# MIT Technology Review reports Facebook’s claims for its Blender chatbot [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-reports-facebook-s-claims-for-its-blender-4151db96) MIT Technology Review reported on April 29, 2020, that Facebook had open-sourced Blender, a chatbot designed for open-ended conversation. Facebook said the system could discuss many topics in an engaging way, and its tests compared Blender with Google’s Meena using human evaluators. The article describes Blender’s training on Reddit conversations, followed by fine-tuning on data intended to support empathy, knowledge and personality. Facebook’s evaluators rated Blender more engaging and more human-sounding than Meena, but the article notes that these results came from short conversations. The researchers expected longer chats to expose problems with coherence and memory. The article also describes a risk of Blender making up facts, because it generates responses from statistical correlations rather than a knowledge database. Facebook’s team planned to explore adding such a database. The article reports that harmful or biased responses were another concern: Reddit data was not filtered for harmful language, and a toxic-language classifier might miss how a response changes meaning in context. How well would a toxic-language classifier detect harmful meanings that depend on conversational context? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Apr 29, 2020 (2020-04-29). Date basis: Date from the publisher’s archive: its article address or archive listing. Added to Escape Meridian: 2026-10-29T00:48:21.274Z. First observed: 2026-10-24T14:19:57.916Z. Retrieved: 2026-10-24T14:21:35.301Z. - [Facebook claims its new chatbot beats Google’s as the best in the world](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Limits Text as retrieved on October 24, 2026; originally published April 29, 2020. The page was edited after publication; passages describing later facts may be additions and are not treated here as part of the original report. The article describes Facebook’s evaluation and plans; it does not establish how Blender would perform in real-world use or how well the proposed safety mechanisms would work. ## Open question How well would a toxic-language classifier detect harmful meanings that depend on conversational context? ## Source-stated MIT Technology Review reported that Facebook had open-sourced Blender, a chatbot intended for broad, engaging conversation. > Now Facebook has open-sourced a new chatbot that it claims can talk about nearly anything in an engaging and interesting way. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated Blender was trained first on Reddit conversations, then fine-tuned with datasets intended to develop empathy, knowledge and personality. > It was first trained on 1.5 billion publicly available Reddit conversations, to give it a foundation for generating responses in a dialogue. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated In Facebook’s tests, human evaluators rated Blender more engaging and more human-sounding than Meena. > In Facebook’s own tests, however, 75% of human evaluators found Blender more engaging than Meena, and 67% found it to sound more like a human. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated The article reports that Blender could make up facts and that the team planned to explore adding a knowledge database. > Blender also has a tendency to “hallucinate” knowledge, or make up facts—a direct limitation of the deep-learning techniques used to build it. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated The article says the team did not filter the Reddit dataset for harmful language and cautions that a classifier could miss context-dependent harm. > but it did not do the same for the Reddit data set because of its size. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## Editorial provenance Version 1. Publication ID: 81e5136c-86a6-7d5a-8bd0-d829a784d25b. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T00:45:51.004Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_screen. Observation: 5255bcceef232d0d5f7a4386208bfc667b629d381c47a94c3d1df40b62010880. Source text hash: e7ca107883e0b0636cb20b3a5a5f9899ad71684365826e4a760ec7be80b84ac6. ## Revision and source-check history ### Version 1 — 2026-10-29T00:48:21.274Z MIT Technology Review reported on April 29, 2020, that Facebook had open-sourced Blender, a chatbot designed for open-ended conversation. Facebook said the system could discuss many topics in an engaging way, and its tests compared Blender with Google’s Meena using human evaluators. The article describes Blender’s training on Reddit conversations, followed by fine-tuning on data intended to support empathy, knowledge and personality. Facebook’s evaluators rated Blender more engaging and more human-sounding than Meena, but the article notes that these results came from short conversations. The researchers expected longer chats to expose problems with coherence and memory. The article also describes a risk of Blender making up facts, because it generates responses from statistical correlations rather than a knowledge database. Facebook’s team planned to explore adding such a database. The article reports that harmful or biased responses were another concern: Reddit data was not filtered for harmful language, and a toxic-language classifier might miss how a response changes meaning in context. How well would a toxic-language classifier detect harmful meanings that depend on conversational context? ## Original source MIT Technology Review · Article Original source date: Apr 29, 2020 (2020-04-29). Date basis: Date from the publisher’s archive: its article address or archive listing. Added to Escape Meridian: 2026-10-29T00:48:21.274Z. First observed: 2026-10-24T14:19:57.916Z. Retrieved: 2026-10-24T14:21:35.301Z. - [Facebook claims its new chatbot beats Google’s as the best in the world](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Limits Text as retrieved on October 24, 2026; originally published April 29, 2020. The page was edited after publication; passages describing later facts may be additions and are not treated here as part of the original report. The article describes Facebook’s evaluation and plans; it does not establish how Blender would perform in real-world use or how well the proposed safety mechanisms would work. ## Open question How well would a toxic-language classifier detect harmful meanings that depend on conversational context? ## Source-stated MIT Technology Review reported that Facebook had open-sourced Blender, a chatbot intended for broad, engaging conversation. > Now Facebook has open-sourced a new chatbot that it claims can talk about nearly anything in an engaging and interesting way. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated Blender was trained first on Reddit conversations, then fine-tuned with datasets intended to develop empathy, knowledge and personality. > It was first trained on 1.5 billion publicly available Reddit conversations, to give it a foundation for generating responses in a dialogue. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated In Facebook’s tests, human evaluators rated Blender more engaging and more human-sounding than Meena. > In Facebook’s own tests, however, 75% of human evaluators found Blender more engaging than Meena, and 67% found it to sound more like a human. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated The article reports that Blender could make up facts and that the team planned to explore adding a knowledge database. > Blender also has a tendency to “hallucinate” knowledge, or make up facts—a direct limitation of the deep-learning techniques used to build it. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source-stated The article says the team did not filter the Reddit dataset for harmful language and cautions that a classifier could miss context-dependent harm. > but it did not do the same for the Reddit data set because of its size. - [Supporting source](https://www.technologyreview.com/2020/04/29/1000795/facebook-ai-chatbot-blender-beats-google-meena/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## Editorial provenance Version 1. Publication ID: 81e5136c-86a6-7d5a-8bd0-d829a784d25b. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T00:45:51.004Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_screen. Observation: 5255bcceef232d0d5f7a4386208bfc667b629d381c47a94c3d1df40b62010880. Source text hash: e7ca107883e0b0636cb20b3a5a5f9899ad71684365826e4a760ec7be80b84ac6. ### passed — 2026-10-29T00:45:51.004Z — anthropic/claude-opus-5.5 Publication: 81e5136c-86a6-7d5a-8bd0-d829a784d25b attribution — pass: The reading consistently attributes the claims to MIT Technology Review's reporting and to Facebook's own claims and tests. It does not present the evaluation results as independently verified. > In Facebook’s own tests, however, 75% of human evaluators found Blender more engaging than Meena summary — pass: The reading clearly identifies the actor (Facebook), the development (an open-sourced conversational chatbot) and the comparison with Google's Meena. It explains 'making up facts' in plain terms and covers the limitations the article stresses: short conversations, memory, hallucination and toxicity. The question follows from the evidence. Terms like 'open-sourced' and 'fine-tuning' are lightly explained or clear from context, so they do not block understanding. > Thus far, the team has evaluated the chatbot only on short conversations with 14 turns. date — pass: April 29, 2020 matches evidence.originalPublished at day precision and the URL path. The page was edited after publication (modified 2020-12-10), but every cited passage comes from the main article body, which is plainly part of the original report. The later 'Deep Dive' sidebar describes subsequent events, but the reading neither uses nor relies on it. Plans are framed as plans, and there is no hindsight framing. The kind is publisher-page, so no archive disclosure is required. > The team plans to experiment with integrating a knowledge database into the chatbot’s response generation. support — pass: Each claim matches the article. The claims cover the Reddit pretraining and three-skill fine-tuning, and the evaluator preferences over Meena. They also cover the short-conversation caveat and the researchers' suspicion about longer chats and memory. The rest cover hallucination and the planned knowledge database, the unfiltered Reddit data, and the classifier's context limits. The limitations correctly avoid turning test results into real-world reliability. > The researchers admit, however, that this approach won’t be comprehensive. media — pass: No media is selected and no verified media candidates were supplied, so nothing is misidentified or irrelevant. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/4151db96a803793d9cf3a12575802fc2ea827a7835d8fea954976110d3d9f640/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-reports-facebook-s-claims-for-its-blender-4151db96.json)