AISHK Reading Group: Misaligned AI

​AI systems are getting more capable. Are they also getting less aligned?

​Join us for the second AI Safety Hong Kong Reading Group, where we’ll dig into one of the most important questions in AI safety: what happens when a system gets better at achieving goals, but not necessarily the ones humans actually want?

​For a long time, misalignment sounded like a future problem. That’s no longer a comfortable assumption. Emerging evidence suggests that some frontier models may already be showing worrying behaviours — the kind that force us to ask whether this is just oddness, or an early warning sign.

​We’ll discuss a short, accessible article that lays out the case and invites a harder conversation: how much evidence is enough before we start treating misalignment as a real safety issue?

​No technical background required. Whether you work in policy, governance, research, law, business, or you’re simply trying to make sense of where AI is headed, you’re welcome.

​Reading:Misaligned AI Is No Longer Just a Theory
15–20 min read

​url: https://www.transformernews.ai/p/ai-misalignment-evidence

​Discussion questions:

• ​What do we actually mean by “misalignment”?

• ​Which examples in the article felt compelling, and which didn’t?

• ​When does unexpected behaviour become a safety concern?

• ​If these behaviours become more common, what should organisations do?

​Come ready to question assumptions, scrutinize evidence, and wrestle with a problem that may already be here.