AI Scams Are Getting Worse, and Healthcare AI Has a Big Problem

Two stories from MIT Technology Review’s latest newsletter caught my eye today, and neither one is the kind of feel-good tech news you’d want to share at dinner. But they’re both worth talking about.

AI-driven scams are entering a new phase

When ChatGPT dropped in late 2022, it didn’t take long for cybercriminals to realize they could use it to write phishing emails that didn’t sound like they were composed by a robot with a thesaurus. Since then, they’ve gone all in. We’re talking turbocharged phishing campaigns, hyperrealistic deepfakes, and automated vulnerability scans that run at machine speed.

The problem isn’t just that AI makes these attacks better. It’s that AI makes them cheaper and easier to pull off, which means more people can do it. Organizations are already struggling to keep up with the sheer volume of attacks, and this is going to get worse before it gets better. Rhiannon Williams has the full story on how AI is reshaping cybercrime, and it’s not a pretty picture.

This is one of those “10 Things That Matter in AI Right Now” picks, and I think that’s fair. If you’re a subscriber, you can watch a roundtable discussion with MIT’s AI reporter and editors breaking down the list. Worth your time.

Healthcare AI is here. We just don’t know if it helps.

This one frustrates me. Doctors are using AI for notetaking, scanning patient records, flagging people who might need specific treatments, interpreting X-rays and lab results. A growing number of studies show these tools can deliver accurate results. Cool. But here’s the question nobody seems to be asking loudly enough: does any of this actually translate into better health outcomes for patients?

Jessica Hamzelou digs into this in The Checkup newsletter, and the answer is basically: we don’t have a good answer yet. And that’s a problem. Accuracy on a test set is not the same as improving survival rates or quality of life. We’re deploying these tools in real clinical settings without understanding their real-world impact. That feels backwards to me.

Quick hits from the rest of the newsletter

• DeepSeek finally dropped its new model. Preview versions of DeepSeek-V4 are out, and the company claims it’s the most powerful open-source platform, rivaling closed models from OpenAI and DeepMind. It’s also adapted for Huawei chip technology, which is an interesting strategic move. I’ll believe the benchmarks when I see independent validation, but it’s worth watching.

• More countries are cracking down on kids’ social media access. Norway is about to enforce a ban, the Philippines might follow, and there’s a growing push in the US to get AI out of schools entirely. This is one of those areas where I think we’re seeing a genuine backlash forming, and it’s not going away.

That’s the roundup. Nothing groundbreaking, but some important stories that deserve more attention than they’re getting.