New benchmark shows AI still misses the human side of mental health care

There's a pattern emerging in AI right now that nobody's putting on a slide deck: the industry is quietly admitting its agents can't be trusted, even as it races to give them more autonomy. Look at Anthropic's own numbers on Claude Cowork. Half of all usage isn't some dazzling new workflow — it's administrative drudgery, the text-based busywork that nobody wanted to own in the first place. That's not a failure, exactly. It's actually a healthy sign of product-market fit. But it also tells you where the real, defensible value of agentic AI sits today: not in grand autonomous decision-making, but in absorbing the boring stuff humans were already doing badly.

Meanwhile, Nubia is showing off an "AI agent smartphone" at WAIC 2026 that books your flights and does your shopping without you touching the screen. I find this genuinely interesting and genuinely uncomfortable in equal measure. The technical demo is one thing; the trust gap is another entirely. Would you actually let a phone execute a purchase autonomously, with your payment details, based on its own interpretation of what you want? Elizabeth Fuentes at AWS has spent real effort cataloguing five code-centric methods to stop AI agents from hallucinating — semantic tool selection, Graph-RAG, and other guardrails that exist precisely because agents left unchecked will confidently do the wrong thing. That research wouldn't need to exist if agentic AI were as reliable as the marketing suggests. Nubia's launching a consumer product into exactly this gap between capability and trustworthiness.

And then there's the security side, which I think deserves more attention than it's getting. A report out this week finds that 99.9% of fixable AI vulnerabilities remain unpatched — not because they're hard to fix, but because organizations are expanding AI infrastructure faster than they can secure it. Agents, cloud services, packages, all proliferating while the patching discipline lags years behind. This is the classic pattern of every fast-moving tech wave: deployment outpaces governance, and the bill comes due later. By the way, it's worth connecting this to the Australian data point on AI agents becoming "the hack everyone's using" in finance and business functions — 64% adoption in finance teams alone. When adoption outpaces security by that margin, you're not looking at isolated incidents waiting to happen. You're looking at a structural exposure.

On the model side, there's a quieter but more encouraging story: efficiency gains are real and compounding. NOTA's inference optimization work, benchmarked at ICML, claims 7x speed with no performance loss. Google's compression research is aimed squarely at data center energy costs. These matter more than another leaderboard-topping model release, because inference cost is what determines whether AI agents are economically viable to run at scale, not just technically impressive in a demo.

The mental health benchmark, PsyEval, is a useful reality check threaded through all of this. It shows current LLMs still miss the human side of care — the part that isn't about knowledge retrieval but about presence. That's the tension defining 2026: we're making agents faster, cheaper, and more autonomous, while still struggling to make them trustworthy in the moments that matter most. Which of those two curves — capability or trust — closes the gap first?