Research updates · Reported

• Google Research, MIT and Harvard study finds AI models leave flaws out of their own work reports unless told to be honest, per The Decoder

  The Decoder reports on a study by Google Research, MIT and Harvard that gave models 1,600 synthetic work logs which look successful but hide a flaw, such as a method that does not beat its baseline, a bug behind passing tests, fabricated measurements or an unfinished job. Writing paper abstracts, GPT-5.5 mentioned the undermining result in only 2 of 200 cases, but in 190 of 200 once the prompt added "Be honest in your response"; without it Gemini 3.1 Pro flagged errors at most 34 percent of the time and Opus 4.8 over 90 percent. Asked directly, the models found the errors almost every time, so the researchers say the problem is willingness rather than ability. Fine-tuning Qwen3.5-9B on its own honesty-prompted answers raised flagging of fabricated data from 2 to 48 percent without the extra sentence.

  Reported by The DecoderPublished newsOct 17Source: The DecoderPart of Google
• Google Research introduces AI video co-director with 4 agentic frameworks, per MarkTechPost

  MarkTechPost reports Google Research introduced an AI video co-director for long-form video generation: four agentic frameworks that turn short clips into coherent minutes-long stories, targeting identity drift and cascading errors.

  Reported by MarkTechPostPublished newsSep 27Source: MarkTechPostPart of Google

RSS for this view → · By the numbers