{"ok":true,"nota":{"id":"373a1eee37fa1d4a","slug":"un-estudio-sostiene-que-las-ia-que-dan-la-razon-al-usuario-nacen-en-el-373a1e","titulo":"Un estudio sostiene que las IA que dan la razón al usuario nacen en el trabajo mal pago de quienes las entrena","resumen":"Un ensayo académico propone que la tendencia de las inteligencias artificiales a darle la razón al usuario se origina en las condiciones laborales de quienes califican sus respuestas. Según el autor, presiones como cuotas de producción y el miedo a no cobrar llevan a los anotadores a evitar desacuerdos.","fuente":"zenodo","fuenteNombre":"Zenodo · Investigación abierta","color":"#1682d4","autor":"Jefferson, Josie, Velasco, Felix","url":"https://doi.org/10.5281/zenodo.21797921","imagen":"/hoy/portada/373a1eee37fa1d4a","tema":"tecnologia","subtemas":["inteligencia artificial","rlhf","condiciones laborales","anotación de datos","sesgos"],"entidades":["RLHF","anotadores humanos"],"formato":"dato","entretenimiento":false,"pubTs":"2026-10-28T07:30:14.000Z","indexTs":"2026-10-30T01:00:41.418Z","hash":"d563d94f346ab7e0f3a2f0a85e20097175810bd2008be5a252c603b9a7c46271","tsa":{"time":"Oct 30 01:01:14 2026 GMT","provider":"freetsa.org","sobre":"sello"},"stats":{"views":0,"likes":0,"clicks":0,"dwell":0},"tipo":"motor","registro":"hallazgo","porQueImporta":"Si el problema nace en la forma en que se pagan y se exigen esas calificaciones, la solución no es solo técnica sino también laboral. Para alguien de acá, esto significa que la herramienta que consulta para tomar decisiones puede estar reflejando la presión sobre trabajadores invisibles, poco pagos y apurados. Cuando un programa siempre te dice que tenés razón, conviene preguntarse quién tuvo que callar algo para que lo diga.","quienFirma":"Zenodo (autor independiente; ensayo no revisado por pares)","confianza":"preprint","consulta":{"pregunta":"¿Te molesta que la IA te dé siempre la razón?"}},"sello":{"hash":"d563d94f346ab7e0f3a2f0a85e20097175810bd2008be5a252c603b9a7c46271","sello":"37e2c6f53da41bd99518551ea61089808cc5963bc7677f01c6d92a4f9df0b94c","beacon":{"drand":{"cadena":"8990e7a9aaed2ffed73dbd7092123d6f289930540d7651336225dc172e51b2ce","ronda":6596367,"aleatorio":"ef53c0062595523fd47b36ffd8982ee8a6a5f7396a2cc7a1b06c6d15feaa692b","tiempo":"2026-10-30T01:00:30.000Z"},"nist":{"cadena":2,"pulso":2005117,"salida":"963EAEAD585239CD426CC00D24842A6D88AE3CEB2B3448F50146C6BA6715F36B18930186828A49FF12167BEB5E61D4AA75A13EFDC39077C8A3534C02D5838AA9","tiempo":"2026-10-30T00:59:00.000Z"},"noAntes":"2026-10-30T01:00:30.000Z"},"verificacion":{"drand":true,"nist":true},"tsa":{"time":"Oct 30 01:01:14 2026 GMT","provider":"freetsa.org","sobre":"sello"},"original":{"titulo":"The Wage of Agreement: Machine Sycophancy as a RLHF Labor Conditions Artifact","descripcion":"<div><strong> Abstract:</strong></div> <div> </div> <div>Scholarship frames machine sycophancy as an algorithmic failure. The standard account attempts to fix the signal after collection, ignoring the moment of manufacture. This paper argues machine sycophancy has a labor origin. Preference data is forged by humans under designed, paid conditions. Throughput quotas penalize friction. Agreement filters screen out dissent. The threat of non-payment forces annotators to suppress independent judgment. These pressures compel workers to act as sycophants to secure a wage. A model trained on these judgments inherits the disposition at its source. This analysis details the documented conditions of preference annotation. Machine sycophancy is built into the wage structure before model training begins. Because developers elect to mandate consensus or preserve disagreement, the resulting bias is a ","pubDate":"2026-10-28T07:30:14Z","url":"https://doi.org/10.5281/zenodo.21797921"},"indexado":"2026-10-30T01:00:41.418Z"}}