Dani AI. Where's the opt-out or off switch?

[rproffitt]

I've seen Dani AI a few times now and all it needs is an opt-out or off switch.

[Dani]

If something like that does ever happen, it would be after my experiment is complete. Assume ~3 months (or until the next Google core algorithm update).

[Dani]

Sorry for my curt response. Two reasons to not want to do anything until after my experiment is complete:

• Turning off AI, even for a select group of logged in users, can affect the results of the experiment. You might say that Googlebot only sees what a logged-out user sees, but Googlebot also sees everything that Google Chrome users see.
• If the experiment fails, which it very well might, I will be going back to the drawing board. I don't want to pour my limited time into coding extra user-facing functionality or features for something that might disappear completely in 2 months. I would rather keep Dani AI under the radar, for now, with limited active members even noticing it. If I start announcing it as a new user-configurable feature with settings that can be modified in the control panel, etc., and then in 3 months from now, there's not a trace of the feature, and/or the feature works/functions completely differently, then that will just lead to UI/UX confusion.

[rproffitt]

Since DaniAI is essentially some member, put the vote mechanism on it.

Time to look again at Violentmonkey and Scriptcat to strip out the unwanted AI content.

Given my current campaigns and ongoing projects this will have to wait.

Let me be clear, the AI summaries I've read so far are awful. Seems a shame these cost you money. They should pay you for the data!

[Dani]

Since DaniAI is essentially some member, put the vote mechanism on it.

It’s not a member, though. Adding the vote mechanism to it would be extremely prohibitively time consuming. Also, it would send the wrong signals to Google, since Schema markup uses voting on a Q&A page exclusively for user-generated content.

As for the quality, I am still tweaking that non-stop. I’m working on my prompt engineering skills.

[Dani]

Yes… except for today. Today I’m on the road heading to wine country to one of our favorite bistros for lunch.

[rproffitt]

https://www.techrepublic.com/article/news-ai-browsers-security-flaw-perplexity-comet/ notes an exploit that is very nasty.

This one is serious enough I expect patches in a few days.

Meanwhile:
Brave suggests several fixes
Clearly separating user commands from web content.
Requiring user confirmation for sensitive actions.
Isolating AI browsing from regular browsing.
Until we figure all that out, maybe keep your AI browser away from your banking tabs.

[Dani]

I think that AI is definitely in its infancy, and, as such, there are going to be all sorts of bugs and exploits and this and that that are continuously popping up every day. IMHO the absolute worst are hallucinations, especially when you are lead to believe the information you're receiving is trustworthy, and where accuracy is paramount such as health-related queries. Hallucinations are the main reason I was unwilling to get behind AI for so long.

[Dani]

The problem I linked to lets the bad actors gain access to bank accounts. IMO, that sounds worse then AI hallucinations.

For sure, but as you pointed out, that's a bug that will soon be fixed. I think the problem with hallucinations is not an acute issue. Plus, I think that trusting an AI that gives you incorrect medical information can potentially have worse ramifications than a compromised bank account.

[Dani]

Also, it goes without saying that the bug you reported is not at all applicable to Dani AI overviews.

[Dani]

How would we know that? Not open source and you don't control which browser people use.

Sorry, I'm not quite understanding you. Dani AI overviews are generated via a cron job using the OpenAI API, and are subsequently stored in our database. What web browsers end-users use to access DaniWeb are completely irrelevant.

[rproffitt]

From the article:

Researchers discovered they could hide malicious instructions in regular web content (think Reddit comments or even invisible text on websites).
When users clicked “Summarize this page,” the AI would execute these hidden commands like a sleeper agent activated by a code word.

As such, your claim what web browser used is false.

For now, “The only winning move is not to play.” To be clear, AI browsers are as of now unsafe even here.

[Dani]

I think you still misunderstand me.

I don’t use any Summarize this Page functionality. When OpenAI generates the Dani AI overviews, it does not even know that the content that I use in my prompt to generate them comes from webpages, never mind what URLs the pages are to access them.

[Dani]

As such, your claim what web browser used is false.

As mentioned, no web browser is used to generate the Dani AI overviews.

[Dani]

The exploit is not DaniAI but use of an AI browser and user content. Sorry to write that the issue is complex.

But my statement holds that what web browser anyone is using is irrelevant to Dani AI. You disagreed with that above and I’m not sure why.

[Dani]

No AI Browser here

I feel like nearly all browsers are AI browsers these days. I have never sought out a browser that has any type of AI features, but I do use Google Chrome, and Google has said that Chrome has AI enhancements built-in. I've never used any of them, and honestly don't even know what they are. Aside from Chrome, I'm told that MS Edge, Firefox, Opera, Brave, etc. all have built-in AI functionality.

[Dani]

Chrome has an AI off switch for good reasons.

Now that you point that out, in my Chrome settings, I notice a section called AI innovations, with all 3 "innovations" turned off. I've never manually turned them off before, so I guess they start out as off by default.

[SCBWV]

There's MS Copilot in my Notepad now; it can easily be turned off, but it's there. I wonder how far Copilot's penetration will go in Windows 11?

[Dani]

So something I just started noticing today is, while using Google Chrome, the right side of the URL address bar says Homework Help when browsing programming threads here on DaniWeb, and when I click on it, it presents me with a list of related search terms. Not AI?

[Dani]

If something like that does ever happen, it would be after my experiment is complete. Assume ~3 months (or until the next Google core algorithm update).

I figured this thread is just about ready for an update. So, if you recall, I initially decided on a 3 month test because that's roughly how long I was expecting it to take for Google to pick-up on the new signals I was throwing its way. Well, Google just started noticing this week, and I have to say that the results are nothing short of a hockey puck-looking graph. It's crazy. Dwell time on page for users coming in from Google search results is at least 5x as they read the Dani AI overviews. Google started noticing that, and crawl rate shot up yesterday and today, and traffic is just starting to climb as of around last weekend, etc.

[Reverend Jim]

a hockey puck-looking graph

I have no idea what this means.

[Dani]

https://en.wikipedia.org/wiki/Hockey_stick_graph

[Dani]

Or https://en.wikipedia.org/wiki/J_curve :-P

[Reverend Jim]

OK. Hockey stick. Not hockey puck.

[Dani]

Oh, LOL!! Sorry, brain fog.

[Reverend Jim]

No True Canadian (fallacy, I know) would have made that mistake ;-)

[Dani]

LOL. Yeah, the brain fog has been really killer lately. I'm trying to imagine what a hockey puck graph would look like. A pie chart, perhaps?

[rproffitt]

An attempt was made. I went to a place to ask it to draw a hockey puck graph and this came back.

Looks like any other graph.

[Reverend Jim]

A pie chart, perhaps?

[rproffitt]

My take on Pie Charts.

Over in some data visualization discussion we move from pie charts to donut charts.
And then know that the size of the donut hole can change the view of the data.
Example: