A study asked voters in Indonesia what would help them decide freely; it never tested a tool

interviews · Science and engineering ethics · la publicación, 30 abr 2026 · gratis

Un video de dos minutos les mostró a veinte votantes indonesios cómo una IA adivinaría su preferencia política y cómo unos errores a propósito en sus publicaciones podrían confundirla.

Versión breve · la versión detallada sigue, unos 5 min

Pregunte al chat de weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.

El estudio, de un vistazo
Quiénes: ciudadanos indonesios y actores políticos
Cuántos: veinte ciudadanos y cuatro actores políticos
Dónde: Indonesia
Cuándo: 2023 y enero de 2024
Tipo de estudio: entrevistas
Quién lo hizo: TU Delft, Países Bajos
El límite que importa: Un solo país, una sola plataforma, veinte ciudadanos; la herramienta nunca se construyó.

La próxima vez que reciba un anuncio político que parezca hablarle directamente a usted, no se pregunte si es falso. Pregúntese qué sabe esa campaña de usted que usted no le contó a nadie.
Lectura de weeklyAI

Ver · 0:34

Qué pasó. Entre 2023 y enero de 2024, en Indonesia, los autores del estudio entrevistaron a veinte ciudadanos y cuatro actores de campaña. A los ciudadanos les mostraron un video con una herramienta propuesta: primero les decía qué preferencia política les atribuía una IA con sus datos, y luego metía errores a propósito en sus próximos posts para engañar a esa IA. Unos lo aceptaron, otros dejaron de querer usarlo al pensar que los verían como poco creíbles, y quienes sabían de lenguaje dudaron de que los errores engañaran a los modelos actuales.

Qué puede revisar usted mismo esta noche. El documento menciona una página web, Bijak Memilih, con dos cuestionarios: uno muestra los temas que más le importan a quien responde, y el otro compara sus respuestas con lo que han hecho los partidos.

Qué dice la autoridad. El documento menciona reglas de campaña de la Comisión Electoral General de Indonesia (KPU), dentro de la ley UU No. 7/2017, sobre publicidad de campaña. También menciona que la ley indonesia de protección de datos, de 2022, ya permite objetar decisiones tomadas solo por sistemas automáticos.

Qué significa para usted

Ese video no probó que la herramienta cambie un voto: solo mostró reacciones, y nada más. Usted puede hacer los dos cuestionarios de Bijak Memilih para ver qué temas le importan y cómo se comparan sus respuestas con lo que han hecho los partidos. Cuando llegue la próxima elección, fíjese en si su autoridad electoral publica informes sobre publicidad dirigida.

Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9

Quién pagó: El Fondo de Dotación de Indonesia para la Educación (LPDP) financió a la primera autora con la subvención número KEP581/LPDP/LPDP.3/2022; el artículo no menciona otros financiadores ni préstamo de equipos o programas.

Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 913 palabras · unos 5 minLeerla →Cerrar

La regla y la evidencia: un video de dos minutos les mostró a veinte votantes indonesios cómo una IA adivinaría su preferencia política y cómo unos errores a propósito en sus publicaciones podrían confundirla

### EL ACTO

Un estudio de Auliya, Kudina, Ding y Van de Poel, publicado en Science and Engineering Ethics en 2026, propone que las herramientas tecnológicas para ayudar a decidir el voto deberían combinar la privacidad con otras condiciones: la reflexión sobre los propios sesgos, la conciencia de la manipulación política, el pensamiento crítico, la búsqueda de información diversa y la apertura a otras perspectivas¹.

El texto describe la tecnología propuesta así: la "AML, which uses an AI to create small non-random perturbations to change the output of another AI's system"². Y añade que la AML "may also be used to counterattack targeted micro-targeted campaigns based on AI-driven political profilers by using AML to obscure social media data, thereby impairing the accuracy of AI systems in profiling users' political preferences"³.

La herramienta concreta que se mostró a los participantes se llamó APP-AI, con dos funciones: una les decía cómo la inteligencia artificial infería sus preferencias políticas, y la otra introducía errores deliberados en sus publicaciones para enturbiar los datos⁴⁵.

### EL CASO QUE RESPONDE

Esta semana no se encontró ningún caso real que encaje con lo que el estudio describe.

### LO QUE SE MIDIÓ

Los autores piden que los principios de diseño propuestos se evalúen y validen más a fondo, y que futuros estudios los complementen con mediciones cuantitativas. Se entrevistó a veinte ciudadanos indonesios y cuatro actores políticos durante las elecciones de Indonesia de 2024, principalmente en Twitter, y a los ciudadanos se les mostró un vídeo de dos minutos con la herramienta propuesta para provocar la conversación.

Los veinte ciudadanos sabían que su contenido en redes estaba adaptado a ellos, aunque no sabían hasta qué punto se analizaban sus datos⁶. Algunos lo aceptaban como parte del negocio de las plataformas⁷. Pero reaccionaban de forma distinta ante la política que ante otros temas⁸.

Cuando algo les incomodaba —la privacidad, las burbujas de filtro, la dificultad para decidir, la desinformación—, desarrollaban sus propias defensas⁹. Una participante escribía palabras clave distintas en el buscador para confundir al algoritmo de Twitter sobre su preferencia política¹⁰.

La función que informaba sobre el perfilado entusiasmó a todos los participantes⁴. La que introducía errores recibió reacciones encontradas⁵. Los participantes con conocimientos de procesamiento de lenguaje natural dudaban de que los errores sirvieran para engañar a la máquina¹¹.

### LA RESPUESTA DE OTRA REGIÓN A LO MISMO

Esta semana no encontramos ningún tribunal, ley o autoridad de otro país que se pronuncie sobre este mismo método.

### LA LECTURA DE LA BIBLIOTECA

Así lo leemos nosotros. La propaganda política siempre ha funcionado igual: busca la grieta que ya existe en una sociedad, la alimenta y la lleva al extremo. Lo que cambia ahora es la escala y el silencio. Antes el rumor pasaba de boca en boca y dejaba rastro; ahora se deposita en cada pantalla sin que nadie vea el reparto. Un votante puede aprender a desconfiar de un panfleto, a buscar la fuente de una foto, a preguntar quién paga un anuncio. Lo que no puede verificar solo es lo que una máquina dedujo de él sin decírselo: qué infirió de sus silencios, de las cuentas que sigue, de las horas en que publica. Eso no está en ningún papel que él pueda consultar.

Por eso esperamos que en las elecciones que cubrimos la mayoría no viva los anuncios hechos a su medida como un problema democrático, sino como un riesgo más entre muchos: el trabajo, los ahorros, la seguridad de los hijos. Si eso ocurre, las campañas que apelen al miedo cotidiano moverán más que las que expliquen cómo funciona la microsegmentación. Sabríamos que nos equivocamos si viéramos a los votantes rechazar esos mensajes precisamente por considerarlos una amenaza a su vida privada, y que ese rechazo pesara más que cualquier preocupación económica al decidir. Usted puede comprobarlo con una pregunta sencilla: cuando lea encuestas o cobertura electoral, fíjese si el encuadre dominante es el riesgo personal o el riesgo cívico. Y pregunte a su autoridad electoral si el escrutinio y las auditorías se comunican en términos que la gente común pueda vincular con su vida diaria.

### LO QUE NO SE SABE

El estudio es de un solo país, Indonesia, y de una sola plataforma, Twitter, con veinte ciudadanos y cuatro actores políticos; no puede decirle qué permitió, auditó o encontró su propia autoridad electoral. La herramienta propuesta nunca se construyó: fue un vídeo de dos minutos para provocar la conversación, no un producto que usted pueda usar. Los propios autores advierten que la técnica de los errores deliberados es una carrera de gato y ratón: la inteligencia artificial que perfilea puede adaptarse y cambiar sus algoritmos cuando aprenda que existe¹². Y no hay ninguna medición de si la herramienta cambia la precisión del perfilado o el comportamiento del votante. Los autores piden repetir el estudio en otros países y plataformas antes de sacar conclusiones.

Y aquí está lo que usted puede hacer con todo esto. La próxima vez que reciba un anuncio político que parezca hablarle directamente a usted, no se pregunte si es falso. Pregúntese qué sabe esa campaña de usted que usted no le contó a nadie. Esa pregunta no necesita que nadie se la responda: es suya, y puede hacerla cada vez que abra el teléfono.

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Thus, these empirical observations and normative thought motivate the following design principles. We conclude that deliberative decision-making in elections can be enhanced when people (1) recognize and reflect on their personal predispositions and biases, (2) are aware of political manipulation, such as on social media, (3) think critically, (4) seek diverse information, and (5) are open to alternative perspectives."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "AML, which uses an AI to create small non-random perturbations to change the output of another AI’s system (Szegedy et al., 2014 ), offers a promising avenue for mitigating the risk associated with some AI applications (P.-Y. Chen, 2022 )."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "AML may also be used to counterattack targeted micro-targeted campaigns based on AI-driven political profilers by using AML to obscure social media data, thereby impairing the accuracy of AI systems in profiling users’ political preferences (Auliya et al., 2024 )"
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "All participants were excited about Feature-1 , which allowed them to be informed on how AI inferred their political preferences."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "However, Feature-2 , which introduced typos to blur social media data, received mixed reactions."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "All citizen participants were aware that their social media content was tailored for them, though they were unsure about the extent of data analysis involved."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Some participants accepted it as part of the platforms’ business process. They understood that social media platforms function based on a profit-oriented model, and enabling advertisers to target consumers is one of the methods employed to sustain the business."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Nonetheless, some participants reacted to micro-targeting political campaigns differently than in other non-political contexts."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "When participants feel unease with how social media platforms use their data, either because of concerns regarding privacy, filter bubbles, a hindrance in decision-making, or dissemination of unverified content, they develop some mechanisms to make them more at ease."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A participant using Twitter expressed that she usually manually put different keywords on the search bar to make the Twitter algorithms confusing her political preference"
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Additionally, participants with expertise in Natural Language Processing (NLP) also doubted the effectiveness of typos in ‘fooling’ AI."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Third , the key limitation of AML lies in the evolution of countermeasured AI systems. As noted in prior conceptual work preceding this study by Auliya et al. ( 2024 ), the proposal of using AML to counterattack microtargeting political campaigns based on AI resembles a cat-and-mouse game, in which profiling AI systems may adapt and change their algorithms after learning about the existence of AML."

Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.

interviews · Science and engineering ethics · the paper, 30 Apr 2026 · free

A study asked voters in Indonesia what would help them decide freely; it never tested a tool

Short version · the longer version follows, about 5 min

Ask the weeklyAI chat

Ask me about this study: who was studied, what it found, and what it does not say.

Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.

The study at a glance
Who: Indonesian citizens eligible to vote and political actors
How many: twenty citizens and four political actors
Where: Indonesia
When: 2024 elections
Kind of study: interviews
Who did it: TU Delft, Netherlands
The limit that matters: No tool was built or tested; only interviews about a proposed tool.

All participants were excited about Feature-1, which allowed them to be informed on how AI inferred their political preferences.
From the study

Watch · 0:34

Researchers at a Dutch university interviewed twenty Indonesian voters and four campaign workers during Indonesia's 2024 election, mostly about Twitter, and showed them a two-minute video of a proposed tool. The tool would have told people how AI guessed their politics and added typos to their posts to blur that guess. No tool was built or tested. The paper does name one way for a voter to check this: the website 'Bijak Memilih', which offers quizzes that compare your policy choices with what political parties have done.

The paper states one rule from Indonesia: its Personal Data Protection Law (UU No. 27/2022) gives people the right to object to decisions based only on automated profiling and the right to be told how their data is used, but political micro-targeting is not yet explicitly covered.

You can ask your own electoral authority what it allowed and audited, and whether it publishes campaign ad disclosures.

What this means for you

For your own vote, the study offers no proof that blurring your posts changes how you decide, so treat the proposed tool as an open question, not a fix. Ask your electoral authority what it allowed and audited, and what the campaign ads you saw were required to disclose.

Who paid: The Indonesia Endowment Fund for Education (LPDP) supported the first author under grant number KEP581/LPDP/LPDP.3/2022; the article states no other funders and no equipment or software lenders.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1079 words · about 5 minRead it →Close

The rule and the evidence: A study asked voters in Indonesia what would help them decide freely; it never tested a tool

### THE ACT, quoted exactly and linked

There is no resolution, ruling, audit or bill here to put in force. The document is an explorative study, received 2024 Dec 10, accepted 2026 Apr 9, issue date 2026¹. What it names is a proposal, not a rule: a tool the authors call Adversarial Privacy Preserving AI (APP-AI), shown to participants as a two-minute video with two features. Feature-1 informed users about their political preferences as profiled by AI that learned about them using their social media data; Feature-2 modified their social media data by putting deliberative typos in future social media posts using the concept of AML, where minor alterations to data input change the AI's output². The concept behind it is stated plainly: "AML, which uses an AI to create small non-random perturbations to change the output of another AI's system"³. Nothing in the document says this tool was built, released, enforced or made available to anyone.

### THE CASE IT ANSWERS

No real instance was found this week.

### WHAT WAS MEASURED

Nothing was measured in the units a voter could use. The study interviewed twenty Indonesian citizens eligible to vote in the 2024 elections and four political actors — candidates, campaign managers, or political/media consultants at management levels — with interviews conducted in 2023 and 2024, and analyzed the interviews with Reflexive Thematic Analysis. All participants were excited about Feature-1, which allowed them to be informed on how AI inferred their political preferences⁴. No effect size, no percentage change in profiling accuracy, and no persuasion or turnout figure appears anywhere in the passages. The study was not run in an election as a test of the tool; it was run as interviews during an election period, and the tool itself was never deployed.

What the interviews did record is a split in reaction. "All participants were excited about Feature-1, which allowed them to be informed on how AI inferred their political preferences"⁴. But "Feature-2, which introduced typos to blur social media data, received mixed reactions"⁵, and "participants with expertise in Natural Language Processing (NLP) also doubted the effectiveness of typos in 'fooling' AI"⁶. The authors themselves flag the deeper limit: the approach "resembles a cat-and-mouse game, in which profiling AI systems may adapt and change their algorithms after learning about the existence of AML"⁷.

### THE OTHER REGION'S ANSWER TO THE SAME THING

We found no passage this week giving another country's court, law or authority on adversarial machine learning aimed at political profiling.

### THE LIBRARY'S READING

Here is how we read it. A false rumor about a candidate has always done the same work before a ballot is cast: it plants a doubt that cannot be cleared, and it spreads because people enjoy passing it along. What is new is not the whisper but the aim. A machine can now assemble a picture of a person from what they post, like and follow, and a message can be built for that picture alone.

What this leads us to expect in homes like the reader's is that a political message will arrive already shaped to a mood or a grievance the voter has not announced. We would expect the voter to feel the message simply reflects their own thinking. We would be wrong if personalized political ads turn out not to be timed to emotional states at all, or if voters who notice the targeting still weigh the same information as everyone else.

We also read that people judge a message against what they already believe, and a voter deeply invested in one side may hear even accurate information from the other side as hostile. We would be wrong if voters with strong prior views judge identical messages the same way as voters with weak prior views.

What the reader can do with this: when a political message arrives at a low or heated moment, ask what in your mood might be doing the deciding, and notice whether the same message would move you on an ordinary day. Before judging a claim, ask how far it sits from what you already believe, and try to restate it in its own terms before accepting or rejecting it. When a claim comes from someone you respect, ask what evidence sits behind it, and look for the same claim from a source with no stake in the outcome.

### WHAT IS NOT KNOWN

Not known: whether any such tool keeps working, since the approach resembles a cat-and-mouse game in which profiling AI systems may adapt and change their algorithms after learning about the existence of AML⁷; whether it would work outside Indonesia's 2024 election or outside Twitter, which the authors name as a limit; whether the design principles hold, since the authors call for them to be "further evaluated and validated"; and whether typos survive contact with a system that reads words in fragments rather than whole, a doubt participants with the relevant expertise raised⁶.

Not known from the wider record either: how well people actually tell synthetic images from real ones in the settings they live in, since one experiment found people were, on average, not substantially better than chance at discerning AI-generated images from authentic images, regardless of the context of the image⁸; and how such content performs once posted, since one study of the 2024 European election campaign found AI-generated posts consistently achieve higher user engagement, particularly in reposts and saved posts⁹. Neither of those was measured in an election in the Americas, and neither is about the tool this study proposed.

The record is silent on the thing the reader most wants to know: whether any of this is running against them right now, in their own register, under their own authority's rules. This study cannot tell you that. It can only tell you that twenty people in one country, talking about one platform, wanted to see how they were being profiled — and disagreed about whether blurring their own words was worth the cost.

So here is the one thing to carry to the polling place. Before you share, like or forward anything political in the weeks before a fixed date, ask yourself one question out loud: if this had been built for me alone, what would it be trying to make me feel — and would I still believe it tomorrow?

Where each piece of context comes from, and how much of it we read

Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "The proliferation and pervasive use of artificial intelligence (AI) pose significant challenges to our democracies. In particular, AI leverages microtargeting political campaigns by constructing detailed user profiles and inferring people’s individual susceptibilities from their data."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "AML may also be used to counterattack targeted micro-targeted campaigns based on AI-driven political profilers by using AML to obscure social media data, thereby impairing the accuracy of AI systems in profiling users’ political preferences (Auliya et al., 2024 )"
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "AML, which uses an AI to create small non-random perturbations to change the output of another AI’s system (Szegedy et al., 2014 ), offers a promising avenue for mitigating the risk associated with some AI applications (P.-Y. Chen, 2022 )."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "All participants were excited about Feature-1 , which allowed them to be informed on how AI inferred their political preferences."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "However, Feature-2 , which introduced typos to blur social media data, received mixed reactions."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "Additionally, participants with expertise in Natural Language Processing (NLP) also doubted the effectiveness of typos in ‘fooling’ AI."
Auliya, S. F., Kudina, O., Ding, A. Y. et al. (2026). The Complexities of AI for Social Good: An Explorative Study on Adversarial Machine Learning to Facilitate Deliberative Decision-Making in Elections. Science and Engineering Ethics. https://doi.org/10.1007/s11948-026-00598-9 - the article this story is about — the whole article — the passage: "Third , the key limitation of AML lies in the evolution of countermeasured AI systems. As noted in prior conceptual work preceding this study by Auliya et al. ( 2024 ), the proposal of using AML to counterattack microtargeting political campaigns based on AI resembles a cat-and-mouse game, in which profiling AI systems may adapt and change their algorithms after learning about the existence of AML."
Hove MF, Hjorth F, Golovchenko Y, Adler-Nissen R, Bechmann A, van Dalen A, et al. (2026). Seeing Is Believing? Most People Struggle to Detect AI-Generated Images. Social Media + Society. https://doi.org/10.1177/20563051261483376 — only the abstract - the full text could not be fetched — the passage: "In a preregistered survey experiment, we find that, on average, people are not substantially better than chance at discerning AI-generated images from authentic images, regardless of the context of the image."
Pérez-Curiel C, López-Domínguez C, Méndez-Muros S, Mila-Maldonado A. (2026). When populism meets artificial intelligence: AI-generated multimedia content and far-right digital campaigning on Instagram and X in the 2024 European elections. Social Sciences & Humanities Open. https://doi.org/10.1016/j.ssaho.2026.103611 — only the abstract - the full text could not be fetched — the passage: "Moreover, AI-generated posts consistently achieve higher user engagement, particularly in reposts and saved posts."

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.