Summary

Across all three platforms, the findings point to structural gaps in enforcement design that allowed manipulation campaigns to operate with limited interruption throughout the election period.

The most detailed case is of the Facebook page 'Kutyakalandok', which ran 121 paid political advertisements promoting five AI-generated videos between February 20, 2026 and April 10, 2026.

Meta classified all 121 advertisements as political and actioned individual removals, yet the page's operators consistently re-published identical content under new advertisement IDs (in 96% of observed cases on the same day the preceding advertisement ended, or the day after).

This pattern, observable directly from Meta’s own Ad Library data, constitutes evidence that Meta’s enforcement mechanisms were being systematically circumvented without triggering escalated action.

 At peak, 17 concurrent advertisements for a single video ran simultaneously.

Meta's system, designed to process individual advertisements, showed no apparent escalation in response to this cross-ID behavioural pattern, allowing the campaign to maintain a near-continuous presence for 50 days despite repeated removal actions.

Taken together, the four Facebook pages examined in this report point to a single structural problem, rather than a series of moderation errors. As observable in the Ad Library, Meta's enforcement operated at the level of the individual advertisement: it was frequently triggered, often reasonably quickly, and entirely reactively. The operations Meta’s systems face are organised at a different level, at which almost no enforcement response was observable. Nothing in the outcomes indicates that identical videos were linked across advertisements, relaunch chains produced no escalated response, and a page created overnight, from which the recently removed campaign continued, without being connected to its predecessor.

The lack of transparency in Meta’s infrastructure compounds the problem. The Ad Library made this analysis possible, but its omissions, such as the timing of political classification, the distinction between removal and expiry, the presence of AI-content labels, information about page operators, mean that researchers and other external observers working from the public Ad Library alone cannot reconstruct how the platform's enforcement actually unfolded, or verify its compliance with its own stated policies.

Nevertheless, Meta's overall removal rate was the strongest recorded among the three platforms. Response times averaged under one day, dropping to one to two hours on election day itself. However, additional anomalies were noted: three Facebook pages publishing unlabelled AI-generated videos of political figures were handled inconsistently, with two pages removed entirely rather than labelled per Meta's stated manipulated media policy, and the third page apparently seeing no enforcement action at all.

Google's performance was that much more concerning. Political Capital reported 15 advertisements to the platform – with content substantially identical to material which Meta had already removed.

Google issued no written responses and removed none of the advertisements, recording a 0% removal rate. This outcome raises serious questions about Google's compliance with its obligations under the Digital Services Act (DSA).

TikTok received a referral concerning a network of 105 accounts exhibiting coordinated inauthentic behaviour. It issued a determination that no violations had been detected, but the speed of case closure and the absence of any explanatory communication made it impossible for researchers to assess whether a substantive review had taken place.

The lack of transparency is itself a finding: without visibility into TikTok's internal process, the "no violations" conclusion cannot be meaningfully evaluated.

The cumulative evidence from the 2026 Hungarian election monitoring period points not to isolated moderation errors, but to potentially structural deficiencies in the enforcement architectures of all three platforms reviewed.

According to available data, enforcement frameworks responded to individual reported items but lacked the capacity to detect and act on patterns of behaviour that emerged across multiple items, accounts, or campaigns. Such failures warrant further investigations under Articles 34 (Risk assessment) and 35 (Mitigation of risks) of the DSA.

The final months of the 2026 Hungarian campaign thus offer a concentrated case study, within a single EU Member State, of what per-item enforcement yields when confronted with adaptive, AI-based advertising operations – prohibited political content in near-continuous circulation, synthetic depictions of real candidates in paid distribution, and constituency-scale audiences reached before and despite intervention. The evasion methods documented here required little more than persistence, and little suggests that the exploited vulnerabilities have been addressed. The operation across Facebook accounts Kutyakalandok, Nem a mi háborúnk, Nem a mi utunk, and Tiszta Jövőt! ultimately failed to achieve what appears to have been its electoral objective, but the vulnerabilities it exploited - the reupload gap, the multi-ID redundancy strategy, the absence of cross-advertisement pattern detection - remain seemingly unaddressed and available to future actors at the time of writing.