{"id":11714,"date":"2025-09-12T15:29:51","date_gmt":"2025-09-12T18:29:51","guid":{"rendered":"https:\/\/fenati.org.br\/?p=11714"},"modified":"2025-09-15T09:35:50","modified_gmt":"2025-09-15T12:35:50","slug":"ias-falhas-moderacao-discurso-de-odio","status":"publish","type":"post","link":"https:\/\/homolog.fenati.org.br\/en\/ias-falhas-moderacao-discurso-de-odio\/","title":{"rendered":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio"},"content":{"rendered":"

Discurso de \u00f3dio<\/strong> – Com a expans\u00e3o do discurso de \u00f3dio nas redes sociais, empresas de tecnologia t\u00eam recorrido a modelos de intelig\u00eancia artificial<\/a> para filtrar automaticamente conte\u00fados nocivos. No entanto, um estudo publicado no ACL Anthology revela que esses sistemas ainda apresentam falhas significativas.<\/p>\n

Pesquisadores da Escola de Comunica\u00e7\u00e3o Annenberg realizaram a primeira avalia\u00e7\u00e3o comparativa em larga escala de sete modelos de IA usados na modera\u00e7\u00e3o, entre eles sistemas da OpenAI, Mistral, Claude 3.5 Sonnet, DeepSeek V3 e Google Perspective.<\/p>\n

LEIA: A\u00e7\u00e3o judicial contra a Uber acusa empresa de discrimina\u00e7\u00e3o contra PCDs<\/strong><\/a><\/p>\n

O experimento analisou 1,3 milh\u00e3o de frases sint\u00e9ticas envolvendo 125 grupos sociais, formuladas em diferentes contextos, desde termos neutros e positivos at\u00e9 insultos.<\/p>\n

Principais conclus\u00f5es do experimento<\/strong><\/p>\n

\u2022 Classifica\u00e7\u00f5es divergentes para o mesmo conte\u00fado: modelos distintos deram respostas opostas a textos id\u00eanticos, com alguns sinalizando-os como ofensivos e outros os aceitando, o que pode gerar a percep\u00e7\u00e3o de vi\u00e9s e comprometer a confian\u00e7a p\u00fablica.
\n\u2022 Sensibilidade desigual a certos grupos: enquanto alguns sistemas mant\u00eam maior consist\u00eancia, outros apresentaram varia\u00e7\u00f5es mais marcantes, sobretudo em frases relacionadas a escolaridade, interesses pessoais e classe econ\u00f4mica, expondo determinados p\u00fablicos a maior vulnerabilidade.
\n\u2022 Tratamento inconsistente de frases neutras e positivas: modelos como Claude 3.5 Sonnet e Mistral marcaram insultos como prejudiciais de forma categ\u00f3rica, independentemente do contexto, enquanto outros consideraram a inten\u00e7\u00e3o, revelando a aus\u00eancia de um padr\u00e3o intermedi\u00e1rio.<\/p>\n

Segundo os pesquisadores, essas discrep\u00e2ncias refor\u00e7am os dilemas da modera\u00e7\u00e3o automatizada: encontrar equil\u00edbrio entre precis\u00e3o e excesso de restri\u00e7\u00e3o. O estudo conclui que, apesar dos avan\u00e7os, as ferramentas de IA ainda t\u00eam limita\u00e7\u00f5es significativas para lidar de forma justa e confi\u00e1vel com o discurso de \u00f3dio online.<\/p>\n

(Com informa\u00e7\u00f5es de Olhar Digital)<\/strong><\/em>
\n(Foto: Reprodu\u00e7\u00e3o\/Freepik\/user8285578)<\/strong><\/em><\/p>","protected":false},"excerpt":{"rendered":"

An\u00e1lise comparativa mostra que principais modelos de linguagem divergem nas classifica\u00e7\u00f5es de um mesmo conte\u00fado<\/p>","protected":false},"author":10,"featured_media":11715,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[35],"tags":[13],"class_list":["post-11714","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ti","tag-sindical"],"yoast_head":"\nIAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio - FENATI<\/title>\n<meta name=\"robots\" content=\"noindex, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio - FENATI\" \/>\n<meta property=\"og:description\" content=\"An\u00e1lise comparativa mostra que principais modelos de linguagem divergem nas classifica\u00e7\u00f5es de um mesmo conte\u00fado\" \/>\n<meta property=\"og:url\" content=\"https:\/\/homolog.fenati.org.br\/en\/ias-falhas-moderacao-discurso-de-odio\/\" \/>\n<meta property=\"og:site_name\" content=\"FENATI\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/feittinf\" \/>\n<meta property=\"article:published_time\" content=\"2025-09-12T18:29:51+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2025-09-15T12:35:50+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2025\/09\/design-sem-nome-2025-09-12t152852169.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"1200\" \/>\n\t<meta property=\"og:image:height\" content=\"630\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"Julia Stoever\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@FederacaoTiBR\" \/>\n<meta name=\"twitter:site\" content=\"@FederacaoTiBR\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Julia Stoever\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"2 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/\"},\"author\":{\"name\":\"Julia Stoever\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#\\\/schema\\\/person\\\/5b0800dda1ad08bac67cec5eacddc965\"},\"headline\":\"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio\",\"datePublished\":\"2025-09-12T18:29:51+00:00\",\"dateModified\":\"2025-09-15T12:35:50+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/\"},\"wordCount\":331,\"publisher\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/homolog.fenati.org.br\\\/wp-content\\\/uploads\\\/2025\\\/09\\\/design-sem-nome-2025-09-12t152852169.jpg\",\"keywords\":[\"sindical\"],\"articleSection\":[\"TI\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/\",\"url\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/\",\"name\":\"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio - FENATI\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/homolog.fenati.org.br\\\/wp-content\\\/uploads\\\/2025\\\/09\\\/design-sem-nome-2025-09-12t152852169.jpg\",\"datePublished\":\"2025-09-12T18:29:51+00:00\",\"dateModified\":\"2025-09-15T12:35:50+00:00\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#primaryimage\",\"url\":\"https:\\\/\\\/homolog.fenati.org.br\\\/wp-content\\\/uploads\\\/2025\\\/09\\\/design-sem-nome-2025-09-12t152852169.jpg\",\"contentUrl\":\"https:\\\/\\\/homolog.fenati.org.br\\\/wp-content\\\/uploads\\\/2025\\\/09\\\/design-sem-nome-2025-09-12t152852169.jpg\",\"width\":1200,\"height\":630,\"caption\":\"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/ias-falhas-moderacao-discurso-de-odio\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"In\u00edcio\",\"item\":\"https:\\\/\\\/homolog.fenati.org.br\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#website\",\"url\":\"https:\\\/\\\/homolog.fenati.org.br\\\/\",\"name\":\"FEITTINF\",\"description\":\"Federa\u00e7\u00e3o Nacional dos Trabalhadores em Tecnologia da\u00a0Informa\u00e7\u00e3o\",\"publisher\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/homolog.fenati.org.br\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#organization\",\"name\":\"FEITTINF\",\"url\":\"https:\\\/\\\/homolog.fenati.org.br\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/homolog.fenati.org.br\\\/wp-content\\\/uploads\\\/2023\\\/08\\\/logo_300x50.png\",\"contentUrl\":\"https:\\\/\\\/homolog.fenati.org.br\\\/wp-content\\\/uploads\\\/2023\\\/08\\\/logo_300x50.png\",\"width\":350,\"height\":50,\"caption\":\"FEITTINF\"},\"image\":{\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/www.facebook.com\\\/feittinf\",\"https:\\\/\\\/x.com\\\/FederacaoTiBR\",\"https:\\\/\\\/www.instagram.com\\\/feittinf\\\/\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/homolog.fenati.org.br\\\/#\\\/schema\\\/person\\\/5b0800dda1ad08bac67cec5eacddc965\",\"name\":\"Julia Stoever\",\"url\":\"https:\\\/\\\/homolog.fenati.org.br\\\/en\\\/author\\\/julia-stoever\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio - FENATI","robots":{"index":"noindex","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"og_locale":"en_US","og_type":"article","og_title":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio - FENATI","og_description":"An\u00e1lise comparativa mostra que principais modelos de linguagem divergem nas classifica\u00e7\u00f5es de um mesmo conte\u00fado","og_url":"https:\/\/homolog.fenati.org.br\/en\/ias-falhas-moderacao-discurso-de-odio\/","og_site_name":"FENATI","article_publisher":"https:\/\/www.facebook.com\/feittinf","article_published_time":"2025-09-12T18:29:51+00:00","article_modified_time":"2025-09-15T12:35:50+00:00","og_image":[{"width":1200,"height":630,"url":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2025\/09\/design-sem-nome-2025-09-12t152852169.jpg","type":"image\/jpeg"}],"author":"Julia Stoever","twitter_card":"summary_large_image","twitter_creator":"@FederacaoTiBR","twitter_site":"@FederacaoTiBR","twitter_misc":{"Written by":"Julia Stoever","Est. reading time":"2 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#article","isPartOf":{"@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/"},"author":{"name":"Julia Stoever","@id":"https:\/\/homolog.fenati.org.br\/#\/schema\/person\/5b0800dda1ad08bac67cec5eacddc965"},"headline":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio","datePublished":"2025-09-12T18:29:51+00:00","dateModified":"2025-09-15T12:35:50+00:00","mainEntityOfPage":{"@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/"},"wordCount":331,"publisher":{"@id":"https:\/\/homolog.fenati.org.br\/#organization"},"image":{"@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#primaryimage"},"thumbnailUrl":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2025\/09\/design-sem-nome-2025-09-12t152852169.jpg","keywords":["sindical"],"articleSection":["TI"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/","url":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/","name":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio - FENATI","isPartOf":{"@id":"https:\/\/homolog.fenati.org.br\/#website"},"primaryImageOfPage":{"@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#primaryimage"},"image":{"@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#primaryimage"},"thumbnailUrl":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2025\/09\/design-sem-nome-2025-09-12t152852169.jpg","datePublished":"2025-09-12T18:29:51+00:00","dateModified":"2025-09-15T12:35:50+00:00","breadcrumb":{"@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#primaryimage","url":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2025\/09\/design-sem-nome-2025-09-12t152852169.jpg","contentUrl":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2025\/09\/design-sem-nome-2025-09-12t152852169.jpg","width":1200,"height":630,"caption":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio"},{"@type":"BreadcrumbList","@id":"https:\/\/homolog.fenati.org.br\/ias-falhas-moderacao-discurso-de-odio\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"In\u00edcio","item":"https:\/\/homolog.fenati.org.br\/"},{"@type":"ListItem","position":2,"name":"IAs apresentam falhas e inconsist\u00eancias na modera\u00e7\u00e3o de discurso de \u00f3dio"}]},{"@type":"WebSite","@id":"https:\/\/homolog.fenati.org.br\/#website","url":"https:\/\/homolog.fenati.org.br\/","name":"FEITTINF","description":"Federa\u00e7\u00e3o Nacional dos Trabalhadores em Tecnologia da\u00a0Informa\u00e7\u00e3o","publisher":{"@id":"https:\/\/homolog.fenati.org.br\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/homolog.fenati.org.br\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/homolog.fenati.org.br\/#organization","name":"FEITTINF","url":"https:\/\/homolog.fenati.org.br\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/homolog.fenati.org.br\/#\/schema\/logo\/image\/","url":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2023\/08\/logo_300x50.png","contentUrl":"https:\/\/homolog.fenati.org.br\/wp-content\/uploads\/2023\/08\/logo_300x50.png","width":350,"height":50,"caption":"FEITTINF"},"image":{"@id":"https:\/\/homolog.fenati.org.br\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/feittinf","https:\/\/x.com\/FederacaoTiBR","https:\/\/www.instagram.com\/feittinf\/"]},{"@type":"Person","@id":"https:\/\/homolog.fenati.org.br\/#\/schema\/person\/5b0800dda1ad08bac67cec5eacddc965","name":"Julia Stoever","url":"https:\/\/homolog.fenati.org.br\/en\/author\/julia-stoever\/"}]}},"_links":{"self":[{"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/posts\/11714","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/users\/10"}],"replies":[{"embeddable":true,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/comments?post=11714"}],"version-history":[{"count":1,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/posts\/11714\/revisions"}],"predecessor-version":[{"id":11716,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/posts\/11714\/revisions\/11716"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/media\/11715"}],"wp:attachment":[{"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/media?parent=11714"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/categories?post=11714"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/homolog.fenati.org.br\/en\/wp-json\/wp\/v2\/tags?post=11714"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}