The main idea behind ITFLOWS is the same as other EU-funded projects’, such as FUME or HumMingBird: to use Machine Learning and sentiment analysis techniques to provide evidence-based predictions of future migration flows so that potential issues and risks of social tensions can be anticipated and optimally tackled.
The project’s Cordis page states that “the EU-funded ITFLOWS project aims to predict and manage migration flows via the creation of an evidence-based information and communication technology-enabled solution, the so-called EUMigraTool (EMT, ed.). The tool will ease the reception, relocation, settlement and integration of migrants. Intended to be used mainly by first-line practitioners, second-level reception organisations and municipalities, it will also help to predict migration flows and identify the potential risks of tensions between migrants and EU citizens. Ultimately, the project will formulate recommendations and good practices for the attention of policy makers, governments and EU institutions.”
From this solutionist perspective, the technology would be necessary “in order to deal with migration flows in Europe in an optimal way.”
Nonetheless, the project homepage says: “Our goal is to provide predictions of migration flows to enhance humanitarian support.”
The “evidence-based ICT enabled solution” and “prediction tool” EUMigraTool (EMT) is ITFLOWS’ main technological output.
A description is provided on the project’s Cordis page: “The backend of the EMT comprises 2 complementary approaches to simulation and prediction. The first approach is based on agent-based modelling, while the second approach relies on artificial intelligence. The EMT includes analysis of media content from TV-news (video content), web-news and social media (text content) using deep learning and proposing novel deep architectures in generative modelling and forecasting using sequential data. Predictions incorporate algorithms that consider the two key challenges associated with prediction of migration: (a) Adequate selection of relevant data sources, and (b) correct selection of the potential drivers to be monitored and the warning thresholds to be set.”
The tool, which according to its developers will only be used for “humanitarian purposes and will never be made available to governments, to manage migration flows or deliver data to immigration control agencies,” has two complementary functions:
1) SIMULATION: “The Small-Scale Model (SSM) aims to predict the distribution of incoming asylum-seekers/unrecognised refugees arriving to neighbouring countries of conflict origins. It uses a generalised and automated simulation development approach and the Flee agent-based simulation code, which is optimised for simplicity and flexibility. The SSM synthesises data from the United Nations High Commissioner for Refugees (UNHCR), the Armed Conflict Location and Event Data Project (ACLED), OpenStreetMap and population data using the City Population database or other population sources. The conflict model is constructed, run and validated by comparing the simulation results to the existing camp registrations obtained from UNHCR;”
2) FORECASTING: “The Large-Scale Model (LSM) produces monthly predictions of asylum applications in the EU for a variety of bilateral (i.e., from country of origin to the EU Member State) cases. It uses state of the art machine learning approaches, including neural network architectures and time series analysis. Its techniques allow for correlation analysis between raw data sources and simulation. Furthermore, the LSM provides intuitions on attitudes towards migration (“intuitions”, ed.) among populations in all European destination countries, using the Twitter Sentiment Analysis model data as input, and the most influential or relevant determinants of attitudes towards migration. The LSM combines a set of different inputs and methods from Topic Modeling by monitoring national press and asylum seeker data from Eurostat (the official EU statistics office).”
Two reports describe the EMT in detail. The first was published in October 2022, the second in February 2023. A list of data sources is also available.
At an Annual OECD Conference in Paris, Konstantin Boss (UAB) “explained the use of Google Trends data to predict bilateral refugee flows and how the UAB team implemented that within the ITFLOWS Project.” Deliverable 3.3 illustrates “details of the big data used in predicting country instability and model description,” while D6.1 details the specifications and architecture of the EMT.
Deliverable D1.3 illustrates “synergies with other projects,” such as HiDALGO, which “makes use of real world data from UNHCR, ACLED, and Bing Maps” for an “agent-based” “simulation framework” to “predict possible destinations of refugees coming from conflict regions.” According to D1.3, WP6 will integrate existing solutions — such as “Flee agent-based modelling code and FabSim3 automation toolkit” used in HiDALGO, or solutions from easyRIGHTS, METICOS, and TRAFIG to “investigate long-lasting displacement situations at multiple sites in Asia, Africa and Europe and analyse options to improve displaced people’s lives,” — and “develop solutions for protracted displacement situations (PDS) that are better tailored to the needs and capacities of persons affected by displacement.”
Non-EU funded projects involving the prediction of migration flows are also listed, including GDELT (privately funded), JETSON (UNHCR), and Foresight (Danish Refugee Council).
D7.5 states: “One potential exploitation of the tool would be to integrate it into an existing forecasting or predictive platform or project that is oriented towards humanitarian purposes. The tool could thus complement a pre-existing system and be operated by, for example, a larger NGO that has the capacity to maximise the tool’s capabilities and employ it in their operations that seek to assist migrants and asylum-seekers arriving to Europe.”
Other projects/tools that have been considered for potential integration include the Internal Displacement Event Tagging and Clustering Tool (IDETECT) by the Internal Displacement Monitoring Center, IOM’s Displacement Tracking Matrix System (DTM), Forecasting Reports by Frontex, and the Heightened Risk Identification Tool by UNHCR — plus forecasting statistics adopted in Belgium, Germany, Norway, Ireland, the Netherlands, Sweden, and Switzerland.
“After an exhaustive analysis of these existing predictive tools in the field of migration, the ITFLOWS Consortium attempted to merge with one of them: Foresight, currently funded and operated by the Danish Refugee Council. We approach them with this proposal in May 2022 and, although initially the institution showed interest in such opportunity and they were discussing this possibility internally in the department management team (DMT), they finally declined the collaboration in December 2022. (…) Therefore, for now we are not planning to merge with any other current predictive tool in the field of migration.”
ITFLOWS is also part of the Border External Security (BES) Cluster of EU-funded projects, lead by METICOS.
According to deliverable D2.3, “the EUMigraTool will be tested by a variety of potential end-users from civil society organisations and municipalities. In particular, pilot cases will be validated by experts in human rights and civil society organisations with expert knowledge on migrants, asylum seekers and refugees.”
D7.1 adds: “The tool is to be validated in, at least, three specific EU Member States: Italy, Spain and Greece.” D7.5 confirms: “The validation has been mainly carried out by members of the UB in three EU countries, i.e. Spain (September 2022), Italy (February 2023) and Greece (June 2023). These countries have been chosen according to the number of arrivals and asylum seekers in the EU, and considering the nationality of the practitioner’s part of the ITFLOWS consortium.”
The EMT was used to analyze “the Nigeria conflicts.” The Third report on the EuMigraTool specifies: “asylum seekers from Nigeria to Germany” and “attitudes towards migration in Greece” through sentiment analysis of Twitter messages.
Deliverable D1.3 notes that “the EMT has reached Technology Readiness Level (TRL) 6 and is at a point that can be used outside of the testing environment.”
A Disclose investigation based on access to information requests says that ITFLOWS “is due to enter service in August 2023” (the research project ended on 31August 2023, ed.) — and this “despite repeated warnings that its predictive capabilities could end up being misused to control and restrict the rights of refugees on European soil.” According to Disclose, “charities Red Cross and Oxfam are helping to supply important information for the software used by ITFLOWS. This data comes directly from interviews carried out in migrant camps with Nigerian, Malian, Eritrean and Sudanese refugees. This information could, for example, be about the ethnic origins, sexual orientation or religion of those interviewed.”
Two internal consortium reports obtained by Disclose reached “alarming conclusions”: “The first document (225 pages) reveals that the “ITFLOWS consortium is fully aware of the risks and their potential impacts in terms of jeopardising human rights that both empirical migration research activities and technological developments foreseen in the Project may pose.” The information provided by the algorithm “may pose several risks if misused for stigmatising, discriminating, harassing, or intimidating individuals, especially those that are in vulnerable situations such as migrants, refugees and asylum seekers.”
Five months later the ethical board delivered a second report: “Member States may use the data provided to create ghettos of migrants,“ “discrimination on grounds of sexuality, race, religion, disability, age,“ “the risk that migrants and asylum seekers may be identified and sanctioned for irregularities“. The report also warns: “The risks of reinforcing fear and arguments against migration, or the increasing hate speech in areas where the inhabitants are informed that the inflows will move.” The Disclose investigation contextualizes: “These documents were draft versions, and not final European Commission approved versions. They were not and are not to be public as they were only draft versions, Itflows reacted in a letter sent to Disclose a few days after the publication of our article.”
A member of the ITFLOWS ethical board, Alexandra Xanthaki, argued during a 2021 symposium: “We spent six months working day and night to create a report about the human rights framework (…). And now it seems to me that what the tech members are saying is: we’re not taking it into account. So what’s the point in having it in the project?”
The main concern seems that rather than “enhancing humanitarian support,” the EMT will be weaponized against migrants. Contacted by Disclose, Alexander Kjærum, a senior analyst at the Danish Refugee Council (DRC), who also sits on the users board at ITFLOWS, said: “Here’s a big risk that information gets into the hands of states or governments that will not use it to enhance support and protection for these vulnerable groups, but will use it to throw up more barbed wire” (here is the full response by ITFLOWS).
Deliverable D7.5, the “Final EMT Analysis Report”, discusses “the possibility of expanding the tool within the EU” in Section 3, “Societal Impact Analysis”: “The EMT is expected to be used within the European Union.” It “also examines possibilities of expansion in countries outside the EU.”
As for the validation, the deliverable says that “simulations” managed to “accurately predict more than 75% of the asylum-seekers / unrecognised refugee movements for Mali, Nigeria, Syria, and Venezuela. However, the Ukraine conflict has a higher average relative difference in comparison to other conflict simulations.” The pilots’ countries of origin include Syria, Nigeria, Mali, Venezuela, Iraq, and Morocco.
As objectives, the project homepage states “policy recommendations for receiving migration flows in the EU in accordance with human rights.” How problematic this is showed the Disclose investigation and the project’s Ethical Board.
A second objective is “to identify sentiment that could lead to potential risks of tension between people residing in the EU (EU citizens, their non-EU family members, migrants including refugees) to protect vulnerable people.” The question is whether sentiment analysis is up to the task, and on what scientific grounds “risks of tension” were identified.
The project’s Ethical Board consists of project collaborators and three independent “renowned professors,” a “Gender Committee” (again with researchers from consortium partners and independent experts) that worked on measures to effectively protect from discrimination, as per D7.1, and a “Policy Working Group.” A “data protection advisor” is also included in the team.
But can AI correctly predict migration flows at all? And actually “mitigate risks of tensions between migrants and EU citizens”, as claimed in a project video? Is this proven or assumed?
Data sources are specified for the overall project, and its website also “provides relevant, recent statistics from the Eurostat database extracted from ITFLOWS Deliverable 4.3.”
ITFLOWS organized a conference, “Migration Prediction, Policy and Human Rights” with “the aim to discuss
1) the challenges posed by the misuse of migration flow prediction technology for humanitarian purposes within increasingly politically polarised contexts in the EU,
2) how bias, inaccuracy, and the misunderstanding of the context of use can negatively impact the human rights protection of migrants and refugees, and
3) the human rights policy challenges when using migration prediction technologies for humanitarian purposes”. Minutes are unfortunately not available.
Deliverable D8.1 contains “good practices and asylum policy commitments according to human rights,” and Deliverable D2.3 a “Report on Human Rights, Ethical, Societal and Data protection risks assessments,” which includes an “ethical and societal impact assessment” and a “Data Protection Impact Assessment.”
A project paper highlights the human rights risks of the EU’s interoperability framework for border management.
Another project paper discusses “generalised push-back practices in Europe.” It is unclear what would prevent tools such as the EMT from being exploited for preventive (and illegal) pushbacks (as the #ProtectNotSurveil coalition convincingly argues).
D7.1 explicitly states: “The UB (Users Board, ed.) was also asked whether in their opinion there were ways in which the EMT could be misused, and they all assented. The misuse could entail closing of borders, instigating violence, and misuse for political purposes to gain support and consensus for an anti-migration policy. The main recommendation was to ensure that the tool was shared only with relevant stakeholders.”
The authors of a paper on forecasting migration tools (“The Role of Emerging Predictive IT Tools in Effective Migration Governance”) claim to have “identified 18 tools and projects incorporating AI, which were relevant for EU migration governance purposes” between 2010 and 2020.
Three in particular are detailed in the analysis:
“(1) the Jetson tool, funded and operated by UN High Commissioner for Refugees;
(2) the Early Warning and Preparedness System tool (hereafter EPS‐Forecasting), funded and operated by the European Asylum Support Office;
(3) Foresight, currently funded and operated by the Danish Refugee Council (DRC). It was initially funded by the Danish Ministry of Foreign Affairs, with the model and user interface developed in collaboration with IBM.”
They all use Machine Learning.
Cristina Blasi Casagran and her colleagues also claim that “an ideal model could include behavioural and sentiment analysis collected online,” such as posts, likes, and interactions of migrants on social media.
Sentiment analysis is included in the EMT, as per D5.3, which “describes the architecture and main functionalities of a sentiment analysis prototype for the ITFLOWS EUMigraTool (EMT) policy-making and sentiment monitoring module.”
D7.5, the Final EMT Analysis Report, notes that the tool, born out of the need to predict and prevent social tensions, might actually end up exacerbating them:
“A possible problem that was raised was that a problematic dynamic within the integration process could be exacerbated. Specifically, the goal to eliminate or reduce to their minimum cultural and social differences between asylum seekers and locals with the aim of integrating the former within the host society. This could take the form of sometimes harsh impositions of sociocultural dynamics, accompanied by the need to reduce certain idiosyncratic expressions or behaviours by the migrant community, that can be sometimes perceived as obstacles for their ‘successful integration’.”
D7.5 also notes that “the EMT is not fully run by artificial intelligence. ITFLOWS researchers regularly analyse and synthesise the information processed and feeding the algorithms, ensuring that EMT-generated results are accurate and up to date. Also, human supervision from our ethical team regularly screens for any potential misuse of the outputs for security and migration control purposes; i.e., access to the prediction and forecasting datasets themselves is still restricted.”
The deliverable also claims that the EMT were not available for policy makers, and that access to the datasets and the prediction and forecasting datasets generated by the EMT followed a “restrictive approach” — “because of the risks that could take place if the access of such tool is potentially handed to other end user such as national governments, law enforcement agencies, intelligence agencies or even FRONTEX.”
