Taco Bell, owned by Yum! Brands, began deploying AI-powered voice ordering systems at its drive-through windows starting in 2023. By mid-2025, the technology was active at more than 500 locations across the United States. The system promised to take orders faster, reduce mistakes, improve consistency, and cut labor costs. According to Taco Bell, it successfully processed two million orders. What the company did not emphasize in that statistic was how many of those orders became content for millions of social media viewers.

The Mountain Dew Loop

One Instagram clip, viewed over 21.5 million times, captured the core absurdity. A customer pulled up to the drive-thru speaker and ordered a large Mountain Dew. The AI confirmed the order and then asked: "And what will you drink with that?"

The customer, understandably confused, repeated the order. The AI acknowledged it and asked again what they would like to drink. This continued in a loop. The system could not reconcile the idea that a Mountain Dew was itself the drink. It kept slotting the Mountain Dew into some internal category of "food item" and prompting for the beverage that should accompany it.

Drive-thru AI ordering systems work by parsing spoken language into structured orders - matching items against a menu database, tracking modifications, and building a complete order for the kitchen. When the model's classification goes wrong at the category level, the conversation derails in ways that no amount of polite rephrasing from the customer can fix. The customer knows Mountain Dew is a drink. The system does not. Neither party has a mechanism to resolve the disagreement.

18,000 Cups of Water

In what became the signature prank of the Taco Bell AI experience, one customer ordered 18,000 cups of water. The AI accepted the order. It reportedly stalled mid-interaction, unable to process the absurd quantity, and staff had to intervene manually.

The incident became shorthand for a specific class of AI failure: the system had no sanity check on order quantities. A human employee hearing "18,000 cups of water" would respond with some variation of "no." The AI had no concept of "no, that doesn't make sense." It parsed the request, attempted to fulfill it, and broke.

This particular failure mode is well-known in software engineering. Input validation - rejecting values outside a reasonable range - is a basic practice. But voice ordering systems operate differently from web forms where you can set a maximum value of, say, 20 on a quantity field. The AI has to interpret natural language, map it to structured data, and decide which parts of the utterance are menu items, quantities, and modifiers. Adding hard caps on quantities means making assumptions about what constitutes a reasonable order, and in the fast-food context, large catering orders can be legitimate. The question is where to draw the line, and the Taco Bell system apparently drew no line at all.

Ordering McDonald's at Taco Bell

In another widely shared video, a customer pulled up to a Taco Bell drive-thru and started ordering McDonald's menu items. The AI did not object. It accepted the McDonald's items and, in an act of remarkable accommodation, suggested McDonald's dipping sauces to go with them. A human drive-thru worker eventually cut in and took over the order.

This failure showed something about how the model handled out-of-vocabulary items. When the AI didn't recognize a product name from its own menu, it apparently didn't reject the order or ask for clarification. Instead, it processed whatever the customer said as if it were valid, improvising responses that sounded plausible but bore no connection to the restaurant where the order would actually need to be prepared.

For the customer, it was comedy. For the kitchen staff who would have had to make sense of an order ticket for a Big Mac at Taco Bell, it was a preview of a workflow problem that scales badly. If the AI can be talked into accepting anything, the burden of catching nonsensical orders falls entirely on the humans working the line.

The Broader Pattern

Taco Bell was not the first fast-food chain to face this kind of public failure with AI ordering. In 2024, McDonald's removed its AI-powered drive-thru ordering system after customers shared videos of its own glitches: bacon added to ice cream orders, hundreds of dollars' worth of chicken nuggets appended to a simple meal. McDonald's had partnered with IBM for the technology and quietly pulled the plug after the viral embarrassment.

The pattern is consistent across both deployments. Voice AI ordering works adequately for straightforward orders in controlled conditions. It struggles with ambiguity, unusual requests, heavy accents, background noise, and - critically - customers who are deliberately testing its boundaries. The social media incentive structure guarantees that customers will test those boundaries. A video of an AI drive-thru failing spectacularly is inherently shareable in a way that a video of one working correctly never will be.

Taco Bell's Response

Dane Mathews, a Taco Bell executive, told The Wall Street Journal that the experience was causing Taco Bell to rethink its approach to AI in the drive-thru. "There are times when humans are better placed to take orders, especially when the restaurants get busy," Mathews said. He also offered a candid assessment of the technology: "Sometimes it lets me down, but sometimes it really surprises me." This is not, traditionally, the level of reliability you want from a system responsible for processing customer orders.

Taco Bell paused further expansion of the AI ordering system. The company said it would train staff to monitor the AI and intervene when needed, and would avoid relying exclusively on the technology at high-volume locations. The AI would remain at existing locations, but the plan to roll it out to the rest of the chain was shelved.

This is the same trajectory McDonald's followed a year earlier: deploy, encounter problems, get mocked online, scale back, revert to humans for the hard parts. The difference is that Taco Bell had the benefit of watching McDonald's go through the cycle first and deployed anyway.

The Economics of Not Listening

The argument for AI drive-thru ordering has always been straightforward: labor is expensive, and voice AI can take orders without a paycheck. Yum! Brands' original announcement touted easier workdays for employees, improved order accuracy, more consistent and friendly experiences, reduced wait times, and higher profit growth.

The reality exposed by the viral videos was that the technology wasn't ready. The accuracy improvement existed only for orders that fell within a narrow band of normal: standard menu items, clear pronunciation, no ambiguity, no pranksters. Outside that band, the AI was worse than a human in ways that were not just operationally problematic but actively entertaining to document and share.

Two million successful orders is a real number, and it's worth acknowledging that the AI worked for many customers who ordered their Crunchwrap Supremes without incident. But fast-food operations don't get to choose their customers. The system has to handle the full range of human behavior at a drive-thru speaker, including the teenager who thinks ordering 18,000 cups of water is the funniest thing that has ever happened. A human employee handles that interaction in about three seconds. Taco Bell's AI turned it into a system crash.

The company spent years deploying AI to 500 locations and then discovered, via Instagram, that the technology couldn't tell the difference between a drink and a side dish. The lesson from both Taco Bell and McDonald's is identical: AI voice ordering at the drive-thru is a solved problem only for customers who order exactly the way the AI expects them to, and a content creation opportunity for everyone else.