AI Scams Are Getting Smarter, And We Still Can’t Tell If Healthcare AI Actually Works

I’ve been watching the AI landscape long enough to know when something shifts from interesting to genuinely concerning. Two stories crossed my desk today that capture that feeling perfectly.

AI-Powered Scams Are No Longer a Glimpse of the Future

When ChatGPT dropped in late 2022, a lot of us thought about the cool possibilities. Cybercriminals thought about something else entirely: finally, a way to write convincing phishing emails without having to hire a writer or learn English themselves.

That was just the beginning. Rhiannon Williams over at MIT Technology Review has a solid piece on how AI is now powering everything from hyper-realistic deepfakes to automated vulnerability scans. The volume of attacks is already overwhelming many organizations, and here’s the kicker: it’s only going to get worse. AI makes cyberattacks faster, cheaper, and easier to scale. As more criminals adopt these tools—and the tools themselves improve—we’re looking at a problem that compounds on itself.

This isn’t speculative. We’re already seeing AI-generated phishing emails that pass grammar checks, voice clones that fool family members, and video deepfakes that impersonate executives. The defenses exist, but they’re playing catch-up. And in a game where attackers only need to succeed once, that’s a scary place to be.

If you’re a subscriber, MIT Tech Review has a roundtable on this exact topic, part of their “10 Things That Matter in AI Right Now” list. Worth your time.

Healthcare AI: Deployed, But Not Yet Proven

On the other side of the coin, we have Jessica Hamzelou’s piece on healthcare AI. And this one frustrates me, because it’s a classic case of “we can, so we did” without asking “should we?”

Doctors are already using AI for notetaking, scanning patient records, flagging at-risk individuals, and even interpreting X-rays. Multiple studies show these tools deliver accurate results. Great. But here’s the question nobody seems to be answering: does any of this actually improve patient outcomes?

We don’t know. And that’s a problem.

There’s a gap between “the AI got the diagnosis right” and “the patient got better.” Maybe the AI flags a condition that the doctor would have caught anyway, adding no real value. Maybe it introduces false positives that lead to unnecessary stress and testing. Maybe it works brilliantly in one hospital system but fails in another because the patient population is different.

We’re deploying these tools at scale without the kind of rigorous, longitudinal studies that would tell us whether they’re worth the investment—and the risk. That’s not just sloppy; it’s irresponsible.

Hamzelou’s piece is from The Checkup, their weekly health and biotech newsletter. If this kind of reporting interests you, it’s worth signing up.

Quick Hits From Around the Web

A few other stories caught my eye today:

• DeepSeek finally released its V4 model, and it’s claiming it’s the most powerful open-source platform out there. They say it rivals top closed-source models from OpenAI and DeepMind. I’ll believe it when I see independent benchmarks, but the fact that they’ve adapted it for Huawei chips is an interesting geopolitical angle.
• More countries are cracking down on children’s social media access. Norway is enforcing a ban, the Philippines may follow, and there’s growing American sentiment to get AI out of schools entirely. The pendulum is swinging hard.

I’ll be watching both the scam and healthcare AI stories closely. The former because it affects everyone with an inbox; the latter because it affects everyone who will eventually need medical care. Which is, you know, all of us.