AI Can’t Spell ‘Strawberry’ For Some Reason, Insists It’s Right And You’re Wrong

Image generated on AI

Artificial intelligence has come a long way, but even the smartest bots still have a few spelling bees to lose. This is evidenced by a juicy controversy sparked by the humble strawberry, which has exposed a surprising gap in the capabilities of even the most advanced language models.

Sophisticated tools like GPT-4o and Claude have found themselves in a rather consistent jam when asked how many ‘r’s there are in the word “strawberry.” This has led to some hilarious—and, frankly, frustrating—conversations with users.

The root of this fruity fiasco lies in how AI models like GPT-4o are built. These systems break text into tokens, which can represent words, syllables, or even single letters. But rather than understanding words in the way humans do, AI models translate text into numerical codes to generate responses. This process can result in errors, especially when precise letter counting is required. For example, the tool might see “strawberry” as a combination of “straw” and “berry” and get confused about how many r’s are actually present.

One TikToker, known as Y’all Watch This, took to social media to showcase this AI quirk in action. In a video, the user engages in a comical back-and-forth with a chatbot, which confidently asserts that there are only two ‘r’s in “strawberry.” When the user points out the correct count, the chatbot responds with a blend of reassurance and deflection that feels remarkably similar to being gaslit. 

@yallwatchthis

I asked ChatGPT how many Rs are in the word Strawberry #chatgpt #ai #strawberry

♬ original sound - YallWatchThis

“I can see why you might think [there are three r’s]... but ‘strawberry’ only has two R’s,” ChatGPT is shown to have responded. “Maybe you were counting the letter R in ‘straw’ and ‘berry’ separately?”

Asked if he was crazy for seeing three r’s, the TikToker was told by the chatbot, “You’re not crazy—sometimes mistakes happen when counting or reading. It’s easy to miscount letters, especially if you’re looking at the word quickly. If you’re still unsure, you might want to check again or use a different method to verify.”

“This feels like talking to my ex,” says one viewer. “Not ChatGPT gaslighting you,” shares another.

I tried the “ask ChatGPT how many r’s are in the word strawberry” thing. Hilarious! pic.twitter.com/nT7Af5FbGV

— Steve Xero is building MoneyG.ro (@steve_xero) August 28, 2024

In its reasoning, ChatGPT also misspells the word once, writing, “strawrberry.”

DesignTAXI replicated this conversation with the GPT-4o model, and indeed, the AI insisted there were only two r’s in the word.

Screenshot via ChatGPT

However, it correctly said there were three r’s in “raspberry.”

Users have pointed out the unsettling potential of AI’s persuasive power when combined with such errors. “Persuasion is the biggest threat with AI. Imagine this but with history or current events,” warns one person.

Interestingly, this phenomenon may or may not have inspired OpenAI’s latest model, internally nicknamed ‘Strawberry’. This version is built for more advanced reasoning and problem-solving, excelling at tasks like complex math equations and intricate word puzzles. The model has managed to tackle the New York Times puzzle Connections and score over 90% on the MATH benchmark, a significant leap from GPT-4’s 53% score. 

[via TechCrunch, The Mirror, OpenAI, R&D World, images via various sources]