Two tips and an energy answer

Issue #57

On today’s quest:

— Avoiding Google’s AI overview
— AI for alt-text
— Do reasoning models hallucinate more?
— This is what I’ve been trying to say about AI and energy
— Video generation uses a shocking amount of energy
— AI video just got a lot more realistic

Tip: AI-free search

I continue to find Google’s AI Overviews to be especially useless, so here’s a tip to both free up space on your results screen and to stop wasting the energy that’s used to create the summaries¹ :

Tip: AI for alt text

Adding alt text to images makes them accessible to people who are visually impaired, which can mean anyone from people who use screen readers to people who just can’t see all the tiny details in an image because it’s too small on their phone. Also, some people won’t share images without alt text, so adding it makes it more likely your image will get shared.

But writing good alt text can take a lot of typing time, especially when your image has a lot of text in it, and chatbots do a great job of writing alt text.

I use the following prompt so often I made a keyboard shortcut for it. I almost always edit the result — the chatbot may get a color wrong or be more wordy than I think is necessary. I could probably cut that down by fiddling with the prompt more — but just getting an immediate first draft with all the correct words in an image saves a lot of typing time, and it’s always better than the auto-generated alt text on platforms such as Facebook.

Also, if you take the time to write entertaining alt text, people often notice. I’ve never tried using AI for that kind of alt text, and I don’t imagine I will since the fun is the whole point. But I definitely appreciate AI for work-a-day alt text.

Do reasoning models hallucinate more?

An interesting article in the New York Times reported that researchers are finding that reasoning models actually hallucinate more than older models.

But multiple AI experts posted debunkings claiming the article was misleading. For example, Ethan Mollick posted a chart to back up his claims. And here’s another more detailed debunking of the NYT article.

One thing that crossed my mind as I was reading the New York Times article (and before I saw the posts from critics) is that the source of some of the NYT data didn’t seem clear. I actually wanted to confirm some of it myself, which some people might consider ironic. People say, “How is AI helpful if you have to confirm everything?” But I often have to confirm information I get from humans too.

This is what I’ve been trying to say about AI and energy!

Can I just say, “Thank God!”? The Bennett Institute for Public Policy at the University of Cambridge has come out with a blog post that says everything I’ve wanted to say about AI and climate but haven’t been able to pull together from my 20+ pages of notes.

The bottom line is that AI is a systemic problem, not an individual problem, and we need to focus on solutions at the regulatory and company level.

In the context of an individual’s life, AI doesn’t use that much energy. The calculation I came up with but never felt confident enough to publish was that 100 average ChatGPT queries use about as much energy as heating a cup of water in the microwave. The Bennett report says roasting a tray of potatoes in an electric oven is about the same as 1,379 ChatGPT queries.

You frequently hear that a single ChatGPT query uses as much energy as 10 Google searches — which may no longer be true, but even if it is, have you ever once thought about how much energy you are using when you do a Google search?

We do many things each day without a second thought that use more energy or water than using a chatbot. The rapid growth of AI is part of the climate problem, and I still don’t use it frivolously, but we need to think about all our energy uses and whether they are worth it, not just AI.

Read the whole short Bennett report.

If you’re especially interested, these two medium posts also influenced a lot of my thinking:

• Why using ChatGPT is not bad for the environment—a cheat sheet

• Why ChatGPT is not bad for the environment (long)

AI video generation uses shocking amounts of energy

On a related note, new data shows that AI video uses a shocking amount of energy — far more than text or image generation. So much that it seems like the responsible thing to do is to avoid using it unless you’re sure it’s giving you a huge benefit. I would never use it just for fun or even to test it after seeing the numbers. (Making a 5-second video is equivalent to running a microwave for an hour.)

AI video gets way more realistic

Speaking of AI video destroying the world, Google’s new video generation model, Veo3, is a huge breakthrough, increasing the quality of AI-generated videos so much that it’s now easy to create short videos that are indistinguishable from reality.

PJ Ace on Twitter/X posted a TV commercial for a fake pharmaceutical that cures depression by attracting puppies that he says he made for $500.

Ethan Mollick has a short montage of examples on his LinkedIn page.

Putchuon made a four-minute montage of examples on YouTube illustrating the fringe “prompt theory.”

I was depressed when I saw the following prompt out in the world, but actually, I’m sure it will get much worse:

Remember to always check the source of the media you consume — especially if it’s inflammatory or selling something. And don’t assume you are immune to being fooled.

Quick Hits

Health

Artificial intelligence improves breast cancer detection in mammography screening — Nature Medicine

AI helped cure my chronic pain — WSJ

Business

In April, 40% of U.S. businesses had paid AI subscriptions, compared to 36% in March, and 29% in February — Ramp

AI use damages professional reputations. In a hiring simulation, managers who didn't use AI themselves were less likely to hire candidates who regularly used AI tools. However, managers who frequently used AI showed the opposite preference, favoring the AI-using candidates. — Ars Technica

Duolingo is adding 148 new language courses thanks to AI — The Verge

Klarna abandons its AI-only approach to customer service, says it led to lower quality — Fortune

Education

ChatGPT gave Black students lower scores on essays than Asian students, replicating the bias seen from human graders on the same large essay set — The 74

AI isn’t replacing student writing — but it is reshaping it — The Conversation

Researchers are divided on when it’s OK to use AI to write research papers — Nature

Miscellaneous

New ‘grief tech’ apps allow grieving families to communicate with dead relatives via an AI afterlife — The Observer

Family uses AI to create video for deadly Chandler road rage victim's own impact statement — ABC 15 Arizona

Why AI is better when we don’t let it have opinions — The Atlantic

This ‘AI granny’ bores scammers to tears — PC Mag

Pope Leo XIV compares AI to the industrial revolution and calls it one of the most critical matters facing humanity today. — Independent

AI language models develop social norms like groups of people (which could lead to bias) — Nature

Another paper shows AI is more persuasive than the average human, even when the humans had financial incentives — Ethan Mollick

What is AI Sidequest?

Using AI isn’t my main job, and it probably isn’t yours either. I’m Mignon Fogarty, and Grammar Girl is my main gig, but I haven’t seen a technology this transformative since the development of the internet, and I want to learn about it. I bet you do too.

So here we are! Sidequesting together.

If you like the newsletter, please share it with a friend.

Written by a human.

1. Of course, people will suggest using a different search engine such as DuckDuckGo, Kagi, or Ernie Smith’s udm14, but old habits die hard.