How Often Does Google AI Actually Generate A Response?

[Parmandur]

Who knows, but if you cuss in the search it will circumvent the AI result (like "how do I make a cussing burrito" will give you traditional search results).

 

[The Firebird]

The motivation for my question was about energy efficiency. Generating a new, unique result to every query with AI, no matter how simple and straightforward, seems like it could be horrifically energy inefficient.

Iirc it is about 5 times as much energy to generate an AI response than google search. More than search, but I don't recall hearing that search was an environmental disaster. I think the environmental issues with AI are more because of training and/or widespread adoption for other purposes moreso than adding it to google search.

 

[trappedslider]

The motivation for my question was about energy efficiency. Generating a new, unique result to every query with AI, no matter how simple and straightforward, seems like it could be horrifically energy inefficient.

“Median energy use per AI text prompt is approximately 0.24 Wh, comparable to only a few seconds of television viewing"

Measuring the environmental impact of delivering AI at Google Scale

The transformative power of AI is undeniable - but as user adoption accelerates, so does the need to understand and mitigate the environmental impact of AI serving. However, no studies have measured AI serving environmental metrics in a production environment. This paper addresses this gap by...

arxiv.org

 

[Umbran]

“Median energy use per AI text prompt is approximately 0.24 Wh, comparable to only a few seconds of television viewing"

Sure. Be remember that we aren't really talking about one use.

A million text prompts become like 104 days of TV viewing.

And when there's more than a million people using it, and each person is using it over and over and over in a day, the aggregate isn't great.

 

[Whizbang Dustyboots]

I think that's what it's intended to do, though in reality I find it is frequently inaccurate, and I've seen it have all correct search results, but the summary is some complete off-topic nonsense! So it seems like it must also be drawing from sources which aren't the shown results (indeed sometimes if you click the links it gives, those don't go to any front-page search results).

I am guessing the summary is going off of the user-created page summaries for each page, which are often pretty half-assed by a lot of sites, which would mean it's scraping lower quality data than the actual page (which is usually more complete and up to date).

 

[The Firebird]

Sure. Be remember that we aren't really talking about one use.

A million text prompts become like 104 days of TV viewing.

And when there's more than a million people using it, and each person is using it over and over and over in a day, the aggregate isn't great.

Isn't average TV use in the US like 3 hours a day? At 1 prompt = 5 seconds of TV viewing, you need 720 searches to match 1 hour.

 

[Umbran]

Isn't average TV use in the US like 3 hours a day? At 1 prompt = 5 seconds of TV viewing, you need 720 searches to match 1 hour.

So, the point of comparing to TV is merely to give you an idea of energy scale.

It should not be the basis upon which we choose if the activity is worth the cost - that's the fallacy of whataboutism. We ought to look at the energy use for the value it gets us, not for the value some other activity gets us.

Like, watching TV gets us family time. Common genAI use gets us... removal of humans from generation of content. Those are totally the same, right?

And, as an aside, the article that trappedslider referenced said, "Less than nine seconds." I would be fine with you then choosing 8 seconds. Less accepting of 5 seconds.

 

[aco175]

I was googling about the noble house in Waterdeep and wanted to know the names of people tied to the green dragon in Kryptgarten Forest to the north. There is an adventure tied to the nobles trying to kill it and such and I was planning for an upcoming adventure. I found the info I needed and the timeline of the people was 100-200 years before the current time, so I wanted to check the AI part if there was some NPCs in the current time I could use.

The AI part did show some stuff from an adventure, but it looked to be a 3rd party thing and nothing more 'canon' or 'official' leaving me with just making up some new nobles tied to that family. I could have used the 3rd party stuff, but if I publish it, I would be using their work unknowingly and not give them credit if I just went with what AI told me.

 

[The Firebird]

So, the point of comparing to TV is merely to give you an idea of energy scale.

It should not be the basis upon which we choose if the activity is worth the cost - that's the fallacy of whataboutism. We ought to look at the energy use for the value it gets us, not for the value some other activity gets us.

Like, watching TV gets us family time. Common genAI use gets us... removal of humans from generation of content. Those are totally the same, right?

And, as an aside, the article that trappedslider referenced said, "Less than nine seconds." I would be fine with you then choosing 8 seconds. Less accepting of 5 seconds.

Turns out, I don't want to do this now.

 

[trappedslider]

So, the point of comparing to TV is merely to give you an idea of energy scale.

It should not be the basis upon which we choose if the activity is worth the cost - that's the fallacy of whataboutism. We ought to look at the energy use for the value it gets us, not for the value some other activity gets us.

Like, watching TV gets us family time. Common genAI use gets us... removal of humans from generation of content. Those are totally the same, right?

Value is in the eyes of the user...what use is family time to me if i don't have one or don't like my family? Common genAI use gets us whatever the user uses it for at low to no cost depending on if the user uses the free version or pay the sub cost.

I've hinted at the following but here's the full image and cites:

SOURCES FOR ENERGY COMPARISONS IN THE POSTER

AI ENERGY USE (BASELINE)

Elsworth et al. (2024), arXiv

“Median energy use per AI text prompt is approximately 0.24 Wh, comparable to only a few seconds of television viewing.”

Measuring the environmental impact of delivering AI at Google Scale

VIDEO STREAMING VS AI

National Centre for AI (UK), Jisc

“Video streaming services such as Netflix and YouTube consume substantially more energy per hour of use than generative AI services.”

https://nationalcentreforai.jiscinv...ronment-putting-the-numbers-into-perspective/

NETFLIX / YOUTUBE / ZOOM COMPARISONS

Forbes, John Koetsier

“Streaming video and video conferencing services like Netflix, YouTube, and Zoom use significantly more energy per hour than typical AI interactions.”

New Data: AI Is Almost Green Compared To Netflix, Zoom, YouTube

DATA CENTER CONTEXT (AI IS A FRACTION OF TOTAL LOAD)

Pew Research Center

“AI workloads are only one portion of total data center electricity demand, which is dominated by video streaming, cloud services, and everyday online activity.”

What we know about energy use at U.S. data centers amid the AI boom

GAMING AND HIGH-INTENSITY ONLINE SERVICES

Mucky Paws Analysis (compiled from industry energy data)

“Cloud gaming and high-performance PC gaming can consume orders of magnitude more electricity per hour than AI text-based services.”

Is AI Really the Energy Villain?

GENERAL CARBON / ENERGY CONTEXT

Sustainability by Numbers

“Per-query emissions from large language models are small compared to common digital activities such as streaming video or gaming.”

What's the carbon footprint of using ChatGPT?

As an aside, I am glad however that as far as i am aware no one from here has issued death threats over the use of AI.