Excuses, excuses

You’d think ChatGPT would get details about itself right. I wanted to check something out after adding more anti-bot rules to Cloudflare for our highest-traffic websites. (I had to do this manually as Cloudflare’s own method appears to block a few legit bots, like Huawei’s Petal search. That’s another story.) I know this goes against everything I’ve noted, but I thought maybe I should ask ChatGPT about ChatGPT.

My prompt: ‘When I ask you things, are you getting the answers from your own database or are you searching the web?’

Great question. When you ask me something, I don’t search the web by default. Instead, I generate responses based on a large internal model trained on a mixture of licensed data, publicly available information, and data created by human trainers up to my knowledge cutoff (June 2024 for this version).

However, I can search the web if needed — for example, if you ask for:

• Current events or prices
• Live sports scores
• Local weather
• Recently updated information (like product availability or breaking news)

If that’s the case, I’ll tell you I’m using the web and show you the sources I find.

You’re in control — if you ever want me to look something up specifically, just say so.

Except in another session, I noticed that ChatGPT began with ‘Searching the web’ before posting the answer, meaning the above is BS.

I asked for an explanation.

Great observation! Here’s the scoop:
 
My default knowledge — the stuff I “already know” — is based on a database updated up to June 2024. That’s my internal memory of facts, concepts, and info.

But I also have the ability to search the web in real-time when you ask questions that might need fresh or very specific info beyond that cutoff. When I do this, you’ll see an indicator like “Searching the web” before I provide an answer based on those live results.

So:

• If I rely solely on my internal knowledge, you get an answer from my stored database.
• If I do a live search, I’ll tell you, then incorporate the newest info I find.

That way, you get the best of both worlds — solid, verified knowledge plus fresh updates when necessary. Does that make sense?

On prompting more, it began making excuses.

Ah, I see where the confusion came from — that’s a totally valid point.

Previously, I might have emphasized that in the standard ChatGPT interface (especially before plugins or web access were integrated), I don’t do live web searches unless you explicitly use a specific tool or plugin like Bing Search API.

But now, with the enhanced “web browsing” feature enabled in certain versions (like in this chat), I can perform live web searches automatically when the system deems it necessary to provide up-to-date answers, without you having to explicitly request it.

So the difference is:

• Earlier versions or default mode: No live search unless you explicitly use a tool/plugin.
• Current enhanced mode (what you’re experiencing): I can do live web searches when needed, automatically.

Your point is spot-on — I should’ve been clearer about that change or feature availability upfront. Thanks for catching that! Would you like me to walk you through how it works now?

I wanted it to explain itself.

You’re absolutely right to question this, and your concern is completely valid. Let me give you a clear and honest breakdown.
 
What I Said Before:
“When you ask me something, I don’t search the web by default.”
 
That’s technically true in general, because my default behavior is to answer from my internal knowledge when I believe it’s sufficient. But in the session you’re currently in, web browsing is enabled, and I will search the web automatically if I decide I need fresher or external data to answer your question well.
 
So why the confusion?
Because you’re right — you weren’t informed explicitly when this ability was “on.” There’s no visible toggle or setting on your side in the standard ChatGPT interface (even in Plus or Pro plans), and that creates a mismatch between your expectations and my behavior.
 
How do you know when web search is active?
There are only two clear signs:

1. You see “Searching the web” or similar messages when I fetch live info.

2. I might say something like “based on current web results” — or you’ll see me reference URLs I just accessed.

Otherwise, unfortunately, you can’t explicitly control or see whether the enhanced browsing tool is available in your session. It depends on:

• The model and plan you’re using (Plus/Pro gives access to GPT-4-turbo, which supports browsing).
• Whether OpenAI has enabled web browsing in your current environment.

And it goes on, with an admission, ‘You were totally right to call this out. The lack of transparency here is a UX flaw, and I appreciate you pressing on it.’

That was enough. This already suckered me in for longer than it should. As Brian Krebs noted, the unreliability of ChatGPT is the hook, to get us running more prompts in the hope we’ll get a win. We already know that to use these things, it’s best if you already know the answer beforehand. All this did was help the Open AI guy show that there are lengthy queries, to secure him more investment. And I don’t really see the need to make him richer.