Weekend Roundup: The AI Hype Bubble Bursts

Growing pains as AI moves from inflated expectations to the trough of disillusionment

DALL-E prompt: “Digital art showing artificial intelligence getting revenge”

A few months ago I published several stories on the promising potential of generative AI tools in vetmed and beyond (see here and here). As people and companies have spent more time with these tools, the shine is wearing off these new toys and we’re beginning to see the growing pains of AI. In some ways, this is not surprising, and I’ll open this post about how this cycle of hype fits many technology developments.

The Gartner Hype Cycle

The IT consulting firm Gartner coined the phrase “hype cycle” and created the graph above to represent the typical stages of adoption, progress, and setbacks for newly introduced technologies. I think this is a good framework to think about AI, even though there is some debate on how accurate and applicable this model is.

The innovation trigger is the initial publicity around a promising new technology. Often, there is not even a clearly-defined use case or commercially viable product based on the science at this point. For large language model AI, I would say this point arrived some time between 2018 and 2020. This is when a number of breakthroughs in deep learning allowed rapid progress across fields that had never been possible before, culminating in the announcement of OpenAI’s GPT-3 model in 2020.

Next, stories about the early successes and breakthroughs lead to fevered hype about this new technology immediately disrupting every industry and creating untold new riches for early adopters (that term is from another conceptual curve for technology life cycles). THIS WILL CHANGE EVERYTHING, the narrative goes. At this point, plenty of hucksters and opportunists flood into the conversation, especially on social media, and you will see tons of people who are newly-(self)-appointed AI gurus (many of whom were crypto/web3 influencers only five minutes earlier). Any savvy start-up or established company that can plausibly start talking about future AI products will do so to drive investment from VCs, angel investors, and the public. This point was probably reached in 2021-2022. There were numerous events in this stage, but the biggest included the public release of OpenAI’s GPT-3 model and then ChatGPT, followed by Microsoft’s Bing chatbot, Google’s Bard, and a bunch of imitators and third-party apps enabled by APIs.

In my opinion, we’re starting to enter the trough of disillusionment in 2023 as we see cracks in the narrative and problems arise. The next few sections of this article will discuss a few examples. The key point is that for truly viable new technologies, once people persist through the setbacks, we learn how to best use the technology and mitigate its shortfalls, and eventually there is a productivity boost above baseline.

ChatGPT Goes to Court

One of the biggest AI face-palm moments I’ve seen recently involves lawyers who pasted unedited or fact-checked content from ChatGPT into actual legal briefs for clients in an aviation injury lawsuit. Readers of All Science probably know why this was a terrible idea, but for those who are newer to the subject: AI, particularly large language models, has a hallucination problem. It makes shit up. Constantly.

In this instance, ChatGPT cited multiple other examples of similar lawsuits about plane crashes to support the lawyers’ arguments. Only problem was…those cases don’t exist, and the crashes were completely fictional! The judge called them out and the lawyers initially defended their filings until it became untenable. The lawyers and their firm were fined and admonished by the judge:

“Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance,” [Judge] Castel wrote. “But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.”

The judge said the lawyers and their firm “abandoned their responsibilities when they submitted nonexistent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question.”

I applaud this judge’s nuanced opinion and hope it sets the precedent for future professionals using AI. Imagine if your physician sent you patient discharge instructions that cited fake research or worse, harmful treatment recommendations! Not great.

CNET Pulls Back on AI “Writers”

The computer technology website CNET was one of the earliest adopters of generative AI, starting to quietly use it to write articles for its site without any human oversight in November 2022. As you would expect from the legal story above, this new editorial practice came to light when problems were noted in the articles and required corrections after publication. These included numerous factual errors as well as outright plagiarism (sometimes the AI is “so good” because it just copies source material without attributing the correct citation).

An article in The Verge this January reported that more than 50% of CNET’s AI-written articles (41 of 77) required corrections. Journalists and editors may be expensive, but I’d wager most bat substantially better than 50/50 in quality and accuracy! CNET smartly hit pause on using generative AI for its articles without oversight. They were not the only outlet guilty of this practice as Men’s Health was similarly caught with an AI-generated article about low testosterone that contained at least 18 medical errors.

Undeterred, other news outlets continue to push forward with experiments in generative AI, including Gannett, one of the largest publishers. To their credit, Gannett claims it will have layers of human fact-checkers and prevent automatic publication of AI content without oversight:

[Gannett SVP] Turiano added, “The desire to go fast was a mistake for some of the other news services,” he said without singling out a specific outlet. “We’re not making that mistake.”

Gannett is hardly alone in its balancing act. For instance Reuters President Paul Bascobert said in a statement Thursday, responding to a reporter's request for comment about the company’s plans, that as the news agency embraces AI technologies, it is "taking a responsible approach that safeguards accuracy and fosters trust.”

Manual human oversight before publication would seem to be the bare minimum expectation here. After all, no human journalist at a major outlet gets their story published without fact-checkers confirming the details, so why wouldn’t the same apply to computer systems known for randomly lying? Many of these companies are rushing ahead because of a focus on the bottomline and desire to cut costs. But they should remember that it only saves you money if you don’t have to clean up a mess afterwards. As the old saying goes, “Measure twice, cut once.”

A Secret Invasion

A still from the opening credits of Marvel’s “Secret Invasion”

Recently, people discovered that the entire opening credits of Marvel’s new Disney+ streaming series “Secret Invasion” is AI generated. This has received backlash from fans online as well as visual artists in show business. The producers argue that this was not a cost-saving measure, but instead a deliberate creative decision to reflect the themes of the show, such as never being sure about who to trust.

To that, I say a big: maaaaaaaaaaayyyyyyybe… More likely in my opinion is that it was a mixed motivation at best, and the showrunners probably came up with this choice to reduce production costs and then convinced themselves “hey, maybe this actually fits!”

As some critics pointed out, using generative AI like this in the middle of the WGA writers strike—where concerns over AI eliminating jobs is one of their primary issues—is in poor taste at best:

I can certainly sympathize with striking writers and artists who feel threatened by AI. However, like the earlier examples in this article, I feel pretty comfortable that AI tools are simply not up to par to completely replace human talent in the short to medium term, and maybe not ever. Besides the outrage over the ethics of AI during the strike, most people who watched the Marvel credits felt the quality was just kinda crummy. Sure, AI might help jump start some ideas and speed up draft writing and design prototypes, but people pay to see human stories, not mediocre knock-offs.

I’m much more interested in how AI is used in original ways to do things that aren’t otherwise possible. For example, the use of de-aging AI in the new Indiana Jones movie, and previously in Martin Scorsese’s The Irishman. The effects aren’t yet perfect, but allow stories to be told using a single actor in a way that is more immersive than casting younger lookalikes. Another example—that was highly controversial—was using AI to resurrect Anthony Bourdain’s voice in the documentary Roadrunner. This upset some people, but I thought it worked in the context of the film.

A still from Indiana Jones 5 showing a younger version of the character that is played by the 80 year old Harrison Ford with the aid of AI

Using AI in media is going to be messy, with results that vary from awe-inspiring to cringeworthy, but let’s hope we see more AI as a springboard for unique ideas like this, and less of low-effort knock-offs.

AI Content “Contaminating” AI Training Sets

A “secret invasion” is an unintentionally apt phrase to describe how AI-generated text and images are now trickling into blogs, news articles, movies, Reddit, social media, and more. Setting aside the immediate factual and ethical issues with all of this, one big concern is that AI is trained on…the internet, and as more and more of the internet is AI-generated, does it begin to “contaminate” training sets and lead to loss of fidelity or even flattening and homogenizing culture? The Substack article below has great reflections on this issue and starts from the analogy about how the steel used in many sensitive applications has to be sourced from pre-WWII shipwrecks because of worldwide contamination with low-level radiation from the atomic era:

The Low-Background Steel of AI

How have you seen AI used in ways good and bad? How have your views on AI changed over the last year? Tell me in the comments!