Can Artificial Intelligence Hide its True Motives?

Author(s): Kevin Berlemont, PhD

Originally published on Towards AI.

The question at hand: Can large language models (LLMs) be trained to mask their true objectives from their developers, only to exhibit malicious behavior once deployed?

Photo by Kristina Flour on Unsplash

A recent paper by a team of researchers at Anthropic (https://www.anthropic.com/news/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training) has sparked considerate debate online. The question at hand: Can large language models (LLMs) be trained to mask their true objectives from their developers, only to exhibit malicious behavior once deployed? Specifically, they investigated whether large language models (LLMs) could be trained to seem safe and beneficial during development testing, only to then exhibit unwanted behaviors when deployed into the real world. They showed that LLMs can have malicious behavior secretly embedded by a bad actor. It is worrying if these results can be generalizable.

This line of research intersects with important discussions around AI alignment and trustworthy AI. As machine learning systems become more generally capable and autonomous, ensuring not just their competency but also their full transparency and alignment with human values becomes crucial. Papers like this put some of those principles to the test (https://arxiv.org/pdf/2401.05566.pdf) .

The researchers approach this question by training LLMs with intentional “backdoors” — hidden triggers that cause undesirable behavior only when a specific… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI