Featured Article

Why is AI so bad at spelling? Because image generators aren’t actually reading text

AI is seemingly unstoppable, but it can’t spell ‘burrito’

Comment

Firefly photograph of a street sign on a busy road near a billboard that says hello
Image Credits: Adobe Firefly

AIs are easily acing the SAT, defeating chess grandmasters and debugging code like it’s nothing. But put an AI up against some middle schoolers at the spelling bee, and it’ll get knocked out faster than you can say diffusion.

For all the advancements we’ve seen in AI, it still can’t spell. If you ask text-to-image generators like DALL-E to create a menu for a Mexican restaurant, you might spot some appetizing items like “taao,” “burto” and “enchida” amid a sea of other gibberish.

And while ChatGPT might be able to write your papers for you, it’s comically incompetent when you prompt it to come up with a 10-letter word without the letters “A” or “E” (it told me, “balaclava”). Meanwhile, when a friend tried to use Instagram’s AI to generate a sticker that said “new post,” it created a graphic that appeared to say something that we are not allowed to repeat on TechCrunch, a family website.

Image Credits: Microsoft Designer (DALL-E 3)

“Image generators tend to perform much better on artifacts like cars and people’s faces, and less so on smaller things like fingers and handwriting,” said Asmelash Teka Hadgu, co-founder of Lesan and a fellow at the DAIR Institute.

The underlying technology behind image and text generators are different, yet both kinds of models have similar struggles with details like spelling. Image generators generally use diffusion models, which reconstruct an image from noise. When it comes to text generators, large language models (LLMs) might seem like they’re reading and responding to your prompts like a human brain — but they’re actually using complex math to match the prompt’s pattern with one in its latent space, letting it continue the pattern with an answer.

“The diffusion models, the latest kind of algorithms used for image generation, are reconstructing a given input,” Hagdu told TechCrunch. “We can assume writings on an image are a very, very tiny part, so the image generator learns the patterns that cover more of these pixels.”

The algorithms are incentivized to recreate something that looks like what it’s seen in its training data, but it doesn’t natively know the rules that we take for granted — that “hello” is not spelled “heeelllooo,” and that human hands usually have five fingers.

“Even just last year, all these models were really bad at fingers, and that’s exactly the same problem as text,” said Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta. “They’re getting really good at it locally, so if you look at a hand with six or seven fingers on it, you could say, ‘Oh wow, that looks like a finger.’ Similarly, with the generated text, you could say, that looks like an ‘H,’ and that looks like a ‘P,’ but they’re really bad at structuring these whole things together.”

Engineers can ameliorate these issues by augmenting their data sets with training models specifically designed to teach the AI what hands should look like. But experts don’t foresee these spelling issues resolving as quickly.

Image Credits: Adobe Firefly

“You can imagine doing something similar — if we just create a whole bunch of text, they can train a model to try to recognize what is good versus bad, and that might improve things a little bit. But unfortunately, the English language is really complicated,” Guzdial told TechCrunch. And the issue becomes even more complex when you consider how many different languages the AI has to learn to work with.

Some models, like Adobe Firefly, are taught to just not generate text at all. If you input something simple like “menu at a restaurant,” or “billboard with an advertisement,” you’ll get an image of a blank paper on a dinner table, or a white billboard on the highway. But if you put enough detail in your prompt, these guardrails are easy to bypass.

“You can think about it almost like they’re playing Whac-A-Mole, like, ‘Okay a lot of people are complaining about our hands — we’ll add a new thing just addressing hands to the next model,’ and so on and so forth,” Guzdial said. “But text is a lot harder. Because of this, even ChatGPT can’t really spell.”

On Reddit, YouTube and X, a few people have uploaded videos showing how ChatGPT fails at spelling in ASCII art, an early internet art form that uses text characters to create images. In one recent video, which was called a “prompt engineering hero’s journey,” someone painstakingly tries to guide ChatGPT through creating ASCII art that says “Honda.” They succeed in the end, but not without Odyssean trials and tribulations.

oh. my. GOD.
byu/debiEszter inChatGPT

“One hypothesis I have there is that they didn’t have a lot of ASCII art in their training,” said Hagdu. “That’s the simplest explanation.”

But at the core, LLMs just don’t understand what letters are, even if they can write sonnets in seconds.

“LLMs are based on this transformer architecture, which notably is not actually reading text. What happens when you input a prompt is that it’s translated into an encoding,” Guzdial said. “When it sees the word “the,” it has this one encoding of what “the” means, but it does not know about ‘T,’ ‘H,’ ‘E.’”

That’s why when you ask ChatGPT to produce a list of eight-letter words without an “O” or an “S,” it’s incorrect about half of the time. It doesn’t actually know what an “O” or “S” is (although it could probably quote you the Wikipedia history of the letter).

Though these DALL-E images of bad restaurant menus are funny, the AI’s shortcomings are useful when it comes to identifying misinformation. When we’re trying to see if a dubious image is real or AI-generated, we can learn a lot by looking at street signs, t-shirts with text, book pages or anything where a string of random letters might betray an image’s synthetic origins. And before these models got better at making hands, a sixth (or seventh, or eighth) finger could also be a giveaway.

But, Guzdial says, if we look close enough, it’s not just fingers and spelling that AI gets wrong.

“These models are making these small, local issues all of the time — it’s just that we’re particularly well-tuned to recognize some of them,” he said.

Image Credits: Adobe Firefly

To an average person, for example, an AI-generated image of a music store could be easily believable. But someone who knows a bit about music might see the same image and notice that some of the guitars have seven strings, or that the black and white keys on a piano are spaced out incorrectly.

Though these AI models are improving at an alarming rate, these tools are still bound to encounter issues like this, which limits the capacity of the technology.

“This is concrete progress, there’s no doubt about it,” Hagdu said. “But the kind of hype that this technology is getting is just insane.”

This Week in AI: Midjourney bets it can beat the copyright police

More TechCrunch

Kobo put out a handful of new e-readers a few weeks back: color versions of the excellent Libra 2 and Clara, as well as an updated monochrome version of the…

Kobo’s new e-readers are a sidegrade most can skip (with one exception)

In an interview at his home near Reykjavík, the entrepreneur-turned-VC shared thoughts on his ventures and the journey that led him from Unity to climate tech, a homecoming of sorts.

Unity co-founder David Helgason’s next act: Gaming the climate crisis

Welcome back to TechCrunch’s Week in Review — TechCrunch’s newsletter recapping the week’s biggest news. Want it in your inbox every Saturday? Sign up here. Over the past eight years,…

Fisker collapsed under the weight of its founder’s promises

What is AI? We’ve put together this non-technical guide to give anyone a fighting chance to understand how and why today’s AI works.

WTF is AI?

President Joe Biden has vetoed H.J.Res. 109, a congressional resolution that would have overturned the Securities and Exchange Commission’s current approach to banks and crypto. Specifically, the resolution targeted the…

President Biden vetoes crypto custody bill

Featured Article

Industries may be ready for humanoid robots, but are the robots ready for them?

How large a role humanoids will play in that ecosystem is, perhaps, the biggest question on everyone’s mind at the moment.

21 hours ago
Industries may be ready for humanoid robots, but are the robots ready for them?

VCs are clamoring to invest in hot AI companies, willing to pay exorbitant share prices for coveted spots on their cap tables. Even so, most aren’t able to get into…

VCs are selling shares of hot AI companies like Anthropic and xAI to small investors in a wild SPV market

The fashion industry has a huge problem: Despite many returned items being unworn or undamaged, a lot, if not the majority, end up in the trash. An estimated 9.5 billion…

Deal Dive: How (Re)vive grew 10x last year by helping retailers recycle and sell returned items

Tumblr officially shut down “Tips,” an opt-in feature where creators could receive one-time payments from their followers.  As of today, the tipping icon has automatically disappeared from all posts and…

You can no longer use Tumblr’s tipping feature 

Generative AI improvements are increasingly being made through data curation and collection — not architectural — improvements. Big Tech has an advantage.

AI training data has a price tag that only Big Tech can afford

Keeping up with an industry as fast-moving as AI is a tall order. So until an AI can do it for you, here’s a handy roundup of recent stories in the world…

This Week in AI: Can we (and could we ever) trust OpenAI?

Jasper Health, a cancer care platform startup, laid off a substantial part of its workforce, TechCrunch has learned.

General Catalyst-backed Jasper Health lays off staff

Featured Article

Live Nation confirms Ticketmaster was hacked, says personal information stolen in data breach

Live Nation says its Ticketmaster subsidiary was hacked. A hacker claims to be selling 560 million customer records.

2 days ago
Live Nation confirms Ticketmaster was hacked, says personal information stolen in data breach

Featured Article

Inside EV startup Fisker’s collapse: how the company crumbled under its founders’ whims

An autonomous pod. A solid-state battery-powered sports car. An electric pickup truck. A convertible grand tourer EV with up to 600 miles of range. A “fully connected mobility device” for young urban innovators to be built by Foxconn and priced under $30,000. The next Popemobile. Over the past eight years, famed vehicle designer Henrik Fisker…

2 days ago
Inside EV startup Fisker’s collapse: how the company crumbled under its founders’ whims

Late Friday afternoon, a time window companies usually reserve for unflattering disclosures, AI startup Hugging Face said that its security team earlier this week detected “unauthorized access” to Spaces, Hugging…

Hugging Face says it detected ‘unauthorized access’ to its AI model hosting platform

Featured Article

Hacked, leaked, exposed: Why you should never use stalkerware apps

Using stalkerware is creepy, unethical, potentially illegal, and puts your data and that of your loved ones in danger.

2 days ago
Hacked, leaked, exposed: Why you should never use stalkerware apps

The design brief was simple: each grind and dry cycle had to be completed before breakfast. Here’s how Mill made it happen.

Mill’s redesigned food waste bin really is faster and quieter than before

Google is embarrassed about its AI Overviews, too. After a deluge of dunks and memes over the past week, which cracked on the poor quality and outright misinformation that arose…

Google admits its AI Overviews need work, but we’re all helping it beta test

Welcome to Startups Weekly — Haje‘s weekly recap of everything you can’t miss from the world of startups. Sign up here to get it in your inbox every Friday. In…

Startups Weekly: Musk raises $6B for AI and the fintech dominoes are falling

The product, which ZeroMark calls a “fire control system,” has two components: a small computer that has sensors, like lidar and electro-optical, and a motorized buttstock.

a16z-backed ZeroMark wants to give soldiers guns that don’t miss against drones

The RAW Dating App aims to shake up the dating scheme by shedding the fake, TikTok-ified, heavily filtered photos and replacing them with a more genuine, unvarnished experience. The app…

Pitch Deck Teardown: RAW Dating App’s $3M angel deck

Yes, we’re calling it “ThreadsDeck” now. At least that’s the tag many are using to describe the new user interface for Instagram’s X competitor, Threads, which resembles the column-based format…

‘ThreadsDeck’ arrived just in time for the Trump verdict

Japanese crypto exchange DMM Bitcoin confirmed on Friday that it had been the victim of a hack resulting in the theft of 4,502.9 bitcoin, or about $305 million.  According to…

Hackers steal $305M from DMM Bitcoin crypto exchange

This is not a drill! Today marks the final day to secure your early-bird tickets for TechCrunch Disrupt 2024 at a significantly reduced rate. At midnight tonight, May 31, ticket…

Disrupt 2024 early-bird prices end at midnight

Instagram is testing a way for creators to experiment with reels without committing to having them displayed on their profiles, giving the social network a possible edge over TikTok and…

Instagram tests ‘trial reels’ that don’t display to a creator’s followers

U.S. federal regulators have requested more information from Zoox, Amazon’s self-driving unit, as part of an investigation into rear-end crash risks posed by unexpected braking. The National Highway Traffic Safety…

Feds tell Zoox to send more info about autonomous vehicles suddenly braking

You thought the hottest rap battle of the summer was between Kendrick Lamar and Drake. You were wrong. It’s between Canva and an enterprise CIO. At its Canva Create event…

Canva’s rap battle is part of a long legacy of Silicon Valley cringe

Voice cloning startup ElevenLabs introduced a new tool for users to generate sound effects through prompts today after announcing the project back in February.

ElevenLabs debuts AI-powered tool to generate sound effects

We caught up with Antler founder and CEO Magnus Grimeland about the startup scene in Asia, the current tech startup trends in the region and investment approaches during the rise…

VC firm Antler’s CEO says Asia presents ‘biggest opportunity’ in the world for growth

Temu is to face Europe’s strictest rules after being designated as a “very large online platform” under the Digital Services Act (DSA).

Chinese e-commerce marketplace Temu faces stricter EU rules as a ‘very large online platform’