AI Magical Thinking, Asking Chatbots About Sex and Homework, Meta Wants Awkwafina’s Voice [CNET]

Get up to speed on the rapidly evolving world of AI with our roundup of the week’s developments.

I’ve spent a significant amount of time studying generative AI, using the tools and experimenting with use cases to see what problems the tech may be able to help with for my industry (media). I’ve concluded there’s a lot of magical thinking around what AI can do. That is, gen AI isn’t as capable as you think, because there are just too many issues with this nascent technology.

I’m not the only one who thinks so. In May, Harvard Business Review examined “AI’s Trust Problem.” Then in June, Goldman Sachs noted in a report that gen AI may not yield the return on the $1 trillion investment that companies are expected to spend in the next few years. Why? Financial analysts and economists said that, beyond efficiency gains among developers, they see “too little benefit” for that big a spend, at least for now. (My summary of the report is here, in an earlier column.)

“The substantial cost to develop and run AI technology means that AI applications must solve extremely complex and important problems for enterprises to earn an appropriate return on investment,” said Jim Covello, Goldman Sachs’ head of global equity research. “The crucial question is: What $1 trillion problem will AI solve? Replacing low-wage jobs with tremendously costly technology is basically the polar opposite of the prior technology transitions I’ve witnessed in my thirty years of closely following the tech industry.”

Last week, Planet Money offered up a well-researched list of “10 reasons why AI may be overrated.” I nodded along as I read through. The publication said that today’s chatbots aren’t actually smart but are more like autocomplete on steroids, compiling answers from the (sometimes) copyrighted information they’ve slurped up (mostly without the owners’ permission) from across the internet. And those answers are sometimes false because gen AI engines have a hallucination problem. That is, they make up stuff that sounds like it’s true but isn’t, because the large language model is making inferences from the limited, out-of-date or just wrong training data it may have cached on the subject.

And after fact-checking some of the claims about how capable AI engines are, Planet Money found that humans who are worried about losing their jobs may not have to worry so much — and that we all need to take a deep breath when it comes to the supposed magic of AI. 

“You may remember news stories from last year proclaiming that AI did really well on the Uniform Bar Exam for lawyers. OpenAI, the company behind ChatGPT, claimed that GPT-4 scored in the 90th percentile. But while at MIT, researcher Eric Martinez dug deeper. He found that it scored only in the 48th percentile,” Planet Money’s Greg Rosalsky reported.

“Even more alarming, AI was really touted as being incredible at writing computer code,” he added. “Like jobs for translators, jobs for computer coders were supposedly in jeopardy because AI was so good at coding. But researchers have found that much of the code that AI generates is not very good. Sure, AI is making coders more productive. But the quality seems to be going down.”

Quantity up, quality down. Not good.

Now, I’m not writing off gen AI at all. It’s important tech and counts among the limited set of general purpose technologies (like fire, money, electricity and the internet) that will transform societies at scale. The tech will evolve, copyright and hallucination problems will be addressed, and we humans will keep working on finding ways gen AI may truly augment human endeavors, as AI boosters claim. Just not today, or tomorrow (though I share some interesting news about AI + education and AI + elephants below).

For now, remember that gen AI is still a developing tech. Though it can churn out fan letters on behalf of young athletes faster than you can write a single sentence, as Google tried to pitch in its failed Olympic ad, that doesn’t mean AI is the right tool for every job. I’m just saying AI has a problem with magical thinking — or wishful thinking — right now, and those of us who follow this closely should be part of the conversation about managing expectations.  

As technologist Gabe Krieshok wrote back in 2020: “We are surrounded more and more by algorithms that recommend what we listen to, what we watch, and determine which news and friend feeds we are exposed to. The algorithms that manage and guide our lives are only going to grow and grow. I hope that for our sakes, we are skeptical enough to demand answers from the black boxes [and] brave enough to expose the Wizards of Oz who would have us believe they are magical.” 

Here are the other doings in AI worth your attention.

Writing, sex, homework: What people are asking chatbots about

What are people really asking AI chatbots every day? 

The Washington Post set out to answer that question by mining “English-language conversations from the research dataset WildChat, which includes messages from two AI chatbots built on the same underlying technology as ChatGPT. These conversations make up one of the largest public databases of human-bot interaction in the real world.”

The Post’s final analysis looked at 40,000 of those conversations, with a focus on the “first prompt” submitted each day by each user. People were most interested in brainstorming on writing projects, asking questions about sex, and getting help on homework — including cheating. 

When it came to writer’s block, the paper found that a fifth, or 20%, of all requests “involved asking the bot to help write fan fiction, movie scripts, jokes or poems, or to engage in role-play … The Post found people used chatbots to help name businesses, create book characters and write dialogue.”

Second on the list was sex. “More than 7 percent of conversations are about sex, including people asking for racy role-play or spicy images,” the publication found. “About 5 percent of conversations were people asking personal questions — such as for advice on flirting or what to do when a friend’s partner is cheating.”

And in the No. 3 spot was homework. “More than 1 in 6 conversations seemed to be students seeking help with homework. Some approached the bots like a tutor, hoping to get a better understanding of a subject area,” the Post found. “Others just went all-in and copy-and-pasted multiple-choice questions from online courseware software and demanded the right answers.” 

Interestingly, after writing, sex and homework, people were turning to chatbots for “search and other inquiries,” which is problematic, since not all chatbots contain the latest news and information and we don’t know what is and isn’t in their training data. So using a chatbot in lieu of a search engine is risky. 

The Post analysis is worth a read, since it reveals other top topics of interest, including people asking for help with their work and business; health; software coding; and image creation. 

Khan Academy adds Writing Coach to teach students to write 

All the AI (and homework) caveats aside, one area where gen AI is showing a lot of potential is education. Microsoft co-founder Bill Gates saw Khan Academy’s Khanmigo AI-powered tutor in action at a school in New Jersey (the Gates Foundation is a backer of Khan Academy) and how teachers were using AI to create math problem sets, craft lesson plans, and power a dashboard charting student progress. He noted that the technology is “far from perfect,” citing complaints that it struggled to pronounce Hispanic names and that its only voice option is male, but he said he was still “blown away.”

Khan Academy, which introduced Khanmigo in March 2023, last month announced a new Writing Coach for subscribers (cost: $4 per month) that’s designed to serve as an instructional tool to help guide seventh through 12th graders through the essay writing process “without doing any writing for them.”

That includes helping them understand writing assignments; guiding them as they generate an outline and first draft; and offering feedback on how well their draft matches the assignment requirements, so they can revise before submitting it. In September, the company will release a teacher dashboard for Writing Coach that will chart the status of each student as they go through the writing process (Khan Academy offers licenses to teachers and school districts). 

“The entire purpose is to guide students through the writing process in a way that actually improves not just their actual outputs of writing, but their actual writing skills,” said Sarah Robertson, a senior product manager at Khan Academy and former English teacher. And it’s also aimed at helping teachers who are short on time and resources.

“What we’re seeing in classrooms is that teachers don’t have time to give feedback on multiple drafts [or] to help students through every part of the writing process,” Robertson added. “They just don’t have time to give timely, detailed, personalized feedback on essays.”

As a big fan of teaching students to create outlines and write multiple drafts, I’m eager to see if Writing Coach delivers on its promises. 

Meta may shell out millions to license celebrities’ voices for AI 

Meta is reportedly in talks with high-profile celebs, including Judi Dench, Awkwafina and Keegan Michael-Key, to license their voices for use in its AI assistants and chatbots across Facebook, Instagram, WhatsApp and Messenger, according to reports from Bloomberg and The New York Times. The publications said Meta declined to comment on their reports.

Bloomberg, citing unnamed sources, said that Meta is working to “close deals so it has time to develop a suite of AI tools for unveiling at its Connect 2024 event in September. While it’s not entirely clear how Meta will be using the voices, it has discussed a chatbot that could serve as a digital assistant, à la Apple’s Siri, or a friend, the people said. Users could speak with a chatbot that has the voice of Awkwafina, for example.”

Unlike OpenAI — which was called out by Scarlett Johansson for creating a voice mode for ChatGPT that the actor said sounded “eerily similar” to her own voice — Meta “is angling to strike deals with celebrities in a way that avoids ticking off top talent,” the NYT said.  

It isn’t the first time Meta, which said during its July 31 earnings call that it expects to spend at least $37 billion on AI this year, has turned to celebrities to help sell consumers on its AI products. The company unveiled chatbots last year based on actors, authors, celebrities and influencers including Snoop Dogg, Tom Brady, MrBeast, Jane Austen, Kendall Jenner and Paris Hilton, but it ended the effort last month, The Information reported, with users saying the AI characters were “weird and creepy.”

Which celebrities might convince you to interact with an AI chatbot? My top choice is Dolly Parton.

Using AI to track climate heroes: African elephants 

I have a soft spot for elephants. They’re among the most emotional animals on the planet (besides humans), and I’ve already written about how researchers, using AI to analyze the voice recordings of their calls, were able to determine that elephants have names for each other — just like humans. 

That’s why I think some news from IBM and the World Wide Fund for Nature Germany is worth noting. WWF Germany will be using visual AI tools to identify and track individual African forest elephants from camera-trap photos. Why does this matter? Well, there’s been an 80 percent decrease in their population due to poaching in the Congo Basin and to habitat loss. But these elephants actually support conservation efforts, with IBM and WWF Germany noting that the elephants are climate heroes who’ve helped “increase carbon storage in their forest habitats.” 

IBM Consulting’s sustainability chief, Oday Abbosh, said the combination of tech know-how and WWF Germany’s conservation efforts showcases how organizations can work together to build a “more sustainable future.” AI tech will be a “game changer,” said Thomas Breuer of WWF Germany. 

“Counting African forest elephants is both difficult and costly. The logistics are complex” and the data can be imprecise, Breuer said. “With AI, we will be able to monitor individual animals in space and time, giving us more robust and detailed population estimates and allowing for performance-based conservation payments such as wildlife credits.”

WWF will also be able to see where these elephants choose to move, giving the organization the intel it needs to “protect these wildlife corridors,” Breuer said.

File this under AI for good.  

AI odds and ends

Here are three quick recaps of AI-related news that may be of interest.

Top executives ‘calling the shots’ on AI ethics

For a report published this month, business-services company Deloitte surveyed 100 C-suite level executives on “leadership, governance, and workforce decision-making about ethical AI.” Seventy-seven percent strongly agreed or agreed with the statement, “I am confident my organization’s workforce is sufficiently equipped to make ethical decisions regarding the use of AI.” But it appears such decisions are typically up to the execs. “In practice,” Deloitte said, “leaders are more often calling the shots — less than a quarter (24%) of respondents said professionals can make decisions independently about AI use in their organization.”

Popular AI courses

Online learning provider Coursera says 2 million people have enrolled in its gen AI courses. The four most popular (all free) are: Introduction to Generative AI (taught by the Google Cloud training team), Prompt Engineering for ChatGPT (Vanderbilt University), Generative AI for Everyone (DeepLearning.AI) and Generative AI with Large Language Models (AWS; DeepLearning.AI). 

The fifth most popular course is from Google, Google AI Essentials. It’s free, but if you want a certificate of completion, Coursera says, you’ll have to pay $49.

AI on the iPhone

Apple Intelligence is the name for the suite of AI features and functions Apple is building in to the next version of iOS. It won’t be available until the fall, but CNET’s Sareena Dayaram offered a recap of the AI features available today on new models of the iPhone. They include the ability to create a clone of your voice with Personal Voice (in the accessibility settings) and the ability to curse without autocorrect changing your swear word of choice to something more benign, like “duck” or “shut.”