survey · PNAS Nexus · la publicación, 30 sep 2026 · gratis
Ensayos con residentes del Reino Unido encontraron que investigar temas políticos con un chatbot aumentó la creencia en información verdadera y redujo la creencia en información falsa, en la misma medida que buscar por Google.
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
- adultos británicos con derecho a voto y residentes del Reino Unido
- Cuántos
- 2,499 en la encuesta; 1,147 y 1,711 en los ensayos; 2,858 en total en los ensayos
- Dónde
- Reino Unido
- Cuándo
- en los cuatro días posteriores a la elección general del 4 de julio de 2024 y en la semana previa a esa elección
- Tipo de estudio
- encuesta y experimentos
- Quién lo hizo
- UK AI Security Institute, AI Policy Directorate, Collective Intelligence Project y una investigadora de Anthropic
- El límite que importa
- Un país, un ciclo electoral, cuatro temas y sesiones cortas: no se ha demostrado que se traslade a otros países.
Porcentajes de la encuesta nacional representativa del público del Reino Unido (N = 2,499) en la semana previa a la elección de 2024; no dicen nada sobre otros países ni sobre otras elecciones.
Investigar temas políticos con un chatbot frente a buscar en Google
el cambio en lo que la gente creía fue casi igual en los dos casos
investigar aumentó la creencia en lo verdadero y bajó la creencia en lo falso
el chatbot redujo el tiempo de búsqueda entre un 6% y un 10%: 17,94 minutos contra 19,82 minutos en promedio
Un estudio hecho en el Reino Unido no dice cómo se comportarán los chatbots en su país ni en su elección.
En los cuatro días siguientes a las elecciones del 4 de julio de 2024, en el Reino Unido, 2,499 votantes respondieron una encuesta. De los que usaban chatbots, 32 por ciento buscó información sobre temas políticos, y 13 por ciento de los votantes elegibles lo hizo. En un primer ensayo, 1,147 residentes fueron asignados al azar a investigar dos de cuatro temas con un chatbot o con Google. En un segundo ensayo, 1,711 residentes usaron un chatbot con distintas instrucciones. Quienes investigaron creyeron más la información verdadera y menos la falsa, y el cambio fue casi igual si usaban chatbot o Google. El estudio mismo advierte que no cubre temas nuevos o poco documentados, ni búsquedas personalizadas, ni chatbots más recientes.
Este documento no menciona ninguna página de su autoridad electoral, ningún registro de anuncios y ninguna herramienta de verificación. No ofrece ningún modo de comprobar un mensaje o un anuncio concreto.
El documento no cita ninguna resolución, auditoría ni fallo de autoridad electoral. No hay regla que pueda citarle aquí.
Qué significa para usted
Estos ensayos fueron con residentes del Reino Unido, así que no prueban nada sobre su país, su autoridad electoral ni la próxima votación. Si usted investiga un tema político con un chatbot, puede esperar creer más lo verdadero y menos lo falso, igual que si buscara en Google. Ningún estudio todavía mide si eso cambia su voto o el conteo.
Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305
Quién pagó: El artículo no indica quién financió el estudio; enumera afiliaciones con el UK AI Security Institute, la AI Policy Directorate, el Collective Intelligence Project y Anthropic, y declara el empleo de Saffron Huang en Anthropic como conflicto de interés.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 922 palabras · unos 5 minLeerla →Cerrar
La regla y la evidencia: la conversación con IA aumentó el conocimiento político en la misma medida que la búsqueda en Google
### EL ACTO
Bajo el título "Conversational AI increases political knowledge as effectively as self-directed internet search", un equipo del UK AI Security Institute y del AI Policy Directorate de Londres, junto al Collective Intelligence Project y a una investigadora de Anthropic, publicó en arXiv un estudio que no dicta regla, no impone sanción y no obliga a nadie: su único texto vinculante para el lector es la medición que reporta1.
El diseño es un ensayo aleatorizado. Cada participante fue asignado al azar a investigar dos de cuatro temas —cambio climático, inmigración, justicia penal y política de COVID-19— usando un chatbot conversacional o la búsqueda de Google, mientras los otros dos temas quedaban como control dentro de la misma persona; las creencias se midieron antes y después2.
Lo que el acto deja en firme es una equivalencia, no una ventaja: la conversación dirigida a una tarea aumentó la creencia en información verdadera y redujo la creencia en información falsa, y lo hizo en la misma medida que la búsqueda autodirigida en Google2. No hay autoridad que lo haga cumplir, no hay fecha de entrada en vigor, y ningún votante queda obligado a nada por este documento2.
### EL CASO QUE RESPONDE
No se encontró esta semana ningún caso real, con nombre, lugar, fecha y segunda fuente, que este acto venga a responder.
### LO QUE SE MIDIÓ
La encuesta se hizo a 2,499 adultos británicos con derecho a voto en los cuatro días inmediatamente posteriores a la elección general del Reino Unido del 4 de julio de 2024, y en la semana previa a esa elección hasta un 32% de quienes usaban chatbots —y un 13% de los votantes habilitados— los usó para buscar información política relevante para su decisión electoral1.
El ensayo principal reunió a 1,147 residentes del Reino Unido y un segundo ensayo, con la misma estructura, a 1,711; en total, 2,858 personas en los ensayos aleatorizados2. Investigar los temas aumentó la creencia en lo verdadero y bajó la creencia en lo falso, con un parámetro de .26 y un intervalo de alta densidad posterior de [.19; .32]3.
La diferencia entre investigar con chatbot y buscar en Google fue casi nula: .02, con intervalo [–.09; .12]4. El patrón se sostuvo incluso cuando se instruyó a los chatbots a ser persuasivos o serviles, y los controles de temas no investigados muestran que se trata de adquisición de conocimiento y no de simple participación5.
En tiempo, la conversación con IA redujo la búsqueda de información entre un 6% y un 10% frente al buscador: unos 17,94 minutos contra 19,82 minutos en promedio6. La encuesta se hizo en los días posteriores a la elección general del Reino Unido y los ensayos fueron experimentos en línea con residentes del Reino Unido2.
### LA RESPUESTA DE LA OTRA REGIÓN A LO MISMO
No encontramos esta semana ningún tribunal, ley o autoridad de otra región que se haya pronunciado sobre este mismo método.
### LA LECTURA DE LA BIBLIOTECA
Así lo leemos nosotros. Un rumor sobre un candidato siempre ha trabajado con lo que ya estaba roto en una sociedad: encuentra la grieta y la ensancha. Lo nuevo no es el engaño; es la velocidad con que una respuesta llega redactada, con tono seguro y sin que nadie firme por ella.
Lo que un público alcanza a verificar es lo que puede volver a encontrar por su cuenta; lo que se le queda en el aire es el origen de un dato que solo aparece una vez, bien dicho, sin responsable a la vista. Por eso conviene distinguir entre saber más datos y entender un problema: lo primero se consigue en una tarde de consultas, lo segundo exige poder explicarlo con palabras propias.
El estudio se limita a un país, un ciclo electoral, cuatro temas y sesiones cortas con una muestra pequeña de modelos, de modo que no se ha demostrado que un resultado de este tipo se traslade a otros países, otras poblaciones u otros sistemas. Es sobre llevarlo de un lado a otro, no sobre lo que aquí se midió.
### LO QUE NO SE SABE
El estudio no captura los efectos de equilibrio de la IA sobre el conocimiento o la participación política en general7; se limita a un país, un ciclo electoral, cuatro temas y sesiones cortas con una muestra pequeña de modelos8; no incluyó modelos señalados como muy sesgados hacia un extremo del espectro político, de modo que los efectos de modelos más sesgados podrían estar subestimados9.
Tampoco puede generalizarse a temas nicho o emergentes donde la información confiable escasea y fuentes de baja calidad dominan los resultados, ni a entornos de búsqueda personalizada donde la curación algorítmica puede dirigir al usuario hacia información de menor calidad8. Las búsquedas de Google se hicieron en una ventana limpia, sin historial ni cookies, lo que controla la personalización algorítmica que en el mundo real puede agravar las burbujas de filtro y los vacíos de datos10.
Un trabajo experimental previo mostró que rankings de buscador sesgados pueden desplazar las preferencias de voto de votantes indecisos en un 20% o más, en gran medida sin que lo adviertan11. El estudio no midió actitudes políticas posteriores, comportamiento de voto ni compromiso cívico, y sus conclusiones no deben leerse como que la IA sea segura o beneficiosa en conjunto7.
Aquí termina lo que el documento registra y empieza lo que usted puede hacer con ello.
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "First, in a representative national survey of the UK public (N = 2,499), we find that in the week before the 2024 election as many as 32% of chatbot users – and 13% of eligible UK voters – have used conversational AI to seek political information relevant to their electoral choice."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "we find that across issues, models, and prompting strategies, task-directed conversations with AI to research specific political topics increase political knowledge (increase belief in true information and decrease belief in misinformation) to the same extent as self-directed Google search."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Researching (vs not researching) political issues increased belief in true information and decreased belief in misinformation across time points (Fig. 2B, solid vs dotted lines; (βPOST×TRUE×RESEARCHED = .26, 95%-HPDI [.19; .32]"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Importantly, we found that belief change was nearly identical for participants who researched using conversational AI or Google search (Fig. 2B, left vs right panel), reflecting in a close to zero parameter estimate (βPOST×TRUE×RESEARCHED×CONVAI = .02, 95%-HPDI [–.09; .12]"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "even when the chatbots were prompted to be persuasive or sycophantic. Our within-subject controls for non-researched topics and secondary outcomes (trust, private beliefs, extremity) further demonstrate that these effects represent genuine knowledge acquisition rather than generic engagement or politically-biased persuasion."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "We found that the use of conversational AI reduced the information procurement time by 6-10% in comparison to self-guided Google search (average time spent on both research tasks (± standard deviation) in minutes: 17.94 (±8.64) for conversational AI vs 19.82 (±8.91) for Google search"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Our experimental design examines task-directed information-seeking, where participants were instructed to research specific political topics. Real-world AI use is considerably more varied: people may use chatbots to seek validation for pre-existing views, to discuss politics in unstructured ways, or conversely, AI may redirect users away from political engagement entirely"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "our findings cannot be assumed to generalise to more niche or emerging political topics where reliable information may be scarce and low-quality sources may dominate search results, nor to personalised search environments where algorithmic curation may direct users toward lower-quality information."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This sample of models did not include models that have been shown to be highly controversial and biased towards specific ends of the political spectrum. In this sense, our results might underestimate the belief changing effects that more biased models may have."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "participants in our study conducted their Google searches through an embedded browser window within our experiment app, which used a clean session without access to participants’ personal browsing history, cookies, or prior search data."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Epstein and Robertson [2015] showed in an experimental setupthat biased search engine rankings can shift voting preferences of undecided voters by 20% or more, largely without their awareness, an effect they termedthe Search Engine Manipulation Effect (SEME)."
Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305
Quién pagó: El artículo no indica quién financió el estudio; enumera afiliaciones con el UK AI Security Institute, la AI Policy Directorate, el Collective Intelligence Project y Anthropic, y declara el empleo de Saffron Huang en Anthropic como conflicto de interés.
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.
survey · PNAS Nexus · the paper, 30 Sep 2026 · free
AI chatbots and Google gave UK voters the same factual gains in a 2024 trial; the test covered four well-known UK topics only.
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
- UK adults eligible to vote, and UK residents in trials
- How many
- 2,499 in the survey; 1,147 in the first trial; 1,711 in the second; 2,858 across both trials
- Where
- United Kingdom
- When
- four days after the UK general election of 4 July 2024
- Kind of study
- survey and experiment
- Who did it
- UK AI Security Institute, London
- The limit that matters
- Only four well-known UK topics, one country, one election cycle, clean search session
Researching political topics with a chatbot compared with Google search
Belief change was nearly identical; the difference between the two was close to zero
Chatbot use cut information-gathering time by 6-10% (17.94 minutes against 19.82 minutes across two research tasks)
The paper names no way for a voter to check a specific video, ad or message before a vote.
In the days after the UK general election of July 4, 2024, researchers surveyed 2,499 eligible UK voters. Nine percent had used chatbots for political information, and among chatbot users 32 percent had used them for current affairs. In trials with 1,147 UK residents, people researched two of four topics (climate, immigration, criminal justice, COVID-19) using a chatbot or Google search. Both groups increased belief in true information by about the same amount. People agreed with true statements about one point more than false ones on a seven-point scale.
The paper names no way for a voter to check a specific video, ad or message before a vote.
The paper states no electoral authority rule, audit or ruling. It reports only the trial result: chatbot and Google search produced nearly the same change in belief, with the conversational AI result faster by 6 to 10 percent on time spent.
What this means for you
What this means for you, as a voter in Latin America or North America, is narrower than it sounds: the trial ran only in the UK, on four well-known topics, with no ballots involved. Nothing yet tells you whether the chatbot or the search results you meet in your own feed are steering you. Watch for the audit or observation report from your electoral authority, and read it before you decide whom to believe.
Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305
Who paid: The article does not state who funded the study; it lists affiliations with the UK AI Security Institute, the AI Policy Directorate, the Collective Intelligence Project, and Anthropic, and declares Saffron Huang's employment at Anthropic as a conflict of interest.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1172 words · about 6 minRead it →Close
The rule and the evidence: AI chatbots and Google gave UK voters the same factual gains in a 2024 trial; the test covered four well-known UK topics only
### THE ACT
What is in force here is not a law or a ruling but a study, published on arXiv on 25 June 2026 as arXiv:2509.05219v5 [cs.HC]. It carries no enforcement, no deadline and no authority over anyone. There is no resolution to quote, no bill, no order, and nothing a person is required to do by a given date.
The only "act" the document records is what people did on their own: "in the week before the 2024 election as many as 32% of chatbot users – and 13% of eligible UK voters – have used conversational AI to seek political information relevant to their electoral choice"1. That is a description of behaviour, not a rule.
The survey behind that figure ran in the four days immediately after the UK general election of 4 July 2024, and the trials that followed measured what happened to beliefs, not what anyone was permitted or forbidden to do.
### THE CASE IT ANSWERS
No real instance was found this week.
### WHAT WAS MEASURED
The trials measured belief, not knowledge of a scoreboard. Participants rated their agreement with statements — some true, some false — on a seven-point scale before and after research, and the change was the outcome23.
The headline result: researching political issues raised agreement with true statements and lowered agreement with false ones, with an effect of βPOST×TRUE×RESEARCHED = .26, 95%-HPDI [.19; .32]3. Picture a scale running from one to seven: that is a shift of roughly a quarter of a point on average, not a conversion.
The comparison that matters to the reader: the difference between chatbot users and Google-search users was close to zero, βPOST×TRUE×RESEARCHED×CONVAI = .02, 95%-HPDI [–.09; .12]4. A range that includes zero means the two groups learned about the same amount.
The sample: 2,499 UK adults eligible to vote in the survey, 1,147 in the first trial, 1,711 in the second, 2,858 across both trials. The country is the United Kingdom, and the moment is one election cycle.
It has been measured in an election year, in the sense that the survey ran days after a national vote, but the belief experiments were not the election itself. They were randomised sessions in which people were told which topics to research5.
One further measurement: chatbot users spent less time getting the information, a reduction of 6-10% against Google search — about 17.94 minutes against 19.82 minutes across two research tasks6. The same knowledge, slightly faster.
### THE OTHER REGION'S ANSWER TO THE SAME THING
We found none this week.
### THE LIBRARY'S READING
Here is how we read it. A false rumour about a candidate has never needed to be believed by everyone. It has needed only to arrive early, in a form that feels like it came from somewhere, and to leave a residue that is hard to wash off. What is new is not the rumour. What is new is that the rumour now answers back, in your own language, at the hour you ask, and that it does so without a masthead, a byline or a page you could point to later.
We also read, elsewhere, that concerns have been raised about chatbots producing unreliable or biased answers on political issues, and that models may lean in a progressive or libertarian direction. We say this plainly: those findings come from other places, on other systems, and they tell us about carrying over, not about the result in front of us. They are a reason to watch, not a reason to doubt what was measured.
What a public can learn to check together is the source, the independence of the source, and who answers for it. What no voter can check alone is whether the answer they were given came from a page that exists, a page someone stands behind, or a sentence assembled to sound like one. That is not a failing of the reader. It is a gap in what the reader is handed.
So what should a household expect? Expect that the smooth answer will feel more settled than it is, and that the feeling of directness will outrun the checking. You would know we are wrong if the people around you name where their election answers came from as readily as they name who told them, and if they ask for a second source as a habit rather than a correction.
And expect that a new place to ask does not simply sit beside the old ones. When a chatbot becomes the first place a voter goes, something else quietly stops being the first place they go. You can test that in your own house tonight, without any study: ask what you used to reach for before you reached for this.
### WHAT IS NOT KNOWN
The study cannot say what happens on niche or emerging political topics where reliable information is scarce and low-quality sources may dominate, nor in personalised search environments where what you are shown depends on what you have looked at before7. The searches in the experiment ran in a clean session, without the participant's browsing history or cookies8.
It cannot say what happens with models known to lean hard toward one end of the political spectrum; those were not in the sample, and the authors say the effects of such models may be underestimated9.
It cannot say what happens outside the United Kingdom, or beyond one election cycle, or among people other than UK residents5.
It cannot say whether the belief changes last, whether they touch how anyone votes, or what they do to trust in institutions over time. The study measured a single session.
It cannot say whether the prompting that was meant to make a chatbot sycophantic or persuasive changed the model's behaviour as strongly as intended — the authors flag that the participants' own ratings of the models did not differ between the prompted and unprompted conditions, so those null results deserve caution.
And it cannot say what the wider effect of these tools on political life will be. The authors write that their findings speak to the direct effects of using AI as a research tool, and should not be read as the full picture of what AI availability does to political knowledge or engagement5.
One more thing we hold in view: a search engine's rankings can move undecided voters by a fifth or more without their knowing, as an experiment on biased rankings found. That was not this study, and it was not these tools. It is simply the reminder that the order in which answers arrive is itself a choice someone made.
The next time a chatbot tells you something about a candidate, a date or a rule before you vote, ask it one thing it cannot answer smoothly: where did you get that, and can you show me the page?
Where each piece of context comes from, and how much of it we read
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "First, in a representative national survey of the UK public (N = 2,499), we find that in the week before the 2024 election as many as 32% of chatbot users – and 13% of eligible UK voters – have used conversational AI to seek political information relevant to their electoral choice."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "we find that across issues, models, and prompting strategies, task-directed conversations with AI to research specific political topics increase political knowledge (increase belief in true information and decrease belief in misinformation) to the same extent as self-directed Google search."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "Researching (vs not researching) political issues increased belief in true information and decreased belief in misinformation across time points (Fig. 2B, solid vs dotted lines; (βPOST×TRUE×RESEARCHED = .26, 95%-HPDI [.19; .32]"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "Importantly, we found that belief change was nearly identical for participants who researched using conversational AI or Google search (Fig. 2B, left vs right panel), reflecting in a close to zero parameter estimate (βPOST×TRUE×RESEARCHED×CONVAI = .02, 95%-HPDI [–.09; .12]"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "Our experimental design examines task-directed information-seeking, where participants were instructed to research specific political topics. Real-world AI use is considerably more varied: people may use chatbots to seek validation for pre-existing views, to discuss politics in unstructured ways, or conversely, AI may redirect users away from political engagement entirely"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "We found that the use of conversational AI reduced the information procurement time by 6-10% in comparison to self-guided Google search (average time spent on both research tasks (± standard deviation) in minutes: 17.94 (±8.64) for conversational AI vs 19.82 (±8.91) for Google search"
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "our findings cannot be assumed to generalise to more niche or emerging political topics where reliable information may be scarce and low-quality sources may dominate search results, nor to personalised search environments where algorithmic curation may direct users toward lower-quality information."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "participants in our study conducted their Google searches through an embedded browser window within our experiment app, which used a clean session without access to participants’ personal browsing history, cookies, or prior search data."
- Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305 - the article this story is about — the whole article — the passage: "This sample of models did not include models that have been shown to be highly controversial and biased towards specific ends of the political spectrum. In this sense, our results might underestimate the belief changing effects that more biased models may have."
Luettgau, L., Kirk, H. R., Hackenburg, K. et al. (2026). Conversational AI increases political knowledge as effectively as self-directed internet search. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgag305
Who paid: The article does not state who funded the study; it lists affiliations with the UK AI Security Institute, the AI Policy Directorate, the Collective Intelligence Project, and Anthropic, and declares Saffron Huang's employment at Anthropic as a conflict of interest.
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.