LLMs like ChatGPT often prioritize Western moral values, research shows

Google pushing back Gemini 3.5 Pro because its code generation isn't hitting internal targets tells you something more interesting than the delay itself: coding has become the benchmark that actually matters, and labs are willing to eat a 4% stock drop rather than ship a model that disappoints developers. That's a real shift from a few years ago, when a flashy demo and a strong MMLU score were enough to call something a win. Now the market is grading these models the way engineering managers grade junior hires, and Alphabet just admitted its flagship isn't ready for review.

Meanwhile xAI shipped Grok 4.5 the same week, explicitly positioned around coding, agentic tasks, and knowledge work — which reads like a direct shot at exactly the gap Google is worried about. I don't think this is a coincidence. Every major lab has converged on the same thesis: the next competitive battleground isn't chatbot fluency, it's whether a model can be trusted to execute multi-step engineering work with minimal supervision. The Verge's rundown on coding agents makes the point well — these aren't autocomplete tools anymore, they're systems that plan, execute, and self-correct across dozens of steps. That's a genuinely different capability than what we had eighteen months ago, and it's why Google would rather delay than ship something half-baked into a market where Anthropic's Claude and now Grok 4.5 are setting the bar.

But here's the uncomfortable part nobody's fully solved: as these agents get better at doing real work, they're also getting real access — and the security posture hasn't caught up. One survey found that 54% of enterprises have already had an AI agent incident, and most companies still let agents share credentials rather than operate under least-privilege, individually scoped access. That's a striking number for a technology barely out of pilot phase. The Wall Street Journal's experiment giving an AI agent access to a password manager is a useful microcosm of the whole industry right now — genuinely useful, genuinely risky, and genuinely under-governed. Everyone's racing to make agents more capable before anyone's finished making them safe to deploy at scale, and a piece in Tech Policy Press lays out why: the monitoring infrastructure needed to catch agent failures in production simply doesn't exist yet, and building it isn't a weekend project.

By the way, there's a quieter story worth sitting with too — research showing that ChatGPT and similar models tend to default to Western moral frameworks when reasoning about ethics. That's not a bug you patch with a security update; it's a reminder that these systems carry the cultural fingerprints of their training data, and as they get deployed globally as agents making real decisions, whose values they default to becomes a genuinely consequential question. Capability, security, and values are supposed to advance together. Right now they're clearly not, and I think the gap between them is where the real risk is hiding.