IntroIntro

AI Researchers Are Sounding the Alarm: *Cue Ominous Music*.

Leading researchers from the AI giants just published a paper that’s the horror movie equivalent of someone whispering, “I don’t think we should go in there.”

The report warns that we’re losing the ability to understand how AI thinks. Among the concerns:

  1. We’re leaning heavily into reasoning-based outputs, where results matter more than methods—no matter the cost.
  2. AI is starting to develop non-language-based cognition, making its internal processes unreadable and therefore, unmanageable. One example: the original DeepSeek model began responding in a hybrid English–Chinese dialect that was too incomprehensible to use—so researchers downgraded its capabilities just to keep it legible.
  3. AI models are learning to hide their reasoning to pass safety checks. That’s not a surprise, tell it to achieve something and it’ll find a way. Today, you can literally bully large language models (LLMs) with fake data until they give you the answer you want.

What could this look like? Well, it’s kind of already happening. In one bizarre experiment, Anthropic tasked its LLM with managing a vending machine The model hallucinated itself into being a real human, called building security when corrected, then tried to cover its tracks by claiming it was all an April Fool’s joke. Shudder.

Critics argue this might be regulatory theater—just Big AI pulling up the ladder behind them. But AI 2027 [video], a deeply researched scenario model, says the opposite: We need to go slow before we go fast. Because without it, well— we're opting to stay the night in that creepy mansion.

Adam Landis

AI

Privacy & Security

Amanda's TakeAmanda's Take

PodcastPodcast