The Case For AI: What It's Actually Good At

I've spent two posts being hard on this technology. I told you it's pattern-matching dressed up as magic, and that the buildings powering it are being built with too little transparency and too few rules. I stand by all of it.

Now let me tell you why I still work in this field.

But first, a ground rule, because this is where most pro-AI writing goes wrong. I'm not going to show you demos, predictions, or press releases. Everything in this post is a documented result: published trials, real incidents, measured outcomes. If AI is as useful as I claim, it shouldn't need borrowed hype. Watch whether I keep that promise.

First, the pattern

Here's the frame that makes everything else make sense. AI shines when four things line up: the task is narrow and well-defined, there's a mountain of data to learn from, there's a checkable right answer, and a human expert stays in the loop. Every genuine success story you're about to read fits that shape. Almost every AI disaster you've heard about broke at least one of those four.

Keep that checklist. It's more useful than anything a salesperson will ever tell you.

It's catching cancers doctors miss

The strongest evidence in medicine right now comes from Sweden. The MASAI trial enrolled over 105,000 women in a national breast cancer screening program and randomly split them: half got standard screening, where two radiologists read every mammogram, and half got AI-supported screening. This is a randomized controlled trial, the gold standard of medical evidence, not a company benchmark.

The final results, published in The Lancet in 2026, are hard to argue with: the AI-supported group had substantially higher cancer detection, caught more of the aggressive cancers that would otherwise surface between screenings, added no extra false alarms, and cut the radiologists' screen-reading workload by 44%.

Notice the checklist at work. Narrow task: is there a tumor in this image? Mountains of data: millions of prior mammograms. Checkable answer: biopsies and follow-up either confirm it or they don't. Human in the loop: the AI flags, a radiologist decides. Nobody got replaced. The doctors got a second set of eyes that never gets tired at 4pm, and some women are alive because of it. That's the honest headline.

It cracked a problem biology was stuck on for 50 years

Proteins are the machinery of life, and what a protein does depends on the 3D shape it folds into. For half a century, figuring out that shape from a protein's chemical sequence was brutal work: months or years of lab effort per protein. Then an AI system called AlphaFold learned the folding patterns from every known structure and started predicting new ones in minutes, accurately enough that researchers now treat its predictions as a standard starting point. Its creators won the 2024 Nobel Prize in Chemistry for it, and the predictions are used daily in labs working on drugs, vaccines, and diseases.

Same checklist. Narrow question, huge dataset, verifiable answer (the lab can check the structure), scientists still driving. This is what it looks like when the technology is pointed at the right kind of problem.

And in my world, it's already stopping attacks

Cybersecurity is where I live, so let me tell you what defense actually looks like now. A city, hospital, or company network generates millions of events per day: logins, file transfers, connections. No human team can read that. Machine learning can, and it's been quietly flagging the weird stuff — the login from two countries in one hour, the printer suddenly uploading gigabytes at 3am — for years. My own doctoral research was on exactly this: intrusion detection, and making the AI explain why it flagged something, so the human analyst can judge the call instead of blindly trusting it.

But the newer story is better. In 2024, an AI agent from Google called Big Sleep found a previously unknown, exploitable flaw in SQLite, a piece of software running quietly inside billions of devices, including probably several in your house. Years of conventional testing and expert review had missed it. Then in 2025 it went a step further: tipped off by threat intelligence, Big Sleep found a critical SQLite flaw that, per Google, was known only to attackers and about to be exploited. The hole was patched before it could be used. An AI found the burglar's key before the burglar could turn it.

I'll be straight with you about the shadow side, because you've earned it by reading this far: the same capability that finds flaws for defenders can find flaws for attackers, and that arms race is very real. It's one of the reasons the governance vacuum in post 5 should bother you. But "this tool is powerful enough to matter on both sides" is an argument for taking it seriously, not for pretending it doesn't work.

Also: you already like AI, you just don't call it that

One more honest observation. The spam filter that keeps your inbox usable? Machine learning. The text from your bank asking "did you really buy this?" thirty seconds after a sketchy charge? Machine learning, watching billions of transactions for patterns no human could hold in their head. Nobody protests these, because they're boring, narrow, and quietly good at their jobs. A lot of "AI" is exactly this: invisible pattern-matching that works. The controversy mostly attaches to the loud new stuff — fair enough — but the ledger should include the boring wins too.

What I'm not claiming

I'm not claiming AI will cure cancer, replace your doctor, or secure the internet. I'm claiming something narrower and better-supported: on well-chosen problems, with humans in charge and results actually measured, this technology is already saving lives, accelerating science, and stopping attacks. Those are facts you can check — every claim above has a paper trail.

The trouble starts when AI gets deployed on problems that fail the checklist: fuzzy tasks, thin data, no way to check the answer, no expert watching. That's where the confident nonsense, the bias, and the expensive failures live.

Next in the series: that failure mode, up close — where AI goes wrong, and why. Because if I only told you what this technology gets right, I'd be just another salesperson, and you've got enough of those.