interviews · Science and engineering ethics · la publicación, 30 abr 2026 · gratis
Votantes de Indonesia vieron en un video una herramienta que no existe: solo un estudio de entrevistas
Versión breve · la versión detallada sigue, unos 7 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
- Votantes y actores políticos de Indonesia
- Cuántos
- veinte ciudadanos y cuatro actores políticos
- Dónde
- Indonesia
- Cuándo
- elecciones de 2024
- Tipo de estudio
- interviews
- Quién lo hizo
- Universidad Tecnológica de Delft, Países Bajos
- El límite que importa
- Nunca se construyó ni se probó: solo vieron un video de dos minutos veinte personas.
La herramienta nunca se construyó ni se probó.
Investigadores de la Universidad de Delft, en Países Bajos, entrevistaron a veinte votantes indonesios y a cuatro actores políticos durante las elecciones de 2024 en Indonesia. Les mostraron un video de dos minutos sobre una herramienta propuesta, llamada APP-AI, que avisaría a la persona cómo la inteligencia artificial adivina sus preferencias políticas y agregaría errores de escritura a sus publicaciones para confundir a quien la perfila. La herramienta nunca se construyó ni se probó. Algunos entrevistados dudaron de que los errores de escritura engañen a los sistemas actuales, y otros temieron perder credibilidad ante los demás. Los autores concluyen que la privacidad por sí sola no basta para que alguien vote con reflexión.
El documento no nombra ninguna forma para que un votante revise esto por su cuenta. Cita el sitio Bijak Memilih como ejemplo de una herramienta que muestra a la persona sus propias inclinaciones. Ese sitio ofrece dos cuestionarios: uno señala los temas que más le importan al usuario según sus respuestas, y el otro compara sus posturas con lo que han hecho los partidos. Aparte, para fomentar el pensamiento crítico, propone herramientas que invitan a verificar los datos y avisos cuando alguien solo ve información falsa.
El estudio no menciona ninguna norma, resolución ni auditoría de una autoridad electoral. Solo cita la ley indonesia de protección de datos (UU 27/2022) y una regulación de campaña (UU 7/2017), sin fecha de aplicación ni hallazgo de auditoría.
Qué significa para usted
Ninguna autoridad electoral de Indonesia aparece en este estudio, así que no hay norma, resolución ni auditoría que revisar sobre la herramienta del video. Lo que sí puede hacer es mirar sus propias inclinaciones antes de votar, como propone el sitio que el documento cita. Manténgase atento a los estudios que midan si estas ideas cambian algo en la práctica.
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 Indonesio de Dotación para la Educación (LPDP) financió a la primera autora (subvención KEP581/LPDP/LPDP.3/2022); el artículo no menciona otro financiador ni dice nada sobre el papel de los financiadores ni sobre equipos o programas prestados.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1375 palabras · unos 7 minLeerla →Cerrar
La regla y la evidencia: votantes de Indonesia vieron en un video una herramienta que no existe: solo un estudio de entrevistas
### EL ACTO, CITADO Y ENLAZADO
En esta historia no hay una resolución, ni una sentencia, ni una auditoría, ni una ley nueva. Lo que hay es un estudio publicado en 2026 por cuatro investigadores de la Universidad Tecnológica de Delft, en los Países Bajos, y su propuesta de diseño: una herramienta llamada APP-AI, con dos funciones. La primera "informaba a los usuarios sobre sus preferencias políticas tal como las perfilaba una IA que había aprendido de ellos usando sus datos de redes sociales"1. La segunda "modificaba sus datos de redes sociales metiendo errores deliberados en sus futuras publicaciones, usando el concepto de aprendizaje automático adversario, donde alteraciones pequeñas a la entrada de datos cambian la salida de la IA"2.
No hay autoridad que la aplique ni fecha en que entre en vigor, porque nunca se construyó. A los veinte ciudadanos indonesios entrevistados se les mostró un video de dos minutos con la propuesta, y luego se les preguntó qué opinaban. Eso es todo lo que existe.
Lo que el estudio sí deja como petición a otros es una frase dirigida a los reguladores: "las personas cuyos datos se usan deberían ser informadas de manera explícita cuando sistemas automatizados infieren sus preferencias políticas para mensajes políticos. Esa información debería incluir la lógica general de los métodos de perfilado y los datos involucrados"3. Es una recomendación, no una obligación.
### EL CASO QUE RESPONDE
Esta semana no encontramos ningún caso real que contar con nombre, lugar, fecha y segunda fuente: ni un video falso, ni un avatar, ni una red de cuentas falsas, ni un acta alterada. Preferimos decirlo así, en una sola frase, antes que rellenar el hueco con algo que no verificamos.
Lo que el documento sí nombra es el escándalo de Cambridge Analytica, que "demostró cómo el mal uso de la IA podía representar riesgos para el sistema electoral"4. No es un caso que este estudio haya investigado: es el ejemplo que los autores usan para explicar de qué hablan. Parecería que es esa clase de cosas la que tienen en mente.
El mecanismo que describen es más modesto y más concreto. La IA "aprovecha las campañas políticas de microsegmentación construyendo perfiles detallados de usuarios e infiriendo las susceptibilidades individuales de las personas a partir de sus datos"5. Y eso "permite mensajes políticos muy dirigidos que pueden influir de manera sustancial en las decisiones de voto, y potencialmente socavar la capacidad de los ciudadanos de tomar decisiones deliberativas en las elecciones"6.
Un dato ayuda a ver el tamaño del asunto. Un trabajo citado en el artículo mostró que características demográficas simples servían para predecir el voto de grupos cada vez más pequeños de ciudadanos7, y que una fundación holandesa contratada por más de setenta partidos logró afinar la información agregada "de 1,200 a 35 individuos"8. Piense en eso: de un bloque del tamaño de un pueblo a un grupo del tamaño de un salón de clases.
La consecuencia de la consecuencia, hasta donde el documento permite: si un mensaje se puede afinar hasta ese punto, el vecino de al lado puede estar recibiendo otro distinto, y ninguno de los dos lo sabe. La conversación previa al voto, que es donde se compara y se discute, se adelgaza sin que nadie la vea adelgazar.
### LO QUE SE MIDIÓ
No se midió ningún efecto. El estudio no reporta tamaño de efecto, ni porcentaje, ni significancia estadística. Entrevistó a veinte ciudadanos con derecho a voto en las elecciones indonesias de 2024 y a cuatro actores políticos, en Indonesia, con Twitter como plataforma principal. Es un estudio cualitativo de entrevistas semiestructuradas, analizado con análisis temático reflexivo, sin grupo de comparación.
Lo que sí registró fueron reacciones. "Todos los participantes se entusiasmaron con la Función 1, que les permitía estar informados sobre cómo la IA infería sus preferencias políticas"9. Con la segunda función hubo división: "algunos participantes que al principio aceptaban el uso de errores para protegerse del perfilado político cambiaron de postura tras preguntas de seguimiento que les pedían imaginar a una persona con errores frecuentes en sus publicaciones"10.
Y hubo una objeción técnica que vale la pena poner en palabras llanas. Un participante con conocimientos de procesamiento de lenguaje explicó que los errores se saltan con facilidad, y propuso en su lugar la sátira, el humor, la paráfrasis y la negación, que los modelos actuales de IA todavía tienen dificultades para entender11.
La conclusión de los propios autores es una negativa: "nuestros hallazgos revelan una realidad más compleja: atender problemas sociales con tecnología es intrínsecamente difícil; ninguna solución sola puede servir de bala de plata"12. Nunca se ha medido en una elección real si estas herramientas cambian el perfilado o la deliberación; tampoco en un laboratorio, con estos datos.
### LA RESPUESTA DE OTRA REGIÓN AL MISMO ASUNTO
Esta semana no encontramos ninguna corte, ley ni autoridad de otro país que se haya pronunciado sobre este método en los documentos que revisamos.
### LA LECTURA DE LA REDACCIÓN
Así lo leemos nosotros. La propaganda política siempre ha trabajado con lo que ya está en la cabeza del votante: el rumor que confirma el miedo, la foto sacada de contexto, la frase que se repite hasta sonar comprobada. Lo nuevo no es el contenido sino el reparto. Antes el mismo cartel se pintaba para todo un barrio; ahora el mensaje puede llegar distinto a cada casa, por el canal donde llegan los mensajes de la familia.
Espere algo concreto en su propio teléfono: que el anuncio parezca escrito para usted, que mencione su colonia, su oficio o su enojo, y que lo que le ofrezca se sienta como un favor. Cuando eso pasa, la pregunta útil no es "¿será falso?", sino "¿qué le dieron a cambio y qué esperan de vuelta?". Anótelo ese mismo día: qué le ofrecieron y cómo se sintió después. Si en una semana no logra recordar ni una sola propuesta concreta y solo recuerda frases pegajosas, ahí tiene una señal.
Podríamos estar equivocados. Si los votantes tratados con anuncios hechos a su medida no sienten ninguna gratitud ni deuda con quien los manda, y su decisión no se mueve aunque el mensaje parezca un regalo, entonces esta lectura no sirve. También fallaría si la gente común consulta y compara tanta información detallada como los mensajes emotivos.
Lo que usted sí puede hacer es pequeño y verificable: antes de votar, busque una fuente que contradiga lo que más le gustó oír, y coméntelo con alguien de confianza que no piense igual. El estudio relata que dos participantes que al principio tenían opiniones políticas firmes impulsadas por influencias externas empezaron a reflexionar sobre perspectivas alternativas solo después del estímulo de personas de confianza de su red social. Pregúntese con quién hablaría esto y qué le diría esa persona.
### LO QUE NO SE SABE
No se sabe si la herramienta funcionaría: nunca se construyó, solo fue un video de dos minutos visto por veinte personas. No se sabe si los errores deliberados engañan a los sistemas que perfilan, y los propios autores advierten que la limitación principal del método "está en la evolución de los sistemas de IA contrarios"13: es un juego del gato y el ratón, y ninguna protección debe darse por permanente.
No se sabe si esto pasa en su país. El estudio es de Indonesia, con veinte ciudadanos y cuatro actores políticos, en una sola plataforma, y sus autores no pretenden generalizar. Tampoco se sabe qué permite o audita su propia autoridad electoral, porque el documento no habla de ninguna.
No se sabe cuánto pesa esto frente a lo demás que decide un voto: el precio de la tortilla, el trabajo, la familia. El documento no lo mide.
Y no se sabe si el entusiasmo que los participantes mostraron por ver su propio perfil se convertiría en desconfianza al verlo de verdad, con nombres y datos que quizá no reconocen como suyos.
La próxima vez que un mensaje político le hable al oído, pregúntese quién se lo mandó y qué sabe de usted; y si no puede averiguarlo, trátelo como lo que es: una voz que no puede identificar.
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: "Feature-1 informed users about their political preferences as profiled by AI that learned about them using their social media 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 - el artículo del que trata esta nota — el artículo completo — el pasaje: "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 output."
- 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: "Data subjects should be explicitly informed when automated systems infer their political preferences for political messaging. Such information should include the general logic of the profiling methods and the data 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: "A prominent example is the Cambridge Analytica scandal, which demonstrated how the misuse of AI could pose risks to the electoral system."
- 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: "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 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This capability enables highly targeted political messaging that can substantially influence voting decisions, potentially undermining citizens’ ability to make deliberative decisions in elections."
- 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: "In 2018, Hazenberg et al. showed that simple demographic characteristics could be used to predict the voting behaviors of increasingly smaller groups of citizens."
- 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: "This allowed them to refine the aggregated information from 1,200 to 35 individuals, enhancing the precision of targeted campaigns to better align with the characteristics of the intended smaller groups"
- 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, some participants who initially accepted the use of typos to protect themselves from political profiling changed their stance after follow-up questions asking them to imagine a person who had frequent typos in their social media posts."
- 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: "Instead of typos that were very easily bypassed, he suggested some alternatives to alter social media data, such as satire, humor, paraphrasing, and negation, which current AI models still struggle to understand."
- 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: "Instead, our findings reveal a more complex reality: addressing societal issues through technology is inherently challenging; no single solution can serve as a silver bullet."
- 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."
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 Indonesio de Dotación para la Educación (LPDP) financió a la primera autora (subvención KEP581/LPDP/LPDP.3/2022); el artículo no menciona otro financiador ni dice nada sobre el papel de los financiadores ni sobre equipos o programas prestados.
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 proposed tool against tailored political ads was only shown as a video, and the voters who saw it were not sure about it
Short version · the longer version follows, about 6 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 voters 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 researchers
- The limit that matters
- The tool was only shown as a two-minute video; nothing was built or tested.
The tool was never built.
Researchers at TU Delft interviewed twenty Indonesian voters and four political actors during Indonesia's 2024 election, mostly about Twitter. They showed the voters a two-minute video of a proposed tool called APP-AI. One part told people how AI had profiled their political preferences. The other put deliberate typos in their posts to confuse the profiling. Voters welcomed the first part. They split on the typos. Some worried that frequent typos would make other people see them as less credible, and one said current AI reads past typos easily. The interviews yielded design principles for deliberation tools. Nothing was built or tested.
The article does name ways voters check or manage their own feed and data: some typed different keywords to confuse Twitter's algorithm, and one used TikTok's refresh feature to reset the 'For You' feed. The article also describes the 'Bijak Memilih' quizzes and APP-AI's first feature, which tells users how AI profiled their political preferences.
The article describes no rule, ruling or audit by an electoral authority. It says Indonesia's data protection law lets people object to decisions based only on automated profiling, and that rules on political micro-targeting are not yet explicit.
What this means for you
For now this is a study of twenty Indonesian voters and four campaign staff, and it tested no tool and watched no count. What it offers you is a question to hold up to any site or app that guesses your politics: does it tell you what it inferred about you? Until a real audit or a rule from your own electoral authority says otherwise, watch what gets disclosed about political targeting where you vote.
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
Who paid: The Indonesia Endowment Fund for Education (LPDP) supported the first author (grant KEP581/LPDP/LPDP.3/2022); the article states no other funder and says nothing about funders' role or any equipment or software lent.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1155 words · about 6 minRead it →Close
The rule and the evidence: A proposed tool against tailored political ads was only shown as a video, and the voters who saw it were not sure about it
### THE ACT
There is no act to quote this week. What the record carries is a proposal, not a rule in force: a research team at Delft University of Technology set out to identify "the design principles of technological tools that could facilitate citizens’ deliberative decision-making in elections." The tool they put in front of people, called Adversarial Privacy Preserving AI, was a two-minute video. Nothing was enacted, nothing was scheduled, and no authority enforces anything here.
The proposal had two parts. "Feature-1 informed users about their political preferences as profiled by AI that learned about them using their social media data"1. "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 output"2. The method behind the second part: "AML, which uses an AI to create small non-random perturbations to change the output of another AI’s system"3. Read that as one machine nudging another machine's answer off course, on purpose, in a way the second machine does not notice.
What a person could do with it, had it been built: see the guess a profiler had made about their politics, and decide whether to spoil that guess by typing badly on purpose. Nobody in this study could do either. They watched a video and answered questions.
### THE CASE IT ANSWERS
No real instance was found this week.
### WHAT WAS MEASURED
The study interviewed twenty Indonesian citizens eligible to vote in the 2024 elections and four political actors: people running in the 2024 elections, their campaign managers, or political or media consultants at management level. The citizens watched the two-minute video and were asked what they thought. The analysis was qualitative, through reflexive thematic analysis. No numerical effect was reported, and none was tested.
What the interviewees said divides cleanly. "All participants were excited about Feature-1, which allowed them to be informed on how AI inferred their political preferences"4. The typo feature split them. "However, some participants who initially accepted the use of typos to protect themselves from political profiling changed their stance after follow-up questions asking them to imagine a person who had frequent typos in their social media posts"5. Consider what that question asked them to picture: a person who writes sloppily in public. They did not like how that person looked.
One participant with expertise in Natural Language Processing doubted the trick would work at all, and offered alternatives instead: "satire, humor, paraphrasing, and negation, which current AI models still struggle to understand"6. The authors' own conclusion is blunt. "Instead, our findings reveal a more complex reality: addressing societal issues through technology is inherently challenging; no single solution can serve as a silver bullet"7.
The authors also name the ceiling on the method itself: "the key limitation of AML lies in the evolution of countermeasured AI systems"8. A profiler that learns the trick can adapt to it.
For scale, the paper cites earlier work on how finely political sorting can go. "In 2018, Hazenberg et al. showed that simple demographic characteristics could be used to predict the voting behaviors of increasingly smaller groups of citizens"9. "This allowed them to refine the aggregated information from 1,200 to 35 individuals, enhancing the precision of targeted campaigns to better align with the characteristics of the intended smaller groups"10. Picture a hall of twelve hundred people. Picture the same hall sorted down to thirty-five.
### THE OTHER REGION'S ANSWER TO THE SAME THING
We found none this week. No passage in the record carries another country's court, law or authority addressing this method.
### THE LIBRARY'S READING
Here is how we read it. A false story about a candidate has always done one thing well: it leaves a mark that is hard to wash off, even after the story is corrected. What is new is not the lie. It is the fitting. A message can now be shaped to one person's worries before it arrives, so the voter never sees the seam where the shaping happened.
That leads us to expect something in homes like yours. The part of the tool people welcomed was the mirror, not the shield: being told what a machine had guessed about them. Expect the demand to land there, in disclosure, before it lands in scrambling. You would know we are wrong if the people around you shrug at being profiled and only want the ads to stop, and never ask how they were sorted.
We also read a limit worth carrying. People rarely drop a strong view because a stranger hands them a contrary fact; they often hold tighter. What has moved people is someone they trust, close by, saying the other thing out loud. Watch whether any tool here arrives alone or with a person attached, and expect the lonely version to fade.
What can you do with this? When a message fits your mood too neatly, treat the fit as a signal to slow down, and ask who chose what you are seeing and on what grounds. Then ask your electoral authority one plain question: must a campaign say when a message was matched to you, and on what data. That question is answerable. Whether the answer exists is the thing to find out.
### WHAT IS NOT KNOWN
The tool was never built. What the twenty citizens judged was a two-minute video, so nothing here shows the typo trick would fool a real profiler, or that a real profiler would fail to adapt.
The record is silent on enforcement. The authors recommend that "data subjects should be explicitly informed when automated systems infer their political preferences for political messaging," including "the general logic of the profiling methods and the data involved"11. They note Indonesia's data protection law already gives people a right to object to decisions based on automated profiling, and that political micro-targeting is not yet explicitly covered. That is a gap they describe, not a fix anyone made.
The study sits in one country and mostly one platform, and the authors say so: Twitter users are a minority of Indonesian social media users, which constrains the scope. Twenty people chosen for variety cannot stand in for a nation, let alone for yours.
The authors also report that the political actors they spoke with did not run large-scale personal-data operations, citing cost, and that targeting in Indonesia's 2024 race ran at the level of groups rather than individuals. Whether that holds where you vote is not something this paper can tell you.
On the day you stand in the booth, the useful question is not whether a machine has a file on you. It is whether anyone is required to tell you what is in it.
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: "Feature-1 informed users about their political preferences as profiled by AI that learned about them using their social media 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: "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 output."
- 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"
- 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, some participants who initially accepted the use of typos to protect themselves from political profiling changed their stance after follow-up questions asking them to imagine a person who had frequent typos in their social media posts."
- 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: "Instead of typos that were very easily bypassed, he suggested some alternatives to alter social media data, such as satire, humor, paraphrasing, and negation, which current AI models still struggle to understand."
- 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: "Instead, our findings reveal a more complex reality: addressing societal issues through technology is inherently challenging; no single solution can serve as a silver bullet."
- 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."
- 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: "In 2018, Hazenberg et al. showed that simple demographic characteristics could be used to predict the voting behaviors of increasingly smaller groups of citizens."
- 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: "This allowed them to refine the aggregated information from 1,200 to 35 individuals, enhancing the precision of targeted campaigns to better align with the characteristics of the intended smaller groups"
- 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: "Data subjects should be explicitly informed when automated systems infer their political preferences for political messaging. Such information should include the general logic of the profiling methods and the data 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
Who paid: The Indonesia Endowment Fund for Education (LPDP) supported the first author (grant KEP581/LPDP/LPDP.3/2022); the article states no other funder and says nothing about funders' role or any equipment or software lent.
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.
