Researchers want to monitor AI's 'thoughts'

Leading researchers want tech industry to monitor AI’s ‘thoughts’

In a paper published yesterday, AI researchers from OpenAI, Google DeepMind, Anthropic and a group of nonprofits proposed chain of thought monitoring to observe how reasoning models “think.”

Quick note: Remember that current AI models don’t actually think. They generate responses by replicating patterns in their training data to predict the next word or action. It’s really just a whole lot of math disguised as humanlike thinking. 

What you should know:

• Your average large language model (LLM) is kind of a black box. Users input a prompt and the model spits out a response — but we don’t really know what happens in between.

• Reasoning models like OpenAI’s o3 work much differently. They use chains of thought* to work through problems step-by-step before providing an answer. 

• This paper’s authors argue that because these chains of thought are visible to users, researchers could monitor them to better understand how AI makes decisions. 

Why it matters: Chain of thought monitoring could give tech companies a way to flag potentially harmful interactions, and enable them to refine their models to correct that behavior. Researchers say this technique could help keep AI under control as it becomes more capable. 

What makes chain-of-thought reasoning models different from standard LLMs?

Training data. Standard LLMs are trained to predict the next word in a sentence, and they usually jump right into this process when users ask a question. Reasoning models are trained on problems that include the full step-by-step reasoning processes that lead to each answer. This teaches them to “think out loud,” making them more accurate and easier to evaluate. Again — there’s no real thinking happening here — even reasoning models are just replicating patterns from their training data. 

xAI launches Grok companions

Elon Musk’s xAI has launched Grok companions, allowing paid subscribers to interact with different “3D animated AI personas” powered by the company’s Grok chatbot. 

Key background: AI personas are quickly becoming a controversial topic in the industry. Character AI, one of the largest platforms that hosts chatbots with “personalities,” is facing a lawsuit from the mother of a 14-year-old user who committed suicide. 

The details:

• Grok currently offers at least two unique companions: a flirty anime girl named Ani and a red panda named Bad Rudi. 

• The paid feature reportedly supports NSFW interactions that allow users to build romantic relationships with these characters. 

• There’s not much research in this space, but some academic writing suggests (unsurprisingly) that relationships with AI bots can worsen loneliness. 

Going deeper: The lawsuit mentioned above is one of the most compelling cases against this type of product. The 14-year-old Florida boy became “noticeably withdrawn” after forming an unhealthy attachment to a chatbot based on Daenerys Targaryen from “Game of Thrones.” He committed suicide less than a year later, which some reports suggest the chatbot actually encouraged him to do. 

The bottom line: AI companions can be dangerous products, especially in the hands of children. 

More trending news

xAI claims it has fixed Grok 4’s problematic responses

Meta commits hundreds of billions of dollars to building new AI data centers

Nvidia will resume exporting H20 AI chips to China after U.S. reverses trade restrictions

Thanks for reading! How was today’s newsletter? My goal is to make these emails as easy to read as possible — if there’s anything I can do for you to achieve that goal, let me know!

See you tomorrow,

Chain of Thought