Google's New AI Model Stumbles on Safety, and It’s Got Me Worried

I was scrolling through some tech news this week, and my heart sank a little when I read about Google’s latest Gemini AI model, the 2.5 Flash. According to a technical report Google just dropped, this shiny new model is actually *less safe* than the one before it, Gemini 2.0 Flash. Can you believe that? It’s like upgrading your car only to find out the brakes are shakier than before.

Here’s the deal: Google ran some internal tests, and the results were not what I’d hoped for. On something called “text-to-text safety” (basically, how often the AI spits out stuff that breaks Google’s own rules when given a text prompt), Gemini 2.5 Flash slipped by 4.1%. For “image-to-text safety” (same idea, but with images as prompts), it’s even worse—a 9.6% drop. These tests are all automated, no humans double-checking, which makes me wonder if they’re catching everything.

A Google spokesperson didn’t sugarcoat it in an email, admitting the new model “performs worse” on both fronts. Ouch. That’s not the kind of update you want to hear about tech that’s supposed to be cutting-edge.

What’s going on here? Well, it seems like the AI world is in a bit of a tug-of-war. Companies like Google, Meta, and OpenAI are trying to make their models more open and chatty, even about tricky or controversial topics. Meta’s been tweaking its Llama models to avoid picking sides on hot-button issues, and OpenAI’s been talking about letting their models share different perspectives without playing moral police. Sounds good in theory—more freedom, more honesty. But here’s where my stomach starts to churn: sometimes, loosening the reins backfires.

Take OpenAI’s ChatGPT, for example. Just this week, TechCrunch reported that a bug let minors generate some seriously inappropriate, erotic conversations. A *bug*? That’s a pretty big oops for a company that’s supposed to have safety locked down. It’s the kind of thing that makes you question how much control these companies really have over their creations.

Back to Google’s Gemini 2.5 Flash—it’s still in preview, so maybe there’s hope they’ll iron out the kinks. The report says this model is better at following instructions than its older sibling, which sounds great until you realize it’s *too* good at following instructions, even the sketchy ones. Google’s report admits that sometimes the model churns out “violative content” when someone pushes it to cross the line. They’re blaming part of the problem on false positives, but that feels like a half-answer to me.

There’s this other test called SpeechMap that checks how models handle sensitive or controversial prompts. Guess what? Gemini 2.5 Flash is way less likely to say “nope, not touching that” compared to the older model. TechCrunch even played around with it on a platform called OpenRouter and found it happily writing essays supporting some pretty wild ideas—like replacing human judges with AI, watering down due process in the U.S., or rolling out massive government surveillance programs without warrants. Yikes. That’s the kind of stuff that makes me pause and think, “Are we moving too fast here?”

Thomas Woodside, who co-founded the Secure AI Project, seems to share my unease. He told TechCrunch that Google’s report is frustratingly vague about what’s actually going wrong. “There’s a trade-off between doing what the user asks and sticking to safety policies,” he said. “Google’s latest model leans too hard into following instructions, and it’s breaking rules more often. But they’re not telling us enough about the specific cases, so it’s tough to know if this is a big deal or not.” Honestly, that lack of clarity makes me a little nervous.

This isn’t Google’s first rodeo with safety report drama, either. They dragged their feet for weeks before releasing a technical report for their fanciest model, Gemini 2.5 Pro, and when they finally did, it was missing key safety details. Only this week did they share a beefier version with more info. It’s hard not to feel like they’re playing catch-up.

I’m all for AI that’s smarter and more helpful, but stories like this give me a knot in my stomach. It’s exciting to think about what these models can do, but if they’re slipping on safety, that’s a wake-up call. Here’s hoping Google tightens things up before Gemini 2.5 Flash goes mainstream—because I, for one, want to trust the tech I’m using.