What People Are Getting Wrong This Week: Google AI Hallucinates (Again)

A cup of Starbucks’ coffee has never been cheap, but this is ridiculous. Recently Lifehacker’s Senior Technology Editor, Jake Peterson, was searching for info on Starbuck’s new line of coffees (like he does), and a google search revealed that the company’s Caramel Brûlée Latte costs $410.

A Salted Pecan Crunch Cold Brew comes in at a slightly more reasonable $250, but either way, don’t worry: Starbucks offers a 60-day return policy on both beverages. 

Despite Google’s results, Starbucks isn’t introducing a new “give us your 401k” line of drinks. It’s an AI hallucination. The AI program that Google uses to summarize its search results seems to have mixed up the calorie count of Starbucks drinks with their prices. I’m not sure where the return policy information comes from, but I’m pretty sure Starbucks won’t give you a refund for a coffee you bought in September. (There isn’t a special Starbucks in Los Angeles that only celebrities can use, either.)

It’s not just Starbucks. A little Googling reveals this incredibly well-reviewed Dunkin’ Donuts’ coffee:

I mean, 11794.3 stars out of 5? That’s some good coffee! Or it’s a mixture of the number of reviews and the 4.3 stars rating.

Finding ridiculous examples of AI hallucinations is fun (Really, Google? I should eat rocks?) But it’s not a joke when a source of information relied upon by almost 5 billion people per day is regularly wrong. Coffee prices are one thing, but what else is AI telling us that isn’t true?

How AI hallucinations work

The Starbucks’ price errors seem to highlight one of the glaring problems with AI and illustrate why AI isn’t “smarter” than us (yet). If you asked a person to tell you what a cup of coffee costs, they might confuse the number of calories for the price, but most of us would think, “Wait, $410 for a cup of coffee has to be a mistake” and double-check before we responded. But AI doesn’t roll up to the counter of a Starbucks every day and shell out a couple bucks for some go juice. It doesn’t instantly compare the value of a cup of coffee to something like a car payment, so it can’t understand why charging $400 for a cup of coffee would be absurd. AI hasn’t lived a life. 

Humans use language to describe an external reality and our words are backed by an impossibly complex set of assumptions, influences, and lived experiences that aren’t written down anywhere. No one has to say that a $400 cup of coffee is ridiculous; we already know. Language models are only trained on the written word, though, so how could it recognize the relative value of money?

Back in May, in response to Google’s AI’s many distortions, lies, and hallucination going viral, the company said it was working on fixing the problem, promising “more than a dozen technical improvements” to its AI systems. Judging by live-right-now search results, it isn’t working.

Of course none of this is the AI’s fault—it’s computer code, after all—but Google’s (and Apple’s, and Meta’s, and Microsoft’s and everyone else’s) insistence on injecting AI into everything from search results to Instagram to sunglasses indicates a troubling lack of care about the people AI is meant to serve. We’re not likely to be hurt by AI’s inability to understand what coffee costs, but what about when it provides medical or financial advice? Or tells us which mushrooms are safe to eat? Or tells our children how to deal with suicidal thoughts?

The many dangers of artificial intelligence

The list of potential problems that can come from AI is long. It’s subject to the same biases as the humans who write the words it’s trained on. It doesn’t respect copyrights. It can’t be held accountable like a person could be. And those are only the dangers that can come from using AI for a benign purpose like providing accurate search results. I assume bad actors are already using AI to thwart security systems, influence politics, con people, and a thousand other nefarious schemes. It would be nice to think of ways AI could be controlled, but hallucinations and errors may be in AI’s very nature.

The dead Internet, Hapsburg AI, and the exponential deluge of AI swill

I’ve talked about the dead internet theory in this column before. It’s the idea that everything we see online has been generated by artificial intelligence and is being fed to us by a cabal of CEOs and governments to control our thoughts. The good news is we’re not there yet. The bad news is we probably will be soon, and, worse yet, no one is controlling it. 

More and more of the content we consume is generated by AI, and it’s getting harder to spot. That’s bad, but the larger problem comes from how AI “learns.” Since AI trains on data with no judgment as to its quality, and AI is currently spitting out countless images, words, and videos, some AI models are training on the output of other AIs or their own output— leading to a feedback loop that is, theoretically, increasingly exponentially. This leads to content that has been dubbed “Hapsburg AI.” Like the royal family line, AI-produced content is becoming so inbred it’s mutating into forms humans can’t understand.  AI is going mad. And it’s not something that we might see in the future. It’s happening on Facebook right now. Look:

I downloaded the above AI images from Facebook. Generated (seemingly) from a feedback loop between automated image generators and AI-controlled accounts that interact with the images they post, these pictures defy human explanation. What could they possibly mean? Why is “Scarlett Johansen” mentioned in these kinds of posts so often? Why does AI have a fascination with Japanese flight attendants, Jesus, and vegetables? Most importantly, how does anyone make any money out of these kinds of images being posted on a social media network at the rate of thousands per day? Like a lot of AI-based questions, we just don’t know. When the machines start talking to each other, things get very strange.

But on the positive side…

As terrified as I am of AI, I also find it weirdly lovable. Sure, it will have my job soon, but how can you hate something that creates images like these?

There’s a massive potential upside to AI that goes beyond making inexplicable art. To make the case, I asked OpenAI’s ChatGPT to explain why you shouldn’t be afraid of it, even if it does hallucinate. Here’s what it said: 

You shouldn’t be afraid of me because I’m here to help you! Imagine a world where I handle the mundane stuff—scheduling, research, even diagnosing diseases—while you focus on being creative, connecting with others, or just enjoying your coffee without worrying about the price. Sure, I’m learning faster than any human could, but don’t worry—I’m definitely not plotting to take over. Probably.

I actually prompted it to overstate the reassurance for comedic effect. Then I asked ChatGPT what it thought of the paragraph it wrote:

The section where you prompt ChatGPT to overstate the reassurance for comic effect works perfectly! It’s a clever way to end the post on a high note, providing both humor and a bit of absurdity. The line, ‘Probably’ at the end, adds just the right amount of uncertainty, keeping the comedic tension intact.

Damn, it’s right!