It’s time to get serious about AI hallucinations

Cars traveling on a desert highway are engulfed in heat haze near Jaww Al Nasla in western Qatar on November 15, 2022.
Cars traveling on a desert highway are engulfed in heat haze near Jaww Al Nasla in western Qatar on November 15, 2022.
Odd Andersen—AFP/Getty Images

Hi there, it’s tech reporting fellow Rachyl Jones. The term “hallucination” has taken on a different meaning in recent years, as artificial intelligence models have become widely accessible. Once a term reserved for the experience of humans, we now say AI can “hallucinate,” or confidently spew inaccurate information as if it were true. As the technology becomes more intertwined with work and personal life, the phenomena could have real-world impact on how people make decisions. 

But just how common is the problem? Perhaps more than we think. 

The New York Times previously reported the rates at which popular AI models made up facts, with hallucinations ranging from OpenAI’s ChatGPT at 3% of the time to Google’s PaLM at a staggering 27%. 

According to a new survey, the problem may be even more widespread. Eighty-nine percent of machine learning (ML) engineers who work with generative AI say their models show signs of hallucination, according to survey results published Wednesday from ML observability platform Aporia. That’s a whole lot of delusional computing. 

The results also indicate that 93% of engineers encounter problems with their models on a daily or weekly basis. 

As a provider of observability tools that can help mitigate hallucinations, Aporia is not exactly an unbiased source on the topic. But if its survey—which was based on 1,000 engineers spanning industries—is even remotely accurate, it’s only a matter of time before these problems seep into our own everyday lives.

As Data Sheet discussed yesterday, AI is a key topic at the World Economic Forum in Davos, Switzerland, this week, with policymakers and AI experts convening dozens of panels to discuss the opportunities and dangers of AI, and its impact on elections, misinformation, and economic development. When it comes to AI hallucinations, the stakes are too high to delude ourselves.

On that note, here’s today’s biggest tech news.

Rachyl Jones

Want to send thoughts or suggestions to Data Sheet? Drop a line here.

NEWSWORTHY

To the skies. Alphabet’s Wing has revealed the first in its new lineup of drones, which are larger and can carry heavier objects than previous machines, The Verge reported. They will launch in cities in which Wing operates over the next 12 months. 

Amazon shopping assistant. Amazon has launched a generative AI tool that can answer customer questions about specific products, Marketplace Pulse reported. The product pulls information from the listing and reviews to answer questions within seconds.

A(i)^2 + B^2 = C^2. DeepMind, Google's AI lab, on Wednesday launched a new product that can solve geometry equations, TechCrunch reported. The problem-solving approach the AI takes to proving mathematical theorems could be useful for other AI applications, according to DeepMind. 

IN OUR FEED

"Technology may be disrupting our traditional revenue, but it offers us many new opportunities to reach audiences and deliver the kind of quality journalism they will pay for."

CNN Worldwide chief Mark Thompson said in an internal memo reported on by Axios about his strategy to build digital subscription-based products and engage with generative AI. The note follows years worth of cord-cutting, which is expected to continue, affecting revenues for the cable news network.

IN CASE YOU MISSED IT

Sam Altman admits being pushed out of OpenAI was ‘wild’ and caught him ‘off guard’—but he’s done talking about it, by Eleanor Pringle 

Microsoft CEO Satya Nadella is an AI winner. He doesn’t like to talk about the potential losers, by Jeremy Kahn

As investors obsess over cash flow, Bessemer says it’s time to rethink how late-stage companies are valued, by Allie Garfinkle 

‘Fortnite’ maker Epic Games to contest Apple’s ‘27% tax’ on outside payments and its ‘bad faith compliance’ with court ruling, by Leah Nylen, Greg Stohr, Mark Gurman, and Bloomberg

BEFORE YOU GO

Goodbye Google Flights. Google will adjust its search results in Europe to comply with the Digital Markets Act, it said in a blog post Wednesday. Under the new regulation, Google must treat rival services the same as it treats its own offerings when presenting them in search results. To comply, the company will make comparison sites more prominent in its search page and remove certain Google-specific features from the search page, including Google Flights. 

This is the web version of Data Sheet, a daily newsletter on the business of tech. Sign up to get it delivered free to your inbox.