Monday, 6 July 2026

I find myself thinking about a peculiar tension running through this week's AI developments: we're simultaneously becoming more cautious about how AI behaves and more eager to deploy it everywhere.

Start with Anthropic's recent research paper, which makes an unexpected argument—that we should, in fact, anthropomorphize AI in certain contexts. This flies against the tech industry's long-standing taboo, yet their reasoning appears sound. If we're building systems that interact with humans in increasingly sophisticated ways, understanding them through a human-centered lens might actually help us predict failure modes. By the way, this matters because the alternative—treating these systems as pure black boxes—hasn't exactly been working out. The more opaque our relationship with AI becomes, the less we understand what goes wrong.

Which brings me to the real crisis in deployment, one that Wandero AI's CEO articulated clearly this week: shipping an agent is trivial now. The hard part is knowing what it's doing after launch. This is the operational blindness I keep warning about. Companies can spin up agents that handle customer service, procurement, or trading in minutes. But observability? Explainability? That's where the friction lives. We've solved the "go build" problem; we're still fumbling with the "go understand" problem.

Meanwhile, the regulatory environment is forcing a different kind of reckoning. ByteDance and Alibaba are disabling customized features in Doubao and Qwen ahead of Beijing's July 15 rules on humanlike AI interaction services. This is less about safety engineering and more about compliance theater—but it signals something real: governments are no longer waiting for the industry to self-regulate on anthropomorphism. They're mandating it. By the way, this creates an interesting asymmetry. Western companies are debating whether to anthropomorphize; Chinese companies are being told they can't. That gap will likely shape competitive dynamics.

Microsoft's move to merge its consumer and enterprise Copilot apps, now running on GPT-5 with an adaptive "smart mode," reflects a different impulse entirely—consolidation and user experience. There's nothing wrong with that, but it's worth noting we're building these systems without full clarity on what they're doing at scale, while simultaneously tightening the regulatory screws in some markets and loosening them in others.

The real question isn't whether anthropomorphization is good or bad. It's whether we can build observability fast enough to match our deployment speed. Right now, we're moving quicker on capability than on understanding.