[jdlyga]

It's tough convincing people that Google AI overviews are often very wrong. People think that if it's displayed so prominently on Google, it must be factually accurate right?

"AI responses may include mistakes. Learn more"

It's not mistakes, half the time it's completely wrong and total bullshit information. Even comparing it to other AI, if you put the same question into GPT 5.2 or Gemini, you get much more accurate answers.

[alex1138]

It absolutely baffles me they didn't do more work or testing on this. Their (unofficial??) motto is literally Search. That's what they're known for. The fact it's trash is an unbelievably damning indictment of what they are

[danudey]

Testing on what? It produces answers, that's all it's meant to do. Not correct answers or factual answers; just answers.

Every AI company seems to push two points:

1. (Loudly) Our AI can accelerate human learning and understanding and push humanity into a new age of enlightenment.

2. (Fine print) Our AI cannot be relied on for any learning or understanding and it's entirely up to you to figure out if what our AI has confidently told you, and is vehemently arguing is factual, is even remotely correct in any sense whatsoever.

[bethekidyouwant]

Testing what every possible combination of words? Did they test their search results before AI in this way?

[AlienRobot]

My favorite part of the AI overview is when it says "X is Y (20 sources)" and you click on the sources and Ctrl+F "X is Y" and none of them seem verbatim what the AI is saying they said so you're left wondering if the AI just made it up completely or it paraphrased something that is actually written in one of the sources.

If only we had the technology to display verbatim the text from a webpage in another webpage.

[gowld]

That's because decent (but still flawed) GenAI is expensive. The AI Overview model is even cheaper than the AI Mode model, which is cheaper than the Gemini free model, which is cheaper than the Gemini Thinking model, which is cheaer than the Gemini Pro model, which is still very misleading when working on human language source content. (It's much better at math and code).

[WarmWash]

I have yet to see a single person in my day to day life not immediately reference AI overviews when looking something up.