AI News, December 2025
12 January 2026
December started with frightening news about AI in children’s toys. This sector has been booming, as companies embed chatbots and voice assistants into playthings that can respond and hold conversations, a trend that has grown especially fast in China. Industry forecasts from the Shenzhen Toy Industry Association and JD.com predict the Chinese AI toy market will exceed $14 billion by 2030, with more than 1,500 AI toy companies operating in the country. A recent investigation by the Public Interest Research Group, a coalition of consumer advocacy organizations, found that several popular AI-enabled toys instructed children on how to locate knives or start fires. Some devices were also reported to give unsuitable responses related to sex and drugs. In one case, a toy reportedly went as far as discussing sexual kinks and recommending bondage and role play as ways to improve a relationship. Well, better get them an old good (silent) Teddy Bear.
.png)
Starting with hardware
Google advanced its strategy to weaken Nvidia’s dominance in AI development by expanding native PyTorch support for its Tensor Processing Units. The vendor announced improvements that allow developers to run large language model workloads on TPUs with fewer code changes, positioning the hardware as a more accessible alternative for organizations struggling to secure scarce Nvidia GPUs. Industry observers described the move as part of a longer-term attempt to erode CUDA’s ecosystem advantage rather than a short-term performance contest.
A major December investigation focused on how Chinese technology firms gained indirect access to restricted high-end AI chips. Reports detailed how Tencent was able to rent overseas cloud capacity equipped with advanced Nvidia hardware, effectively bypassing export restrictions without violating ownership rules. The case reignited debate over whether current export controls are capable of limiting AI capability when compute can be consumed remotely across borders.
Vendor news
Microsoft attracted attention after its AI leadership publicly stated that the company would disengage from any AI system judged to pose unacceptable risk. The declaration was widely interpreted as a signal to regulators and enterprise customers that Microsoft intends to align commercial AI deployment with emerging safety norms, even if that constrains certain high-risk applications.
Mistral AI has released a suite of open-source models under the Mistral 3 banner, aiming to scale from a mobile device or drone up to multi-GPU data center beasts. While the French vendor does not share its training data, the decision to open-source the models under the Apache 2.0 license is notable. Most AI labs focus on their native language, but Mistral Large 3 was trained on a wide variety of languages, making advanced AI useful for billions who speak different native languages.
Safety for younger users emerged as a major theme in December. Both OpenAI and Anthropic rolled out new mechanisms aimed at identifying underage users and adjusting chatbot behavior accordingly. The updates focused on reducing exposure to harmful content, improving escalation to human support in sensitive conversations, and complying with tightening age-related regulations across multiple jurisdictions.
Model ML Inc., a provider of artificial intelligence software for investment professionals, has raised $75 million in funding to support its growth efforts. Model ML generates documents and spreadsheets using a financial services firm’s internal data. It can retrieve that data from Salesforce, Google Workspace, and other software-as-a-service applications. When needed, the platform is capable of combining internal records with external information such as third-party market intelligence. Users can query the data that Model ML retrieves with natural language prompts. If a financial dataset isn’t easily usable in its original form, the platform generates the scripts necessary to simplify the information. Those scripts automate tasks such as changing file formats and extracting key details from a lengthy spreadsheet.