Spain’s Partido Popular published a 51-second AI-generated video on 21 October with AI likenesses of people linked to the governing Socialists, each delivering a line built on the word “zurdo”. The party put its own notice on the video saying it had been made with artificial intelligence. The notice sat at the bottom of the frame in miniature, nearly transparent type. El País described the video as the first time a Spanish party had used AI to impersonate the identity and voice of political figures. Within 11 hours it had nearly 2 million views across the PP’s X, Instagram and TikTok accounts, with 18,000 shares and more than 66,000 likes. The X post alone showed 2.3 million views when checked on 25 October.

A criminal complaint followed. Leire Díez, a former Socialist member whose likeness appears in the video, filed a querella at Spain’s national court. So what did the label buy the PP? On the reading nine jurists gave the newspaper El País, it bought a defensible posture: labelled and satirical, the video sits inside wide free-expression margins. It bought nothing about reach. And it left open whether the notice worked at all. Santiago Mediano, one of the jurists El País named, told the paper the small transparent notice was insufficient: “no es suficiente”.

That split is the story. A label can shape legal exposure. The experimental evidence says it does much less to belief, and close to nothing to sharing.

The notice was the party’s own, and it was easy to miss

Two El País pieces and the Guardian describe the same artefact: a caption the PP itself burned into the video, at the foot of the frame, in tiny transparent lettering that El País says “se hace imperceptible para el ojo durante casi todo el vídeo”: it becomes imperceptible to the eye for almost the whole video. No source in this reporting describes a platform AI-content mark on the post, and none quotes the notice’s exact words. Whether the notice travelled with reposts is unverified. Calling it a warning interstitial, something that interrupts a viewer before playback, would overstate what the evidence shows.

The doubt is not only about wording. Mediano, who leads the robotics and AI section at Madrid’s bar association, argued that a small transparent notice does not meet what disclosure is for, and pointed to the European AI Act’s duty on anyone deploying a deepfake to make clear the content is AI-made. El País reports the same rule applying in Spain since August 2026, citing the jurist Borja Adsuara, with a warning required on AI-generated content and an exception for satire where there is no risk of confusion. No authority has applied it to this video.

The complaint alleges three offences, and no court has ruled

Díez’s querella runs before Judge Santiago Pedraz at the Audiencia Nacional, Spain’s national court (Confilegal, 23 October). It alleges calumnias con publicidad under articles 205 and 206 of the criminal code, an offence against moral integrity, and suplantación de identidad digital. The video has an AI likeness of Díez saying she hired thugs to “get rid of” a judge, a statement the querella says is false. Díez is seeking precautionary removal of the video from the PP’s accounts and anywhere else it has circulated while criminal responsibility is decided. As of 25 October the complaint was pending, with no ruling, no hearing outcome and no removal order reported.

The PP’s on-record answer is that intent was never to deceive. Borja Sémper, the party’s national spokesperson, told reporters the party would have “a problem” if there had been “malicious intent to make people believe that the characters who appear there are real”, and said everyone can see the video is a way of using digital tools to highlight the hypocrisy he ascribes to the Socialist party (Guardian).

Filing is not proof. Confilegal, the legal outlet that reported the complaint in detail, notes that using an AI-generated image is not by itself the offence of identity theft, so a court would have to decide whether the elements of any charge are met. The label does not close that claim.

Several of the video’s lines track matters under investigation rather than established findings. The lines aimed at Begoña Gómez and José Luis Rodríguez Zapatero gesture at cases the Guardian describes as ongoing investigations, and both have denied wrongdoing.

Nine jurists, and what El País actually reported

El País said it consulted “más de media docena de juristas”, more than half a dozen. Counting the jurists it names and quotes gives nine, which matches the Guardian’s “nine legal experts” (Guardian). What El País reports them agreeing on is the breadth of the margin: Spanish and European case law grants “enormes márgenes” to expression, widest in political debate, with room for satire and parody. That describes how far a court may let speech run. It stops short of a ruling that this video is lawful. The Guardian’s gloss that the experts judged the ad “not illegal” is a fair reading of the majority view, but it is stronger than El País’s own words, and it concerns a complaint that is still open.

The dissents carry as much weight as the majority. Mediano faulted the notice. Ofelia Tejerina said the defamation route “se podría intentar, pero no sé si tendría recorrido. Y ese recorrido, además, sería muy corto”: it could be tried but might not travel far, and what distance it had would be short. Borja Adsuara said judges tend to shelve such cases and warned that trying to suppress the video risks a Streisand effect. Juan Antonio Lascuraín raised the European Court of Human Rights’ chilling-effect doctrine. Underlying several of those answers is a distinction that cuts across the majority and the dissents: conduct can be legally permitted and still, in Lascuraín’s phrase, neither exemplary nor ethically correct. On this account the label works as mitigation: it removes the problem the jurists say would matter most, content presented so that people could take it as real, and leaves the rest to a court.

The base rate makes a label a weak constraint

The label’s legal work looks smaller next to the volume of AI political advertising that carries none. The Wesleyan Media Project’s 15 October 2026 release counts 220 AI-generated ads in the 2026 cycle, more than $137m in spending, across 34 states. Republican candidates or groups sponsored 81% of the ads and 88% of the spending; Democrats sponsored 12%. Just 27% of the ads Wesleyan identified carry a disclaimer. The project found less than $50,000 in AI-ad spending in the 2024 cycle, which makes the practice new at this scale.

The project states its own limit: it relies on coders, on the presence of disclaimers and on AdImpact tagging to spot AI ads, and when it is unsure it leaves an ad out, which makes the counts and the spending “likely on the low side”. A disclaimer is also one of the ways an ad gets identified, so ads found that way carry one by definition. Read carefully, 27% is the share of the ads Wesleyan identified that carry a disclaimer, not a base rate of disclosure across AI political advertising.

What a label cannot prove or change

Experimental work gives the label less credit than the campaign argument does. In two preregistered experiments run in 2023 on misleading AI-generated images, with 7,579 US adults between them, Wittenberg and colleagues found that labelling reduced belief in a post’s claims by about 0.21 to 0.25 standard deviations and cut the image’s perceived credibility by 0.34 SD. The plain “AI-Generated” label was the weakest tested, at roughly 0.14 to 0.18 SD. Across the labels, belief effects ran from about 0.1 to 0.35 SD. Engagement thinned further: aggregated across labels, stated sharing fell modestly, but the standalone “AI-Generated” label had no statistically significant effect on sharing, liking, commenting or seeking more information in either study. The labels that did reduce stated sharing were the ones that did not mention AI. The studies used still images, US online samples and self-reported intentions, which limits how far they carry to a video in a feed.

Two newer results point the same way, and both need a caveat. A 2026 preprint on deepfake videos, with a reported N of 7,107, found that labels made viewers rate clips as less deceptive and left a speaker’s arguments equally persuasive either way, while making people less likely to watch or share when they could choose. It is not peer reviewed and its full text was inaccessible, so no effect sizes can be quoted from it. A cross-national experiment in Computers in Human Behavior, spanning 5,435 participants in five countries, found that message-style warnings reduced perceived accuracy without improving people’s ability to tell true videos from false ones, and that individual and country differences predicted responses far better than the format of the warning. That result is abstract-level here, so no effect sizes are asserted. One more finding unsettles a common assumption: a 2024 American Political Science Review study of how politicians deflect scandal found that casting evidence as fake helped against text reports but was largely ineffective against video evidence.

A label is not proof either way about a given artefact. A watermark or disclosure can point to one company’s pipeline while leaving an unlabelled file an open question, the point we made about SynthID-style watermarks. Labelling is not nothing. It is a small, conditional nudge to belief and close to nothing for sharing, weakest in its plainest form, and no guarantee that content is trustworthy.

One ad is not a pattern

El País reports that Vox began publishing AI videos that same afternoon, and that the PP video answered a filmed PSOE video on Pedro Sánchez’s account from earlier that day. In the United States, Mississippi’s attorney general told a Democratic Senate candidate on 23 October that an AI-generated video depicting Republican Senator Cindy Hyde-Smith appeared to violate state law and asked him to take it down; he complied (Mississippi attorney general’s statement). A federal judge blocked enforcement of Montana’s deepfake-ad law on 17 September. Two documented cases and a measured base rate are context for what comes next in Spain. They are not yet a pattern.

Where a complaint can still land

The fastest remedy found in this reporting did not turn on the label at all. On 22 October, Spain’s electoral authority, the Junta Electoral Central, partly upheld a PP complaint about PSOE Instagram memes and videos that imitated PP logos and colours, and ordered the material removed without opening a sanction file. Its reasoning was that another party’s symbols can be used critically or ironically and are not by themselves unlawful when authorship is identifiable. JEC sources told El País, anonymously, that a complaint about the PP video would probably be admissible and would be a more serious case because it came from official party channels. No complaint about the PP video had been reported at the JEC as of 25 October, so there is no JEC decision on it to report.

The rule that would have spoken directly to a labelled AI likeness is not in force. The government sent a bill to Congress in July 2026 that would have made using a person’s voice or image without consent through AI illegitimate, with an exception for public figures in a creative, satirical or fictional context where the use of the technology is indicated. El País reports the bill “ha decaído con la convocatoria electoral”; it lapsed when the election was called. Spain votes on 29 November.

The PP’s X post was still live when checked on 25 October, and no removal order or platform takedown had been reported on any of the party’s accounts. The complaint is pending.