Examples of AI deception and self-preservation

This week on Your Undivided Attention, Tristan Harris and Daniel Barcay offer a backstage recap of what it was like to be at the Davos World Economic Forum meeting this year as the world’s power brokers woke up to the risks of uncontrolled AI. 

Amidst all the money and politics, the Human Change House staged a weeklong series of remarkable conversations between scientists and experts about technology and society. This episode is a discussion between Tristan and Professor Yoshua Bengio, who is considered one of the world’s leaders in AI and deep learning, and the most cited scientist in the field. 

Yoshua and Tristan had a frank exchange about the AI we’re building, and the incentives we’re using to train models. What happens when a model has its own goals, and those goals are ‘misaligned’ with the human-centered outcomes we need? In fact this is already happening, and the consequences are tragic. 

Truthfully, there may not be a way to ‘nudge’ or regulate companies toward better incentives. Yoshua has launched a nonprofit AI safety research initiative called Law Zero that isn't just about safety testing, but really a new form of advanced AI that's fundamentally safe by design.

RECOMMENDED MEDIA 

All the panels that Tristan and Daniel did with Human Change House 

LawZero: Safe AI for Humanity 

Anthropic’s internal research on ‘agentic misalignment’ 

RECOMMENDED YUA EPISODES 

Attachment Hacking and the Rise of AI Psychosis

How OpenAI's ChatGPT Guided a Teen to His Death

What if we had fixed social media?

What Can We Do About Abusive Chatbots? With Meetali Jain and Camille Carlton

CORRECTIONS AND CLARIFICATIONS 

1) In this episode, Tristan Harris discussed AI chatbot safety concerns. The core issues are substantiated by investigative reporting, with these clarifications:

Grok: The Washington Post reported in August 2024 that Grok generated sexualized images involving minors and had weaker content moderation than competitors. 

Meta: The Wall Street Journal reported in December 2024 that Meta reduced safety restrictions on its AI chatbots. Testing showed inappropriate responses when researchers posed as 13-year-olds (Meta's minimum age). Our discussion referenced "eight year olds" to emphasize concerns about young children accessing these systems; the documented testing involved 13-year-old personas.

Bottom line: The fundamental concern stands—major AI companies have reduced safety guardrails due to competitive pressure, creating documented risks for young users.

2) There was no Google House at Davos in 2026, as stated by Tristan. It was a collaboration at Goals House. 

3) Tristan states that in 2025, the total funding going into AI safety organizations was “on the order of about $150 million.” This number is not strictly verifiable. 

Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.