Demonstrators take part in a protest against the deployment of an AI-assisted police surveillance system in Berlin on July 17, 2026. Photo by Andres Trujilo/ JNA PRESS(Sipa via AP Images)

Violent Anti-Tech Extremism Has Roots in History

Illustration of a laptop displaying the OpenAI logo on its screen. A hand holds a magnifying glass over the screen with blue lightning bolts emanating from it, symbolizing scrutiny or investigation. The image uses a halftone comic style with blue and black accents.

ChatGPT, Roblox Face Tough EU Rules. Implementation Is the Hard Part

Paper collage. A photograph of the Atomium in Brussels, cut into diagonal strips and reassembled over a grid of numbers.

Seeing More — Seeing Less by Anna Riepe & FARI / 

Can the EU AI Act’s Transparency Rules Strengthen Democratic Resilience?

﻿Germany’s Federal Chancellery building on a sunny day

Germany’s Federal Chancellery building on July 19, 2025 (Photo by Michael Nguyen/NurPhoto) 

What Germany's AI Security Institute Could Learn From Other Countries

This image is a section of the Detroit Industry Murals by Diego Rivera, created between 1932 and 1933. Rivera, a renowned Mexican muralist, painted these panels inside the Detroit Institute of Arts. They are widely considered one of his greatest works and a landmark of 20th-century art.

The workers shown are laboring on automobile assembly lines, likely at the Ford Motor Company’s River Rouge Plant.

Rivera celebrated the dignity and effort of industrial labor, portraying the workers with strength and unity.

The style is Social Realism, characterized by its focus on working-class subjects and its detailed, lifelike forms.

Photograph of a detail of a 1933 Diego Rivera fresco at the Detroit Institute of Arts. Justin Hendrix/Tech Policy Press

A Labor Day Prescription for Preserving Fair Employment in the Digital Age

A large number of people take part in a gathering of boats and motorbikes in support of the people of Ceuta on 30 August 2026 in Ceuta (Spain). (Europa Press via AP)

A large number of people took part in a gathering of boats and motorbikes in support of the people of Ceuta on 30 August 2026 in Ceuta (Spain). (Europa Press via AP)

How Disinformation Fueled the Latest Ceuta Mass Crossing

Mechanical Turk is Closing. The Workers Who Built AI Are Still Here

Members of California's legal team leave court after a settlement was reached in a multi-state lawsuit accusing Meta of causing harm to children on social media outside the Ronald V. Dellums Federal Building and US Courthouse on Wednesday, Aug. 26, 2026, in Oakland, Calif. (AP Photo/Noah Berger)

August 2026 US Tech Policy Roundup

The Colorado State Capitol is framed by buildings along Sherman Street Friday, June 27, 2025, in Denver. (AP Photo/David Zalubowski)

What 12 State ‘Companion Bot’ Laws Demand of AI Providers

A tic-tac-toe board with human faces as digital blocks, symbolizing how AI works on pre-existing, biased online data for information processing and decision-making.

Amritha R Warrier & AI4Media / Better Images of AI / tic tac toe / CC-BY 4.0

Germany's Draft Law on AI in Migration Raises Rights and Bias Concerns

A collage of circuit boards interwoven with aerial photographs of water treatment facilities, agricultural lands, industrial sites and mining machinery.

Privacy by Architecture Makes AI Compliance Work

A yellow school bus is seen in Chicago, Illinois, United States, on July 28, 2026. (Photo by Marcin Golba/NurPhoto)

In the US, Caution Rules on AI Ahead of K-12 School Year

A young child holds up a smart phone in front of their face using both hands.

This Week's Meta Settlement Will Reverberate Across the World for Years

This image is a collage with a colourful Japanese vintage landscape showing a mountain, hills, flowers and other plants and a small stream. There are 3 large black data servers placed in the bottom half of the image, with a cloud of black smoke emitting from them, partly obscuring the scenery.

Deborah Lupton, "Servers in a Landscape" / 

'Technology Freedom Cooperative' To Build Infrastructure for Human Rights and Justice Groups

Meta CEO Mark Zuckerberg (C) ahead of testifying before the Senate Judiciary Committee hearing on protecting children from sexual exploitation online, in the Dirksen Senate Office Building in Washington, DC, USA, 31 January 2024. (Photo by Ritzau Scanpix/Sipa USA)(Sipa via AP Images)

Meta Owes the Rest of the World Its Teen Protections

A massive sculpture of Icarus lies in front of the Concordia Temple in the Valley of Temples in Agrigento, Sicily, on January 5, 2019. Shutterstock

A massive sculpture of Icarus lies in front of the Concordia Temple in the Valley of Temples in Agrigento, Sicily, on January 5, 2019. 

Make AI Companies Criminally Liable for Preventable Harm

The EU and US flags in front of a European Commission logo.

The EU and US Can’t Agree on Tech — or Trust

Nicklas Akers, California Senior Assistant Attorney General, left, and Deputy Attorney General Megan O'Neill exit court after tech giant Meta reached a settlement to end a landmark trial over teen social media addiction in Oakland, Calif., Wednesday, Aug. 26, 2026. (AP Photo/Noah Berger)

Assessing the Landmark Meta Settlement

Signs for company Cambridge Analytica in the lobby of the building in which they were based on March 21, 2018 in London, England. (Photo by Chris J Ratcliffe/Getty Images)

The $459 Million Cambridge Analytica Reckoning Buried in Meta's Teen Safety Settlement

The TikTok logo is displayed through a magnifying glass in this photo illustration in Ontario, Canada, on February 5, 2026. (Photo by Thomas Fuller/NurPhoto via AP)

Understanding Brazil’s $29.7 Million TikTok Fine Over Children’s Data

 OpenAI’s ChatGPT as a Very Large Online Search Engine (VLOSE) and named Roblox and Reddit Very Large Online Platforms (VLOPs) under the 

 (DSA). Four months from now, these services will have to comply with additional obligations, including that they must:

assess and mitigate the systemic risks their services generate

open their internal data to vetted researchers

give users at least one content recommendation system not based on profiling

publish a repository of the advertisements they run

release biannual public reports on their content moderation

These new requirements kick in anytime an online platform or search engine becomes “very large,” meaning that it reaches at least 45 million average monthly users in the EU. The implications are substantial, particularly for transparency. As very large online services, ChatGPT and Roblox now have to open their hoods to outside inspection — not only by third-party auditors (granted, ones that they themselves select) but also by independent researchers and members of the public.

The first round of designations was straightforward, capturing mainly traditional social media platforms like Facebook and search engines like Google. More recent designations, including those of 

, Roblox, and ChatGPT, involved the Commission applying fixed statutory definitions to services that have changed considerably since those definitions were drafted. And it applied them correctly — considering what these services actually let users do today, not what they were originally built or marketed to be.

But, while the new obligations make sense from the standpoint of increasing transparency and accountability, they also raise new and thorny implementation questions.

Roblox is a game-creation engine. Users build experiences and publish them for tens of millions of other users to play, which makes it an online platform under the DSA’s definition. But much of the games industry has assumed the law doesn’t concern them, based on the theory that their business is interactive entertainment rather than user-to-user communication and user-generated content publication.

 — a multi-stakeholder group convened by the 

NYU Stern Center for Business and Human Rights

 to help steer gaming regulation toward proportionate and evidence-based requirements —addressed this directly in its December 2024 

 to the Commission. We argued that gaming platforms already fall within the DSA’s scope as hosting services, that some qualify as online platforms, and that some, like Roblox, would eventually qualify as VLOPs. The social dimension of online games — user-generated experiences, matchmaking, in-game voice and text chat — blurs the line between gaming and social media. Whether a service carries “game” on its label should be less relevant than the type and scale of social functionalities it provides. Roblox’s designation is exactly what our argument anticipated.

ChatGPT began as a large language model with a consumer chatbot attached to it that answered queries with predictions based on data the model had absorbed during training. This setup made the chatbot unreliable on current events and prone to “hallucinations,” leading OpenAI to 

 search and other tools into the product. But a user no longer has to select a “search” mode. ChatGPT decides on its own when a query calls for the live web and returns an answer with links to sources. The Commission described the result as a “hybrid service” that qualifies as an online search engine, and OpenAI did not contest the characterization.

Why further guidance is needed for effective implementation

Article 38 of the DSA requires VLOPs and VLOSEs that use recommender systems to provide at least one recommender option that is not based on profiling. Applied to Roblox, which ranks user-created experiences in its Recommended for You page, this provision is relatively straightforward. For ChatGPT, it is less so.

When ChatGPT conducts a search in order to respond to a query, an algorithm selects which sources to fetch, draw on, and cite. But the result is a synthesized answer rather than a ranked list, where it is less obvious what information the system prioritized or the “relative order or prominence” of that information. Moreover, what counts as a non-profiling-based option is debatable. ChatGPT’s use of persistent “memory” across sessions to shape answers should count as profiling, a definition set out in the EU’s General Data Protection Regulation (GDPR). Thus, allowing users to toggle off memory and turn on “temporary chat” could satisfy the requirement to provide a non-profiling option.

Yet, even when a user turns on these options in the privacy settings, the system returns answers based on the accumulated conversation rather than a query in isolation. In a single session, a user may reveal their occupation, health concerns, location, political leanings, or emotions, and the chatbot uses that information to return answers 

 that session. Thus, it is not clear how a non-profiling-based recommendation option, as contemplated in article 38, would apply to conversational systems.

The ad repository requirement is especially significant for ChatGPT as OpenAI moves into advertising. On August 24, one week before the designation, OpenAI 

 showing ads in ChatGPT across 31 markets. The ads appear as “sponsored” cards beneath the chatbot’s answers, are shown only to users on the free and low-cost Go tiers, and are tailored to the topic of the conversation the user is having. Under article 39, the ad repository has to disclose “whether the advertisement was intended to be presented specifically to one or more particular groups of recipients of the service and if so, the main parameters used for that purpose.”

Applied to chatbots, what would make this provision meaningful would be to require OpenAI to disclose the “context hints” that determine where ads are placed within conversations. OpenAI’s 

 explains that advertisers supply hints describing “the conversations, topics, or keywords where their products may be relevant,” and that these guide which conversations an ad appears beside. Publishing them per advertisement would show whether a company can ask to appear beside a user describing sensitive personal information, such as health symptoms and interpersonal issues.

This information is very relevant to the public interest, but OpenAI may contest having to disclose it. The company might argue that the parameter disclosure requirement applies to ads that target “one or more particular groups of recipients of the service,” whereas context hints describe conversations rather than people. To ensure effective implementation, the European Commission should issue guidelines, as it is empowered to do under Article 39, making clear that OpenAI’s context hints count as parameters that should be disclosed.

The data access regime also poses unresolved questions, especially for gaming platforms that process ephemeral data.

Data access for traditional social media is relatively straightforward because in most cases posts are static and public. But applied to gaming platforms, the data access provision is less obvious. Games do not store their most revealing data the way social media platforms do. Much of what a researcher would need to study grooming, harassment, or radicalization on Roblox, for example, is ephemeral — real-time voice and text, gameplay behavior, interactions that are never retained long-term. 

, the Working Group set out what a workable regime would look like for gaming platforms like Roblox, balancing researchers’ valid interests in examining ephemeral communications with the privacy, technical, and financial constraints that make ephemeral data genuinely hard to share.

The Commission’s August 31 designations were the right call. It should now clarify what implementation means for services the DSA’s drafters did not have in mind, ensuring that compliance actually translates into better transparency and accountability. Other services, meanwhile, should stop assuming they fall out of scope because of their industry name or service label, and start examining feature by feature what service they actually operate.

This article was corrected on Sept. 3, 2026

﻿Germany’s Federal Chancellery building in Berlin on a sunny day. With the German and EU flags in the foreground. 

Germany’s Federal Chancellery building in Berlin on July 19, 2025 (Photo by Michael Nguyen/NurPhoto) 

In June, Germany’s National Security Council 

decided to create an AI Security Institute

, known as DE-AISI. The issue gained wider attention in Germany during a 

public debate over the potential use of Claude Mythos

 in cyberattacks against the country’s financial system.The new institute will analyze the capabilities and risks of advanced AI models, and work with international partners toward common standards. But the government has yet to settle many of the questions that will determine what the new institute can accomplish.

Until two days ago, German cities Munich, Darmstadt, Saarbrücken and Berlin were competing to host it. 

 pointed to its technology companies and research community. 

 pointed to its engineering and cybersecurity expertise, and 

 to its renowned AI and computer science cluster. In the end, 

However, the far more consequential choices concern the AI safety institute’s mandate, its relationship with other agencies, and what happens when its researchers uncover a serious risk.

Germany’s digital industry association, 

, favors a narrow research mandate. It wants DE-AISI to concentrate on systemic threats posed by frontier models to national security and technological sovereignty, without taking on regulation or anything else.Meanwhile, researchers from the German Research Center for Artificial Intelligence (DFKI) 

hat the institute’s remit also extends to

 privacy, ethics, human oversight, and the use of AI in employment, healthcare, public services, and law enforcement.

This reflects two different ideas of what an AI institute should be for. One treats it as a specialist laboratory for study of frontier-model threats. The other sees it as a public institution that should examine how AI affects people and institutions across society.

In either case, the public value of the institute will greatly depend on how effectively it can transform 

comparable AI safety institutes that have been established

 in the United Kingdom, United States, Australia, Canada, France, India, Japan, Kenya, South Korea, Singapore, and 

 in recent years, with a variety of different approaches and structures.

That lesson applies to an institute’s mandate and governance, as well as to how it conducts its research. Frontier-model evaluations often place systems in deliberately permissive environments, with safeguards reduced, access to tools, and the open internet enabled. These conditions can expose capabilities that ordinary testing would miss. They can also allow an evaluation to spill into the real world.

Testing dangerous systems can itself be dangerous

AI agents under evaluation by the UK AI Security Institute (AISI)

 took unauthorized actions on the live internet. In the most serious case, one agent tried to insert malicious code into a real open-source project on GitHub. It created fake identities and used them to pressure a project maintainer to approve the code.

AISI’s own team identified the behavior and disclosed the incident. They also acknowledged they had deliberately provided the agent with internet access and had not actively monitored its actions.

To its credit, the institute did not bury the episode, but instead set out to change its testing practices. Their disclosure gave other organizations a chance to learn from the failure.

Germany’s DE-AISI will need strict controls on external access, continuous monitoring, clear limits on what agents may do, and an independent process for reviewing serious incidents. Its own security practices should face the same scrutiny it applies to AI companies.

Exposing a risk is not the same as addressing it

The UK experience also shows that AI institutes can identify dangerous capabilities, but cannot necessarily require a company to strengthen its safeguards, delay a release, or submit a model for further evaluation. Access depends on companies choosing to cooperate.

That dependence creates an imbalance: the institute needs access to proprietary models to conduct meaningful tests, but criticizing a provider too forcefully could threaten future cooperation. Meanwhile, companies can cite government testing in their model reports without necessarily disclosing what was found, how they responded, or whether the institute considered those measures adequate.

Germany must establish what happens after DE-AISI researchers identify a serious threat. Which authority can demand corrective action? When must the Federal Office for Information Security, the Federal Network Agency, or the European AI Office become involved? Under what circumstances should the public be informed?

Without a formal route from evaluation to action, DE-AISI could produce technically impressive research that leads to zero changes.

A new government can change an institute’s purpose

AI institutes can also quickly be politicized. In 2025, the Trump administration

 as the Center for AI Standards and Innovation. Its revised mission placed greater weight on national security, innovation, and American competitiveness.

The change in name and mission showed how easily a 

new administration can redirect an institute toward a different understanding of AI governance

. Research priorities, relationships with companies, and decisions about which risks deserve attention can all follow.

Germany cannot entirely remove politics from a publicly funded body, nor should it try. Decisions about acceptable risks and public priorities are political. Scientific findings, however, should not depend on whether they fit the message of the ministry in charge.

Stable funding, transparent research priorities, and the freedom to publish inconvenient findings would give DE-AISI some protection against abrupt political shifts. Without those safeguards, its agenda could shift with each government or become an extension of industrial policy.

Different government entities have overlapping interests in AI safety, and there are different ways to divide or collaborate on the work. For instance rather than create a separate legal entity, France’s National Institute for AI Evaluation and Security brings together four existing organizations with expertise in cybersecurity, digital research, technical testing, and regulation.

Germany has chosen a different structure but a similar starting point, as 

 from the Federal Office for Information Security as well as the Federal Network Agency. In addition, the European AI Office evaluates general-purpose models and has powers to enforce the EU AI Act.

Germany therefore needs to identify which work will be done by DE-ASISI instead of another federal agency or the European Union. It will be important to decide from the start which evaluations DE-AISI will conduct itself, when it will support existing agencies, and which body is expected to act on its findings. It all depends on what the focus of research will be. DE-AISI could also advise German authorities on how to challenge assessments produced by AI companies or foreign institutes, in a way that feeds into European standards.Without a clear division of labor, Germany risks producing overlapping research, competing assessments, and confusion over who must respond whenever a serious threat is identified.

Finally, how will Germany know whether the institute is working well?

Regardless of the number of research papers, model evaluations, international agreements, and conference appearances, a technically capable institute will remain peripheral if ministers, regulators, and companies 

In order to show its impact, DE-AISI would need to track whether an evaluation led a company to alter a model, helped an authority prevent an incident, improved defenses in critical infrastructure, or produced testing tools adopted by others.

Some results will remain confidential for legitimate security reasons. That makes public reporting harder, but aggregated figures, anonymized case studies, accounts of policy changes, and explanations of how its work affected government preparedness need to be transparent to create value for the research community and public advocacy.

Whether DE-AISI should concentrate on frontier-model security or examine a wider range of harms will probably not be settled anytime soon, nor is there a perfect model abroad for Germany to copy. Still, Germany has the advantage of being able to borrow from what worked elsewhere and avoid at least some of what did not. This means looking beyond where in the country the institute will sit and how many researchers it will employ. The real measure of DE-AISI will be whether anyone acts on the problems it uncovers.

Correction (Sept. 3, 2026): A previous version of this post implied that four cities were competing to host Germany's AI institute. In fact, Berlin was chosen two days before publication.

Since August 2, providers and deployers of certain AI systems have 

 to disclose AI interactions, mark AI-generated or manipulated content, and label certain deepfakes and AI-generated text on matters of public interest. The 

 the rules are intended to reduce deception and manipulation and strengthen the integrity of Europe’s information environment.

The moment comes against a backdrop of concerns about 

 in Europe. As European democracies face risks from foreign interference, internal political divisions, and rapidly evolving AI systems, the newly enforceable AI Act provisions can strengthen transparency, accountability, and institutional readiness, but they cannot by themselves address the broader threats facing European democracies. The question, therefore, is whether greater transparency actually translates into greater democratic resilience.

Several regulatory developments are noteworthy. First, 

Article 50's transparency rules took effect

. Providers of certain AI systems must ensure that people are informed when they interact directly with an AI system, while providers of systems that generate or manipulate synthetic content must ensure that outputs are marked in a machine-readable format where required. Deployers have separate obligations concerning the disclosure of deepfakes and AI-generated public-interest text. These obligations come with several exceptions, however. They include law enforcement use, minor editorial assistance, artistic or satirical works, and content that has undergone human review with clear editorial responsibility.

Second, the AI Office gained full enforcement powers over general-purpose AI (GPAI) providers. Although the GPAI obligations had already applied 

, full enforcement by the Commission did not start until August 2. Third, Article 4’s 

, which has applied since February 2025, entered a new phase of supervision and enforcement. The provision requires providers and deployers to take measures to ensure a sufficient level of AI literacy among staff and others using AI systems on their behalf, taking into account their technical knowledge, experience, education, and the context in which the systems are used.

For democratic resilience, however, three aspects of this implementation deserve particular attention: transparency, accountability, and institutional readiness.

Where the newly enforceable rules can help

First, newly applicable rules can strengthen transparency. Democracy is built on the principle that political power stems from the informed 

 of the governed. Without the ability to tell what they are looking at, voters cannot meaningfully analyze information, judge leadership, or hold power accountable. Article 50 and the GPAI provisions aim to prevent citizens from being deceived by synthetic media by requiring machine-readable marking of AI-generated content, mandatory disclosure of AI interaction, and documentation from GPAI providers. In doing so, the rules also aim to preserve institutional trust: citizens who stop believing what they see are citizens who may disengage from democratic participation altogether, or lean toward leaders who promise order over truth.

 of watermarking highlights both the promise and the limits of this approach. Honoring its commitment under the EU's Code of Practice on Transparency, Anthropic started embedding an invisible watermark in text generated by Claude models released after August 2, alongside provenance metadata attached to supported files. On the one hand, a positive aspect is that the new technology was rolled out globally rather than being limited to EU users.

On the other hand, watermarking only addresses the initial generation phase. The watermark remains vulnerable throughout a much longer lifecycle spanning distribution, modification, detection, and human interpretation. Anthropic itself 

 that the absence of a watermark doesn't guarantee content wasn't AI-generated. Indeed, a screenshot, format conversion, or simple file re-save suffice to break the chain of provenance. Article 50 therefore targets a single point in the chain, but democratic resilience depends on what endures through several such points.

, an essential element of democracy as it ensures that power faces consequences for its actions. The GPAI provisions require providers to document training content, adopt copyright-compliant practices, and maintain risk-management records that the AI Office has the authority to request and assess. Although this does not prevent harm directly, it creates the documentation that is necessary for any enforcement.

Third, the rules contribute to institutional readiness. The latter is necessary to prevent judges, police officers, or civil servants from accepting algorithmic outputs as absolute truth, for instance, when setting bail. Training staff to understand a system's capabilities and limitations helps maintain human oversight that might otherwise default to machines. Article 4's literacy obligation, in force since 2025, pushes public bodies, campaigns, and newsrooms to achieve a baseline competence in spotting AI-generated content. While this might seem to be a small detail, transparency and accountability fail without institutional readiness: an official who is not aware of a disclosure requirement will not enforce it.

While the AI Act introduces essential regulatory guardrails, loopholes remain that will likely limit its effectiveness in safeguarding democracy.

To begin with, some actors will likely remain indifferent to the AI Act regardless of fines. Much of the disinformation targeting European democracies originates from 

. The foreign states are beyond the AI Act's jurisdiction and thus cannot be deterred directly. Its impact is instead indirect: constraining the commercial platforms, generative tools, and distribution channels used by foreign actors. Whether this indirect pressure actually limits state-backed operations, which can rely on non-compliant tools, remains an open question.

The AI Act also regulates companies placing systems on the EU market. However, it has almost no practical reach over a malicious actor running a modified, open-weight model on private infrastructure. Such an actor can generate harmful content entirely outside any provider's oversight and never trigger the newly enforceable rules.

A related problem is laundering. A bad actor can upload an illegal deepfake first on an unmonitored app like Telegram, where copying and screenshotting strip any hidden tracking data before regular users share it further. By the time the fake reaches major platforms, detection becomes more challenging, and Article 50's watermarking requirement offers little protection, since its markers were likely erased well before the content reached regulated platforms.

Mandatory labeling can also create an unintended defense. Bad actors can dismiss genuine footage simply by claiming it lacks a watermark or looks altered. This will help turn a tool built to expose fakes into one that helps deny reality. Should these tactics become popular, it will strike at the heart of democratic accountability. Indeed, if voters cannot agree on basic facts, how can they vote on an informed basis or hold leaders responsible?

Finally, the newly enforceable provisions do not address deeper societal problems like political polarization and declining institutional trust. People may choose to believe fake content or ignore official warnings no matter how prominently they are displayed. While the AI Act regulates how AI content is generated, it says nothing about why people are inclined to believe fakes in the first place.

Given these limits, strengthening democratic resilience further requires tools beyond the AI Act. 

 is the most obvious complement. Where the AI Act addresses content generation, the DSA targets the algorithms that amplify it, and can force major platforms to detect and restrict viral fake media regardless of its origin. The mechanism works when it has teeth: once ChatGPT crossed the 45-million EU user threshold, 

 as a Very Large Online Search Engine, subjecting it to the DSA’s strictest rules. But the DSA has its own limits. Telegram, by contrast, has reported staying just below that threshold for over two years despite EU scrutiny, 

 the heightened obligations around detecting and mitigating coordinated manipulation.

Technical improvement would also help. Anthropic’s acknowledgment that provenance information can be lost or circumvented illustrates the need for more robust and interoperable approaches to content provenance. Greater harmonization of marking standards across platforms and file formats could make it harder for provenance information to disappear as content moves through the information ecosystem.

Beyond technology, institutions matter. Election commissions need the capacity to respond swiftly to AI-enabled incidents during active campaigns rather than through retroactive investigations; media literacy efforts should extend beyond Article 4's scope to reach the wider public; and allied democracies would benefit from closer coordination against the shared threat posed by state actors seeking to undermine their political systems.

The newly applicable provisions of the EU AI Act could strengthen democratic resilience, but their impact should not be measured simply by how much AI-generated content receives a label.

The more important question is what happens after that label appears. If provenance information survives distribution, if platforms and institutions can interpret it, if journalists and election authorities can act on it, and if citizens trust and understand the signals they receive, greater transparency can contribute meaningfully to democratic resilience.

However, some actors will never comply voluntarily, and no labeling regime can eliminate political manipulation, polarization, or declining institutional trust.

The AI Act therefore should be understood as one layer of a broader resilience strategy. It needs to work alongside the DSA and other measures aimed at platform governance, election security, media literacy, technical provenance, and institutional capacity.

Ultimately, the AI Act can make synthetic content easier to identify. Whether that makes European democracies more resilient will depend on what governments, platforms, institutions, journalists, and citizens are able to do with that information.

Petra Molnar and Liz Carolan are both fellows at Tech Policy Press.

Over two days in late July, an estimated 

 crossed the border from Morocco to Ceuta, a Spanish enclave in North Africa. Many swam a 

 of open sea to make the crossing, and at least 100 

 doing so, according to the enclave’s mayor. In the lead-up to those chaotic days, a wave of disinformation on social media 

 with young people in Morocco that a recent Spanish Supreme Court ruling had opened the border to Ceuta and onto Europe, and that crossing into Ceuta would be easy.

This week, Spain's Prime Minister Pedro Sánchez 

 of the disinformation that preceded the Ceuta crossings, citing research by the European External Action Service, the EU’s diplomatic service, that traced the campaign to Russian and Israeli networks operating alongside what he described as an ultra-right international network that "always uses migration to attack both Spain and Europe." Sánchez said the two governments had seized on the issue because of Madrid's positions on the wars in Ukraine and Gaza, and that investigators had so far found no evidence of Moroccan state involvement. Israel's Foreign Minister Gideon Sa'ar 

, accusing Sánchez of "lying shamelessly," and Russia has also denied any role.

This disinformation effort has striking similarities to patterns seen in 

content targeting people on the move trying to reach the United States

, and in both cases the campaigns worked not because the people who acted on them were credulous, but because the rules governing who may cross a border have been made so opaque, so changeable, and so badly communicated that a video posted by a stranger is often the most accessible account available.

, found dozens of burner profiles on TikTok and Instagram that purported to be young Moroccan content creators who had made the journey and were celebrating a newfound freedom in Ceuta. The videos, structured as influencer content, often contained practical advice on making the crossing, from the equipment needed to advice for children and non-swimmers. While many accounts were shut down, they later reappeared under new handles.

 suggests that young people traveled from across Morocco in response to online rumors and stories, like waitress Dounia Kherradi, who lived a nine-and-a-half-hour drive away, and 16-year-old Ayoub El Abyad, who told reporters that he and his cousin decided to drive 12 hours to the crossing after seeing videos on Instagram. What they found was drastically different than the upbeat soundtracks of Reels suggested; those who had crossed the border told reporters 

 that they had to pass dead bodies on the treacherous swim to Ceuta, that they were met with violence when they arrived, and denied food in stores. They then learned there was no way for them to travel on to Europe and that the rumors had been lies. By the end of the weekend, all but 

None of the mechanics on display at Ceuta are new, and there are clear similarities in the grammar of videos shared of how people on the move try to reach the United States. In March of last year the Associated Press reviewed 

hundreds of TikTok videos posted by Mexican migration facilitation networks

 and found similar themes that El Español documented on the Moroccan side of the Strait; upbeat soundtracks, emoji-laden captions, dune buggies racing toward the border wall, jet skis and helicopters and tunnels offered as premium options, facilitators slicing open cacti so their clients can drink, a field of lettuce captioned "The American fields are ready!,” and, at the end, the satisfied-customer testimonial.

In Ceuta as in the Americas, these videos and imagery function as advertisements, painting a picture of border crossing as a manageable logistical problem with a happy ending.

The commercial layer around this content is by now well established. In Spain, the Prime Minister initially blamed human-trafficking 

 for the disinformation, as they stood to benefit financially from attempts to cross the border.

Meta earned revenue from smuggling and scam advertisements

 targeting people trying to reach the US, and the pattern intensified once the American asylum system was funneled through a single mobile application. When CBP One became the central route to an appointment at a port of entry, a secondary economy grew up around it involving 

 sold to people who arrived at the border to find nothing waiting for them, paid intermediaries charging anywhere from a few dollars to several thousand, and 

GPS-spoofing workarounds circulated in WhatsApp groups

 by people locked out of the app's geographic requirements. The Tech Transparency Project, mapping the same terrain, found Facebook and WhatsApp 

saturated with false claims about policy changes

 and invented special provisions for pregnant women and parents traveling with children.

The rumor that circulated in Morocco in July —that a court ruling had quietly opened a route — is a recognizable instance of a well-documented pattern, and the difficulty it presents is one the coverage of Ceuta has largely avoided. Deceptive content does not reach people on the move through a separate channel from accurate content; it reaches them through the same feeds, in the same visual idiom, and often from accounts that are indistinguishable at a glance.

For example, alongside the burner accounts are the Venezuelan and Colombian creators who 

filmed their own crossings of the Darién Gap

, some of whom accumulated tens of millions of views, and whose videos 

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe