Governments and corporations are increasingly using artificial intelligence to monitor migrants and to scrutinise the work of civil society groups.
This is happening at borders, in protests, on social media, and in the news that people see every day.
The application of AI can make discrimination look like a neutral, technical process. When a machine makes the decision, it is harder to challenge and easier to hide. But the machine is not neutral and the people that are least able to fight back against its mistakes are consistently the ones it hurts the most.
There are some main problems connected with AI: the first one is in border control; AI is being used to filter and exclude migrants in ways that reproduce racial bias.
The second one is during protests and assemblies, AI is used for surveillance, and tools like facial recognition are being used to track and discourage people from demonstrating.
The third one is about online speech, with automated content moderation overly silencing civil society organisations.
These aren’t separate problems. Together, they form a pattern: the space for ordinary people to engage in public life is quietly shrinking, and AI is the mechanism that is also being used to reduce it.
The Border as testing ground
Artificial intelligence is currently used in border control to identify and catch individuals considered “the Other” in regions lacking physical barriers.
Algorithmic systems at borders do not function as neutral tools of security management. They construct and enforce categories of belonging and exclusion. The EU border regime has become one of the most intensive testing grounds for AI experimentation in the world. Systems assess the credibility of asylum claims, screen travellers for risk, manage flows at hotspots, and identify individuals in movement. These systems have assumptions about what a legitimate refugee looks, sounds, and behaves like, assumptions that are deeply racialised, gendered, and shaped by histories of colonial power.
When an AI assesses the “risk” of a traveller, it is not performing a neutral control. It is reproducing the discriminatory judgments that have always characterised border enforcement, while protecting those judgments from accountability. No citizen input. No timely appeal. No transparency about why the algorithm decided what it decided.
In the area of migration, the AI Act fails to prevent harm. Law enforcement, migration control, and national security authorities are exempted from the most important transparency, oversight requirements. AI systems used in EU migration databases including Eurodac and the Schengen Information System will not have to comply until 2030. The people most directly affected have no meaningful voice in how these systems are designed, and have no effective remedy when they are harmed.
AI also transforms the violence of border enforcement into the neutral language of risk management and efficiency. People do not arrive at borders as “illegal migrants”; they are made illegal through systems that sort people into categories of deserving and undeserving. AI does not simply manage this process; it accelerates and deepens it, making the construction of illegality more automated and harder to contest.
Facial recognition on protesters
Facial recognition used against protesters does not just identify individuals who may have committed offences. It identifies protesters as a class, creating infrastructure for retroactive surveillance and preventing participation before it takes place. This is suppression by anticipation. It requires no arrest warrant and no criminal charge. And it works most effectively against the communities AI systems are already most likely to misidentify.
On 18 March 2025, the Hungarian Parliament passed legislation allowing police to ban LGBTQ+ public events, criminalising participation in prohibited assemblies, and crucially amending the law on facial image analysis to allow facial recognition to identify, fine, and sanction participants in prohibited events. The bill was introduced on 17 March and entered into force on 19 March at a speed that eliminated any possibility of meaningful debate, consultation, or impact assessment. This is democratic exploitation in practice: formal legal procedures used so rapidly that democracy’s own safeguards cannot activate in time.
Budapest Pride was banned for 2025. When the Mayor of Budapest defied the ban and allowed the parade to proceed on 28 June.
The facial recognition provision creates infrastructure for retroactive identification of everyone who attended. A database of faces captured at a Pride parade linked to the legal category of participation in a prohibited assembly becomes an instrument of political persecution deployable long after the event. Bystanders, journalists, supporters, family members could all face legal consequences on the basis of an AI system’s identification, with all the attendant risks of misidentification on future participation.
The rules of assembly are not abolished; they are weaponised. Facial recognition is not deployed as open repression; it is deployed as “enforcement.”
The Silencing of Civil Society Online
Digital platforms have become primary arenas for organising and advocacy, and AI-driven content moderation has become a path of quieter suppression. When civil society organisations cannot reach their communities online, citizens lose access to information, mobilisation, and the networks that make democratic participation possible.
The reproductive rights campaign My Voice My Choice had its Instagram presence more times restricted and disrupted. It suspects coordinated mass reporting by opponents of abortion rights, amplified by algorithms that react to certain words or images without the context needed to distinguish advocacy from prohibited content.
The NGO received more than 130 reports of account restrictions in April 2025 alone. By later in 2025, around 50 reproductive rights organisations, queer groups, and sexual health providers had alleged that Meta had restricted, shadow-banned, or removed their accounts including the banning of abortion helplines in countries where abortion is legal, the removal of queer content, and the restriction of non-explicit sexual health information.
Meta maintains its policies apply equally to all. But the experience of dozens of organisations suggests otherwise: content moderation AI, trained on data reflecting existing social hierarchies, reliably over-moderates speech by and about marginalised communities. A major platform can absorb the reputational damage of wrongly removing a reproductive rights helpline. The organisation whose work has been deleted and the citizens who needed to access it cannot. This is not a coincidence. It is a structural feature of systems built without meaningful participation from the communities they affect.
Conclusion
The picture that emerges is one of convergent crises. Biased AI systems are used in border control, systematically harming the people that are already most harmed before implementation of AI. Facial recognition suppresses protest and enables retroactive identification of protestor. Automated content moderation silences civil society advocating for reproductive rights, sexual health, and LGBTQ+ dignity. These are not separate problems with separate solutions. They are manifestations of a common dynamic: AI used to extend and present existing power structures, separated from accountability by algorithmic opacity and the fiction of technological neutrality. Their common consequence is the narrowing of the space in which citizens can participate in the democracy that is supposed to belong to them.
The EU AI Act is an important step but its effectiveness depends on its implementation and enforcement. Recent amendments under the Digital Omnibus have delayed the application of key obligations for high-risk AI systems, including systems used in areas such as migration, border control and law enforcement, in the name of regulatory simplification and competitiveness. Rather than weakening safeguards the EU should strengthen them. Citizens should have a right to know when they are interacting with AI, how decisions made by AI are affecting them, what data is being collected and where that data is processed or shared. Transparency should not depend on market choice but be a legal requirement for AI systems used in Europe.
The AI Act’s exemptions for law enforcement and migration authorities should also be revisited, with civil society and affected communities that are involved in that process.
As CTOE, we are worried by these wrongful uses of AI, and we insist that the values we hold must be reflected in the systems that govern our lives.
The algorithm does not have values. But we do. And it is time to insist that they count.
Sources:
Bouvier, C. (2024, April 4). A dangerous precedent: how the EU AI Act fails migrants and people on the move – PICUM. PICUM. https://picum.org/blog/a-dangerous-precedent-how-the-eu-ai-act-fails-migrants- nd-people-on-the-move/
Union, B. (2025, October 22). Largest study of its kind shows AI assistants misrepresent news content 45% of the time – regardless of language or territory. Www.ebu.ch. https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste
Haeck, P., & Csongor Körömi. (2025, April 25). Hungary on EU watchlist over surveillance at Pride. POLITICO. https://www.politico.eu/article/hungary-eu-watchlist-facial-recognition-surveillance-lgbtq-pride/
Kreps, S., & Kriner, D. (2023, October). How AI Threatens Democracy. Journal of Democracy. https://www.journalofdemocracy.org/articles/how-ai-threatens-democracy/
Hall, L. (2021). Programming the machine: gender, race, sexuality, AI, and the construction of credibility and deceit at the border. Internet Policy Review, 10(4). https://doi.org/10.14763/2021.4.1601
Milivojevic, S. (2022). Artificial intelligence, illegalised mobility and lucrative alchemy of border utopia. Criminology & Criminal Justice, 174889582211238. https://doi.org/10.1177/17488958221123855