AI Search Overview: Not Ready For Prime Time

AI Search Overview: Not Ready For Prime Time

Google AI search overview is very slowly annoying me. Causing me to scroll past it or ignore it. Forcing me to change habits. Looking at source. Output of search has gotten worse with time. Now it is so obvious it is frustrating.

Let me give you a good example of the problem. Below are two definitions from Google:

Result A: Evaluating is the systematic process of assessing information, actions, or proposals to judge their value, quality, or effectiveness. It involves gathering evidence and analyzing results to determine success and identify areas for improvement.

Result B: Evaluating means to systematically judge, calculate, or assess the quality, value, importance, or significance of something. It is the critical process of analyzing data, performance, or behaviors to determine how successful an action was and identify areas for improvement.

Subtle differences between the two. Same prompt, different result. I was able to get this by using a different browser. Appears to be local cache not causing you to refetch the same thing. Seems the model it uses is quick and creates a new output over caching it. Common searches don’t appear to be stored.

What I’m saying is Google deals with billions of searches per day so they have gone with the cheap route. To my understanding, this is very similar to how Google search works. A machine learning model that tries to predict what words are coming next. These overviews seem to be doing something different. Creating a different output based on the same word as a result. Require more computer power.

My understanding of how it works in a simple explanation: It takes user input, turns it into a token, tries to estimate what is next in terms of letters and words and exports a new token. It does this by cross referencing against its own model, if it can’t get a good answer it will search the web.

That leads us to the next point. If it can’t find it, it makes it up or tries to find it. Errors are built into its design. At no stage does it do an error check.

It’s going ah “Of course I know what evaluating means, it’s in my dictionary. Here’s what I think it means based on what knowledge I have stored.” Sometimes however because it is creating a new output it adds or removes.

It’s similar behavior to humans if you ask them. Can you tell me what this is. Sometimes we misremember. Only difference is a black box, we don’t know when it got that behavior from or how it got the answer. It could be 1,000 models ago the mistake was added. Thing is if you ask it 1,000 times it may give you 700 times the same result, 200 less so and the rest could be rubbish.

I’ve spoken about this sort of thing before. But with it being included in search, it just makes it feel worse. Search was already bad, now it’s even worse and unpredictable. Seems companies are rushing to include it not understanding. Loads of hype. Some simple testing would highlight the problem.

I used to use search as a spell checker. It was quite good at predicting what the next letter is. Now trying to figure out what I mean is less helpful at that. Many ways companies are not using tech correctly. A built in smart spell checker or an improved one at the machine level would be useful. I been using these models for spell checking and oftens grammar checks. What happens is. Oversimplely things or overcorrect things. Ends up losing the meaning. What this shows is prompt matters. Human review matters. And finally until output the same unless it required not to be. I have written about AI before. And been a while yet the flaws remain.