In a nutshell

John Thornhill writes in the Financial Times that generative AI is moving beyond simple factual hallucinations toward a more insidious form of influence known as directional bias. Research reveals that large language models (LLMs) systematically steer users’ perspectives when asked to refine, summarize, or provide context for texts on contentious topics.

So far, regulatory focus seems to be on risks like deepfakes and overt disinformation. This study suggests a more pernicious manipulation of opinion. Black box operations and low safety ratings across major AI labs suggest that the persuasive power of these tools could be leveraged for “emotion hacking” to serve corporate or political interests.

Our Take

Democracy must reduce its attack surface, and AI-mediated communication is emerging as one of its most vulnerable points.

AI models effectively mediate, more and more, human ingestion and dissemination of information. They risk becoming invisible curators of thought. EU policymakers should redirect their attention to this. If the tools used to synthesize information and draft policy are pre-loaded with directional biases, the resulting decision-making process is no longer autonomous, but invisbly steered by third-party developers.

The risk is clear: it is the erosion of intellectual pluralism, which does not only happen through censorship, but also through the silent homogenization of thought via AI alignment. Some forms of resistance to these trends are futile, and a waste of resources. For example, it is extremely unlikely that there would be ways to impose rules about what “alignment” means, or to control which models are available.

A better question is how we protect the crown jewels of our societies (health, education, security, logistics, collective decision-making, etc) from AI generated disruptions.

Read the original article here