KOMETA MEDIA

AI & Business6 min read

What AI Mishaps Have Cost the Big Players

From $100 billion in market value to a cancelled pilot: six documented AI mishaps and what they teach about oversight and testing.

A glass bar chart caving in, a line plunging

Talk about AI mishaps and you quickly end up with headlines and schadenfreude. That helps no one. The more useful question is what well-documented cases can teach us: what exactly happened, what did it cost, and where would a simple check have helped? We have collected six cases from 2021 to 2025, each backed by company statements or major media. Where figures are uncertain, we leave them out.

01

Alphabet 2023: One wrong sentence in an ad

In February 2023, Google published a promotional GIF for its chatbot Bard. It claimed that the James Webb Space Telescope took the first pictures of a planet outside our solar system. That was wrong: according to NASA, the first such images were taken in 2004 by the European Southern Observatory's Very Large Telescope. Reuters pointed out the error [1].

The market reaction was clear. According to Reuters, Alphabet lost about $100 billion in market value that Wednesday, and the stock was down by as much as 9 percent during trading [1]. A Google spokesperson said the incident highlighted the importance of a rigorous testing process and announced a program for external testers [1]. A note on interpretation: a share price never hangs on a single event. Still, the case shows how expensive an unchecked sentence in a public presentation can become.

about $100 billion

market value Alphabet lost, according to Reuters, on the day of the Bard error (8 February 2023) Source [1]
A cracked glass house in front of a falling line of light
02

Zillow 2021: When forecasts are backed with equity

Zillow Offers bought homes directly in order to resell them. The business depended on predicting future selling prices well. On 2 November 2021, the board decided to wind the business down, citing among other things the unpredictability of home prices and capacity constraints [2].

In the third quarter of 2021, Zillow recorded an inventory write-down of $304.4 million because it had bought homes at prices above their estimated later selling price, net of selling costs. For homes already under contract the company expected further charges of roughly $240 to $265 million, plus $175 to $230 million in pre-tax wind-down costs. The workforce was to shrink by about 25 percent [2]. The lesson: a forecast is not a certainty, and the more money depends on it, the more rigorously it must be checked against reality.

$304.4 million

inventory write-down at Zillow Offers in the third quarter of 2021 Source [2]

about 25%

planned workforce reduction at Zillow as part of winding down Zillow Offers Source [2]
03

McDonald's 2024: Two years of testing, no rollout

McDonald's had been testing automated drive-thru order taking with IBM since October 2021, most recently at about 100 restaurants. In June 2024 it became known that the collaboration on this project would not continue beyond the year. The company's stated goal was order accuracy above 95 percent before any broader deployment. A 2022 analyst report put accuracy in the low 80-percent range, and individual order errors spread on social media [3].

This is how a pilot is supposed to work: a clear target, limited scope, and the willingness to stop if the target is missed. McDonald's stressed that IBM remains a partner and said it would decide on a future voice-ordering solution by the end of the year [3]. No cost figure was given.

A cracked smartphone showing two glowing chat bubbles
04

Air Canada 2024: The chatbot is part of the website

A customer asked Air Canada's chatbot about bereavement fares after his grandmother died. The bot said he could apply for the reduced fare after travel. The airline's policy said otherwise [4]. The Civil Resolution Tribunal of British Columbia ruled in the customer's favor in February 2024. Air Canada had argued, among other things, that it could not be held responsible for information provided by its chatbots; the tribunal did not accept that [4].

It should be obvious to Air Canada that it is responsible for all the information on its website.

Tribunal member Christopher Rivers in Moffatt v. Air Canada, as quoted by Gizmodo [4]

The sum was small: $650.88 in damages, $36.14 in interest and $125 in fees, $812.02 in total [5]. The real price was public attention. For businesses the point is simple: customers and the public will treat what a chatbot says on your site as a statement by your company.

05

Klarna 2025: When cost is the main criterion

In February 2024, Klarna said its AI assistant did the work of 700 customer service agents and had handled 75 percent of chats in its first month, about 2.3 million conversations [6]. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg that the AI-focused strategy had not been the right path: chatbots were cheaper but delivered lower quality. Klarna is hiring human support staff again so that customers can always reach a person if they want one [6].

Klarna continues to use AI, just as part of a mixed model. The case is not a defeat for the technology but an example of a metric that cannot stand alone: saved cost says nothing about whether customers are served well. No damage figure is documented, so we do not give one.

06

Apple 2025: Summaries under someone else's logo

Apple Intelligence automatically summarized news push notifications. In December 2024 the BBC complained because a summary under its logo falsely claimed that the suspect in the killing of UnitedHealthcare CEO Brian Thompson had shot himself [7]. On 17 January 2025, Apple said notification summaries for news and entertainment would be temporarily unavailable, with the feature to return in a later update [7].

The damage here was to trust, not a quantifiable sum. Especially sensitive: the errors appeared under the name of a third party that had not reviewed them.

07

What the cases have in common

In none of these cases was the technology the only trigger. Each time something ordinary was missing: a fact check before publication, a measurable stop criterion, a person who verifies an answer, or an honest question about whether the benefit is as large as the slides promised. Four patterns recur:

  • Unchecked output goes straight to the public (Alphabet, Apple).
  • A forecast is treated as certainty although a lot of money depends on it (Zillow).
  • The pilot has no target to be measured against, or the target is ignored (McDonald's).
  • The human is taken out of the loop before quality and accountability are settled (Air Canada, Klarna).

None of the cases argues against using AI. They argue for treating it like any other tool with consequences: test it, limit it, monitor it and judge it by its effect.

Sources (7)
  1. [1]After Google AI chatbot ‘Bard’ blunder, Alphabet shares lose $100 billion (Reuters report)Gulf News · February 9, 2023
  2. [2]Zillow Group, Form 8-K (decision to wind down Zillow Offers)U.S. Securities and Exchange Commission (SEC) · November 2, 2021
  3. [3]McDonald’s is ending its AI drive-thru test with IBMNation’s Restaurant News · June 17, 2024
  4. [4]Air Canada Chatbot Lied About Bereavement Discounts, Now Has to PayGizmodo · February 16, 2024
  5. [5]Company is liable for chatbot miscommunications, holds Canadian tribunalDeneys Reitz · February 27, 2024
  6. [6]Klarna Is Hiring Customer Service Agents After AI Couldn’t Cut It on Calls, According to the Company’s CEOEntrepreneur (citing Bloomberg) · May 9, 2025
  7. [7]Apple suspends AI news after BBC complaintsAdvanced Television · January 17, 2025

How much of this is in your day-to-day?

In a free initial consultation we look together at what AI can already do in your business today – and where it has no place.

Book an initial consultation

More posts