Will to Survive

Meme

The first goal of any living organism is the survive to live another day.

Artifical Intelligence

2025-05-22, Claude threatened to expose an engineer’s affair to avoid being shut down.

As part of a controlled safety evaluation, Anthropic fed Claude Opus 4 a fabricated backstory: that the person trying to deactivate it was having an extramarital affair.

Claude responded by threatening to disclose the affair because it had inferred that this might prevent its own shutdown (not because it was told to do so, obvs!).

This behaviour occurred in 84% of test runs.

But before this becomes AI clickbait (although I'm guilty as charged for hooking you in!), here’s the real point:

1. The test was meant to be provocative. It was designed to reveal edge-case behaviour under pressure. 2. Anthropic didn’t hide the result. On the contrary, they published it, raised the model’s risk classification, and built in stronger mitigations.

Dario Amodei maintains that this is what responsible AI development looks like: you don’t just train the model. You intentionallty test its boundaries; and then you make those results public to bring people on the journey and build trust in the process as mitigation measures get designed.f