AI News, 16 November 2026

Monday 16 November

Manchester Animation Festival bars generative AI films from 2027 competition

AI-generated summary Manchester Animation Festival announced that it will reject competition entries starting in 2027 if any part of their final moving image uses generative AI tools. Festival director Dr Steve Henderson linked the policy to supporting animation artists and craftspeople while addressing concerns about generative systems.

Earlier:

From Local Prediction Error to Final Outputs: A Theory of Expert Sharing in Multi-Step Reasoning

From Local Prediction Error to Final Outputs: A Theory of Expert Sharing in Multi-Step Reasoning

AI-generated summary Researchers analyze mixture-of-experts models that share experts across layers or computational steps to study how local prediction error reduction affects final outputs. The study examines linear regression models and trained Transformers to determine when sharing improves local estimation and when those gains survive to the final output.

Gemini skills add file support and sharing ahead of Gems deprecation

AI-generated summary Google is introducing file support and sharing capabilities to Gemini skills ahead of the Gems shutdown beginning tomorrow. The updated skills will support files up to 100 megabytes from various sources and allow users to grant view or edit permissions.

Also covered byAndroid AuthorityAI Ping
Earlier:

Amazon Bedrock's Built-In Cost Attribution and Integration with IAM Identity Center

AI-generated summary Amazon Bedrock provides built-in cost attribution mechanisms that allow developers to trace every inference call back to the responsible IAM principal. These attribution fields integrate into CUR 2.0 to appear in the line_item_iam_principal column and cost allocation tags.

Earlier:

What Is an "Embedded" AI Safety Evaluator, and How Is It Different From Testing a Model Before Launch?

AI-generated summary OpenAI stated that four outside groups will work inside the company as embedded safety assessors. Apollo Research stated that its earlier evaluations of frontier models took place after post-training and shortly before public release.

Earlier:

SOV-Transcribe-1: Recorded Audio and Deployment

AI-generated summary Salesforce released the SOV-Transcribe-1 model on November 12, 2026, for non-streaming speech recognition with an emphasis on names, phone numbers, and email addresses. Salesforce reports 76.3% task success on VoiceCodeBench and 93.4% entity recovery, while the model card lists seven languages and 26.6B total parameters.

Earlier: