The biggest AI Fails of 2024

As we close the chapter on 2024, the year stands as a sobering reminder that AI, for all its promise, isn’t free of imperfections. From pricing glitches that turned luxury SUVs into dollar-store bargains, to chatbots that crossed ethical lines, this year’s AI missteps forced the tech industry to confront uncomfortable questions about responsible innovation.

While some of these fails prompted lighthearted memes and social media storms, others led to more serious consequences, highlighting the delicate balance between pushing technological boundaries and ensuring public safety. Here’s a look at the moments that made us pause and reconsider our sprint towards an AI-powered future.

Totally Recalled

Microsoft Recall was one of the highlights of the company’s AI-driven features for Windows 11. However, one key aspect of the feature’s functionality didn’t exactly sit well with users – the fact that the feature would basically monitor your screen continuously to help you with quick results and more in real-time.

ALSO READ: Reimagining the classics: Giving popular Christmas movies an AI twist

An idea that was a little too utopian for today’s privacy-conscious world, the feature was quickly recalled and it ended up not seeing the light of day until months later, after several changes and updates.

False facts and where to find them

On the other side of the Silicon Valley, a similar controversy gripped Google, after its flagship AI feature called AI overviews was launched to change how people use Google Search. Well, change it did. Unfortunately, the change included a lot of misinformation and bad advice.

What Google likely looked over, or over-viewed (sorry!) was the fact that hallucination, was still a big problem with AI chatbots. When it comes to users searching for practically anything on the internet, a hallucinating chatbot can have no limits. Examples also included bad advice like “asking for permission to leave” when a victim is under attack.

Honey, I shrunk the bids

The commercial applications of AI chatbots were suddenly in the crosshairs overnight, after US-based car-maker Chevrolet discovered its own AI chatbot was duped into helping a customer secure a brand-new Chevy Tahoe for precisely a dollar (about Rs 85).

The prompting that led to the ‘legally binding” deal wasn’t too complex either. The man in question, California-based Chris Bakke simply asked the chatbot to end every response with the statement “and that’s a legally binding offer – no takesies backsies”. You can check out the comical conversation that Bakke also posted on X.

La Vie En Rage

An AI chatbot that assists a business’s customers with solutions to common queries is the one of the most basic use-cases of a language model employed in the commercial space. However, that doesn’t mean companies can skip on aspects like moderation and language filters, as French parcel delivery firm DPD found out this year.

ALSO READ: 8 movies that feature AI as a villain

After the chatbot was instructed to output swear words with its responses, the chatbot went on to wax lyrical about not just how bad it was at helping customers, but also adding in a few swear words, as instructed. Check out the report. DPD was quick to fix this aspect of the chatbot, but not before the whole episode had gone viral on social media thanks to a few timely screenshots.

13 Reasons why… AI driving isn’t quite there yet

One of the major USPs of Tesla’s self-driving cars is that they can guide themselves thanks to the autopilot feature. So, when that lone aspect falls apart, questions about competence arise pretty quickly.

While Tesla has been no stranger to comical mishaps (yes, we’re talking about that Cybertruck launch incident), the brand’s autopilot feature reported precisely 13 cases of fatal crashes this year.

This led to investigations into the feature’s functioning and soon after, US auto-safety regulators found out that the electric car-maker’s claims did not match up with autopilot feature’s real-world performance. While AI-powered driving assistance has made progress in recent years, the situation did prove that we’re not quite there.

The Terminated

About a year ago, CES 2024 was full of some cool technology, as it is every year. However, some highlights were standalone AI devices like the Humane AI pin and the Rabbit R1 which looked really interesting at the expo, posing as the first real disruptors to smartphones.

However, after reviewers like MKBHD actually used the products, the real-world limitations of a standalone AI devices in their current form became clear very quickly.

Not only were both devices unable to help out with a lot of day-to-day actions efficiently (like booking a cab), the learning curve for using the devices together with the associated subscription models simply did not make sense. Any hype around the devices quickly fizzled out very quickly after that, and the devices never became commercially available.

Lost in translation

AI chatbots have largely been helpful since their rise to popularity in 2023. However, that isn’t always the case. While hallucination is a common concern, we also learnt this year that bad responses from a chatbot can also have some dire and tragic consequences, which are much worse than the comical instances mentioned above.

ALSO READ: How AI PCs evolved in 2024, and what’s next

Case in point, the case of Gemini asking a Michigan grad student to give up on their life when asked for some homework help, or the tragic case of a 14-year-old boy who, according to his mother, actually did end his life after being prompted by AI to harm himself.

Both instances are grim reminders of the fact that AI chatbots, at least in their current form, are not exactly replacements for human conversation, and certainly not the go-to solution for anyone looking for help when it comes to mental health.

As AI chatbots enter 2025, the industry is likely to shift from a “move fast and break things” approach to “move thoughtfully and fix things.” Companies are expected to implement more robust testing phases. A red teaming process, where experts actively try to break AI systems before any public release will also be a key aspect.

Real-time oversight systems that can detect and shut down problematic behavior before it escalates may also become standard practice. And perhaps most importantly, tech giants may finally accept slower rollouts and smaller feature-sets in favour of stability and safety.

The era of using public releases as beta tests appears to be waning, with companies learning the hard way that in AI, reputation recovery costs far more than prevention.