any results that contribute to a certain risk level. The relevant response could be found in different ways depending on the system and the type of AI risk assessment model used. Automated searches of the EU’s databases with visa and travel authorisation applicants’ data are will be introduced under current EU legislation.
Consolidation of data from national and EU databases and other internal and external systems, to provide an overview and highlight key pieces of information which need to be checked manually.
An AI chatbot/virtual assistant can answer any questions the applicant has after they submit their documents. The chatbot can inform the applicant when action is needed to proceed to the next step in the application process.
An AI chatbot/virtual assistant can answer questions from applicants after a visa decision has been made. The chatbot could support applicants in finding the right answer or direct them to the appropriate person/body to speak to. It could also automate scheduling follow-up appointments.
AI can monitor, search and combine data from different sources such as the European Travel Information and Authorization System (ETIAS), the European Entry/Exit System (EES), and passenger name record (PNR) data from airlines. Using this, AI systems can detect ‘irregular’ travelling patterns. Irregularity can be detected from a sequence of stops or from the overall pattern of travel. AI could be used to identify patterns that were not previously observed as strange.
This resembles methods used in fraud detection to analyse spending behaviour, or in cybersecurity to analyse network traffic. AI is commonly used in both cases, particularly within the financial services industry.
AI could be used to create a personalised application form by tailoring questions to the applicant. Natural language processing could also be used for real-time personalisation by using information provided by the applicant as input data. Based on this, models could suggest questions - for instance, if a follow-up is likely based on historic factors.
A concrete example of this would be if the system logged a high chance of a traveller from a specific region overstaying. In this case the applicant could be asked more questions on why they are travelling to the Schengen area and asked to provide documents to support this. A human case worker would then be able to verify this.
It should be noted that this use of AI would raise questions of procedural fairness.
An AI model could be used to predict the risk level of an individual, even if there is no direct hit found in the first automated assessment. The model would perform a risk assessment of ETIAS applications by creating and flagging criteria which could appear ‘risky’. This could be based on the individual’s data, or based on criteria relevant to an ongoing scenario, or developed from reviewing profiles which would return a hit in ETIAS. It is this type of analysis that will be introduced in both the ETIAS and VIS under current legislation.
An AI triaging system could rapidly separate standard applications from more complex ones in need of human review by an appropriate case worker.
An AI chatbot can support individuals requesting ETIAS permits during the process of completing the online form.
An AI chatbot/virtual assistant could answer any questions the applicant has after submitting their documents. It can inform the applicant what action is required to continue to the next stage of the application process.
An AI chatbot/virtual assistant could answer frequently asked questions from an applicant after an ETIAS decision has been made. The chatbot could support the applicant in finding the correct answer or direct them to the appropriate person/entity to speak to. It could also automatically schedule follow-up appointments.
An AI model could determine whether an individual should undergo a travel authorisation or full visa procedure, independently of their nationality. This would shift current procedures from being based on nationality, to individual factors. This would involve the analysis of different risk factors, previous travel history, and so on. It could present recommendations for human review, flagging any significant information.
An AI chatbot/virtual assistant can support individuals requesting a long-term residence permit.
This approach would be similar to VISA-1 and ETIAS-4.
Initially this case was intended to cover only the pre-application and pre-submission phase of an application for long term stay or residence. This might be expanded with support during and after the application (e.g. for renewal).
An AI chatbot/virtual assistant could answer any questions the applicant has after submission of their documents. It informs the applicant when they have taken action to continue to the next stage of the application process.
An AI triaging system could be used to automatically and quickly classify standard applications from more complex ones in need of human review. Classifying means grouping similar applications (e.g. from a certain country or reason for travel) for review by an appropriate expert. By grouping the applicants, the responsible officers would save time sorting the basic applications, allowing more time to be spent on those classified as complex. Technical approaches would be similar to VISA-3.
A chatbot could help at the permit renewal stage. It could find previously submitted documents and help with any questions about what form of renewal to apply for, and so on.
To answer any questions the applicant has after the decision has been made on an application for a residence permit. It can support the applicant in finding the right answer or direct them to the correct person/entity to speak to.
This use case would see facial recognition deployed to determine if two people are related “based on facial characteristics.” This would be because “DNA testing is not always a viable option.” There is no evidence that facial recognition can be used to determine a familial connection between individuals.
AI could be used to assess “success in integration.” It could also analyse the drivers of success in individual cases. Data would be gathered through the use of an AI chatbot in contact with the immigrant, alongside data received from external sources (e.g. tax statements).
A chatbot/virtual assistant could streamline interactions for individuals who have already received permits for long term stay in an EU member state, who wish to move to another member state.
A clustering algorithm could be used to group asylum candidates based on the similarity of their profiles and expected risk level. Risk level could be generated from data including applications, documents and interviews, in the form of both structured fields and/or less structured text/speech.
The aim of the tool would be to enable faster and more informed human decision-making by presenting similar cases or flagging notable pieces of information (e.g. outliers, similarities with other cases). Alternatively, it could be used in a post-decision context to assess historic consistency and quality of decision making.
Applying AI to scan through national legislation to identify what is needed for a compliant asylum procedure. The results would be suitable for individuals looking to quickly gain an understanding of the procedure. This may be useful for training.
Alternatively, legislation could be cross-referenced to identify any differences in process in different Member States, for instance. This could help monitor the operation of the Common European Asylum System (CEAS).
AI could perform real-time analysis of an applicant’s facial movements, spoken language and body language. It could be used to detect “abnormal” patterns (e.g. signs of distress) which can better inform decision-making by a human social worker/expert (e.g. is the applicant should be granted special procedural guarantees).
Techniques would aim to notice and assess the emotional cues displayed by both what the applicant says/does and the way that they do it, either in terms of modelling apparent emotion types or by detecting fluctuating or unusual behaviour. This raises substantial ethical and legal questions. It should be noted that the infamous iBorderCtrl project had similar aims, for the purpose of border checks.
This would use an AI model to assess whether a person is a minor (either as a binary classification or by attempting to predict age). The AI model would likely require an image of a face, but could also include other physical factors in its assessment. The model could then be used in combination with existing techniques to enhance human judgement.
The AI could provide additional value by using other outputs extracted from machine learning models. This could include confidence scores and insights into which input data influenced the model’s output (e.g. which regions of a facial image). There are many ethical questions surrounding collecting samples of facial images from minors, and also regarding the reliability of the technology.
An AI chatbot for the asylum registration process. This would include:
The chatbot would support the applicant by prompting for inputs (e.g. follow-up document requests), and by presenting a convenient way of dealing with frequently asked questions. The chatbot could also be deployed in a training setting - creating ‘virtual interviews’ to familiarise junior case workers with potential scenarios.
A post-entry chatbot/virtual assistant to aid refugee “integration.” It could respond to questions from recognised refugees. It could also provide suggestions to aid integration, such as local language classes and other events. It can also provide a means of monitoring other aspects of integration success.
An AI model could predict the risk of an applicant absconding/leaving during the asylum procedure. It could take into account variables such as country of origin, previous application history, age, travel patterns, and so on.
AI can be used to place individuals in a certain region of a country where they are more likely to find work and integrate smoothly. AI could match an applicant’s skills to the region’s labour market. It could also factor in existing settlement of others from the same country of origin and total flow levels of asylum seekers to predict integration success.
AI could assign individuals seeking asylum to detention centres, “optimising the likelihood of positive integration with other individuals.” It could also consider capacity and cost constraints across the network of detention centres.
AI would consider variables including:
It could suggest assignment to a particular centre based on this data, or streamline the presentation of this data for a human case worker for manual recommendation. It could also contain aspects of ASYLUM-11 (a similar analytics engine performing geographical assignment, e.g. to a particular member state) and ASYLUM-6 (a chatbot).
AI could provide credibility assessment tests and enhance risk assessment for deportations to origin country. The specific assessment of individual risk must remain a human activity, but AI can enhance human decision-making. It could provide extracts/summaries of relevant information. Such a system could also be used to generate questions for asylum interviews.
Using AI in collaboration with Internet of Things (IoT) technology (e.g. sensors, GPS signals, cameras) to monitor asylum seekers. This is presented as a potential alternative to detention. The UK government has used this form of technology. Its own research found it made no difference to the likelihood of people absconding.
The following potential uses of AI were included in a Deloitte report for the European Commission, based on surveys of and interviews with EU and member state officials.[4]
Frequently asked questions chatbot for border officials, helping with interaction and data extraction from databases.
An AI model to flag risk indicators during border check interview, based on data including the individual’s origin country, demographics, reason for travel, and so on.
AI model to triage border crossers into categories for a second line of action (e.g. interview/interrogation), most likely using historic trends based on the individual’s profile. It could also incorporate external factors such as seasonality and macro situation (overall economic conditions within a country/region).
Specifically, the system would analyse the entering travellers and divide them into a group that can proceed without passing a second border check, and a group that should go through the second check.
This would apply predictive analytics to the total migration flows at both land and air borders in order to improve staffing planning. By collecting enough data on travel patterns and expected migration flows, it would be possible to reassign border guards to ensure that there are enough officers during migration peaks to deal with larger number of travellers or vice versa.
AI could run analytics on border guard behaviour/decisions to understand trends, biases and potential inconsistencies. It could cross reference this with rejection/investigation rates to understand trends.
Facial and fingerprint recognition could be used to verify the identity of travellers more seamlessly, enhancing the current passport check. It should be noted that this is foreseen with the introduction of the Entry/Exit System.
Use of machine learning to rotate fingerprint images into the correct orientation for further use.
Through the use of cameras at border crossing points an AI system could apply computer vision to detect Schengen Information System (SIS) alerts, such as identifying a target person or car number plate. The system would capture the image of the border crossers and send a notification to a border guard if there was a match with an alert stored in the SIS. The border guard could then validate/verify the obtained match and perform the activity requested in the alert. In late 2019 the transport agency of the Australian state of New South Wales announced it had installed computer vision technology into roadside cameras to spot offenders, similar to what is proposed here.
An AI chatbot could support SIRENE officers to fill in forms correctly and accurately. For instance, it could help officers to choose the correct form and suggest/provide inputs for the fields in the form. SIRENE stands for supplementary information request at the national entries. SIRENE offices and officers are responsible for exchanging information based on SIS alerts, for example between member states’ border or police agencies.
AI could automatically create reports (natural language generation). The report would automatically create a summary of key indicators/trends found in the data, personalised to the user (police officer, border guard, etc.). For example, this could be used to create reports on alerted individuals, to whom officers should pay special attention. This can be tailored to the officers working in a certain geographical region or job type. Another example would be reports sent to government bodies to make them aware, e.g. increased numbers of Syrian refugees at the Italian border.
AI could be used to enable search and exploration of the SIS database. The AI could be tailored to the user and their ways of searching - through a chatbot interface for instance. It could work with a semantic layer to facilitate searching via natural language queries.
A SIRENE form can be completed using the information from the original alert alongside a report from an officer highlighting the action taken. An AI system could be used to automatically recognise key information from the alert, and to collect and structure information from an officer’s report.
Specifically, using natural language processing, the system would work as a virtual assistant to gather the necessary information from the officer and automatically complete the SIRENE form. This information could then be sent to the other member state.
The following potential uses of AI were included in a Deloitte report for the European Commission, based on surveys of and interviews with EU and member state officials.[5]
AI could collect and analyse sensor data to make energy usage more efficient within infrastructure operations or the data centre. For instance, it could optimise cooling mechanisms based on expected load of systems.
AI can make eu-LISA systems operations and throughput more efficient by applying intelligent routing for load balancing.
AI could apply big data analytics to infrastructure performance metrics. This could automate the process of fault identification and recovery. In particular, the AI would focus on prioritising identified faults, based on expected complexity and impact.
An AI model could predict trends for the efficient provision of IT resources, such as network and storage.
The objective of this case is to reduce the burden on service desks within eu-LISA. Specifically, the goal is to improve efficiency for both internal staff (e.g. eu-LISA help desk) and external stakeholders (e.g. Member States). By improving waiting times and resolution time, ultimately ‘customer experience’ will also improve.
This case aims to improve the accuracy of biometric matching, specifically for facial images. This would ensure that officers do not lose time investigating false positive matches and that citizens are not unnecessarily stopped for an invalid reason.
State of the art AI methods would be used, refined specifically for eu-LISA. It could possibly include dataset generation to improve models in the absence of real training data.
This case intends to speed up the learning process of various stakeholders by answering general questions related to newly-developed Core Business Systems at eu-LISA. For example, carriers will be required to use ETIAS to check if a traveller has a valid travel document. Similarly, border guards must be familiar with the ETIAS and EES when performing checks on TCNs.
As ETIAS will be a new system for these users, it will require learning from their side. Thus, this case intends to target such situations to improve their familiarity with the system, while providing any clarification needed. This case would require a chatbot created specifically for each of the systems and tailored to the user.
The following potential uses of AI were included in a Deloitte report for the European Commission, based on surveys of and interviews with EU and member state officials.[6]
AI can observe the links between regulations, both direct (referred) or indirect (similar regulations).
An AI model can evaluate the extent to which member state legislation is compliant with EU legislation. The model could also potentially provide a downstream recommendation on what infringement procedures could be initiated against non-compliant member states.
AI can search for similar regulations based on substance and format. This would speed up development in terms of both knowledge and process (filling in templates). This would ensure consistency and appropriate dependencies between regulations.
Identify gaps in regulations to focus on areas where regulation can help to better protect citizens and travellers.
AI can assess if terminology is consistent and highlight cases where wording could be improved. Also, a check could be performed to see if a regulation covers all the possible cases.
Analysis of social media, newsfeeds, publications, legislative text by natural language processing and other big data analytics to identify trends.
AI could assist with presenting and communicating new policies to civilians and business through visualisation techniques that use dimensionality reduction (and clustering). This can help communicate complex data/insights.
During discussions between stakeholders (agencies, Member States, citizens, business etc.) a common denominator can be sought to select the policy option that addresses the majority of the demands. AI could extract all requirements from different stakeholders and look for an option which fulfils as many of them as possible.
An AI model could use demographic and political data to assess if citizens or politicians will be in favour of or oppose a new policy regulation.
The following potential uses of AI were included in a Deloitte report for the European Commission, based on surveys of and interviews with EU and member state officials.[7]
AI could provide translation to/from multiple languages of documents provided by the applicant and which need to be understood by an applicant/case worker/other concerned party during assessment of the case.
AI can use optical character recognition (OCR) for converting handwritten forms into digital data that can be used with other computer systems. OCR would convert handwritten text into machine readable text data.
AI can use optical character recognition (OCR) for converting physical forms into digital data that can be used with other computer systems. OCR would analyse the physical forms to create a digital structure containing the interpreted data (text, numbers etc.). The AI could also link into the appropriate downstream systems to further streamline operations.
For instance, instead of a case worker typing in travel details from a physical passport, the OCR system would automatically detect, extract and structure the data, and then store it in an appropriate location.
An AI model could verify if an application is completed correctly before it is submitted. It could engage with the applicant to make them aware of any possible missing/inaccurate data.
AI could detect forged supporting documents such as birth certificates or bank statements. The model would identify signs of fraud:
The model would flag potential cases for human review by highlighting areas of the document requiring further investigation.
An AI case-based reasoning engine could analyse active cases by retrieving similar historic examples. These can then be presented to a caseworker for review when making decisions about a pending application. Similarity could be based on the applicant profile, “macro context,” or more custom indicators which might be developed internally. The tool could include a dynamic self-learning module. It could adapt the model using new knowledge received in the feedback cycle, and take it into account when processing new applications.
By flagging inconsistencies, caseworkers can be made aware of personal biases or identify general biases and receive appropriate feedback.
AI can monitor other analytics and AI systems to uncover undesirable trends, such as a biased decision making. This could either analyse post-hoc results to check for imbalances between various groups of data or individuals. It could also analyse systems when applying explainability/interpretability techniques to machine learning models, to understand what variables are considered significant in decision making.
A chatbot can inform applicants on their rights and the possibilities of appealing a denied application. This is with the aim of ensuring a fair and effective process.
An AI model could monitor human bias when assessing an application by performing a post-hoc quality/fairness assessment. It would search for correlations between applications and outcomes.
By using computer vision, an AI system could detect the use of forged travel documents. It would analyse an image of the provided documents and assess if the physical characteristics of the document match an original one. It assesses if the information provided in the documents is accurate, and the person providing the documents corresponds to the person in the document, rather than a lookalike. The technical approach would be similar to CROSS-6.
AI can provide real-time translations of a discussion between an officer and an applicant during an interview.
AI can scan and monitor different systems to assess if the conditions in which a permit was granted to a third-country national (TCN) still apply. The system would use the data from those systems to assess the likelihood of an applicant not complying with the terms set when issuing the permit.
For example, a TCN might receive a residence permit because of marriage to an EU resident. However, the couple could separate soon after the permit is issued. In this case, the conditions for initially providing the permit no longer apply. The system would try to assess whether conditions for the permit are still valid by analysing various sources of data (e.g. address or tax information) and provide insight on the possibility of fraud.
Another example is to monitor if a TCN is complying with the restrictions of the issued work permit, such as number of days worked. This could be checked by analysing tax statements.
AI could make detention centre allocation more efficient. It would predict when the individual might be deported, which would come from a model fed with the person’s information and risk assessment data. More data can be processed by prioritising individuals who are likely to be deported quickly. In particular, this case is highlighted as being appropriate for irregular migrants/Dublin cases/deportation.
Use of a chatbot could support travellers/citizens with generic questions regarding the rights and obligations within the Schengen area. For instance, information about the right to work for a limited time period and obligation to leave the country after 90 days within the Schengen area.
[1] ‘Cross-border digital criminal justice’, 2020, https://op.europa.eu/en/publication-detail/-/publication/e38795b5-f633-11ea-991b-01aa75ed71a1/language-en
[2] eu-LISA and Eurojust, ‘Artificial intelligence supporting cross-border cooperation in criminal justice’, June 2022, https://www.eurojust.europa.eu/sites/default/files/assets/artificial-intelligence-cross-border-cooperation-criminal-justice-report.pdf
[3] ‘Opportunities and challenges for the use of artificial intelligence in border control, migration and security’, 2020, https://op.europa.eu/en/publication-detail/-/publication/c8823cd1-a152-11ea-9d2d-01aa75ed71a1/language-en
[4] ‘Opportunities and challenges for the use of artificial intelligence in border control, migration and security’, 2020, https://op.europa.eu/en/publication-detail/-/publication/c8823cd1-a152-11ea-9d2d-01aa75ed71a1/language-en
[5] ‘Opportunities and challenges for the use of artificial intelligence in border control, migration and security’, 2020, https://op.europa.eu/en/publication-detail/-/publication/c8823cd1-a152-11ea-9d2d-01aa75ed71a1/language-en
[6] ‘Opportunities and challenges for the use of artificial intelligence in border control, migration and security’, 2020, https://op.europa.eu/en/publication-detail/-/publication/c8823cd1-a152-11ea-9d2d-01aa75ed71a1/language-en
[7] ‘Opportunities and challenges for the use of artificial intelligence in border control, migration and security’, 2020, https://op.europa.eu/en/publication-detail/-/publication/c8823cd1-a152-11ea-9d2d-01aa75ed71a1/language-en
Spotted an error? If you've spotted a problem with this page, just click once to let us know.