The secret story behind the army of people —yes, people— that make AI feel as magic as it does

Covering culture, tech, luxury, travel, and media in 5 minutes, twice a week


Bellustar Tokyo is a 97 key (room) hotel that spans the top 9 floors of Tokyu Kabukicho Tower, a new skyscraper in Shinjuku (Jonas Bjerre-Poulsen / Wallpaper)

🔼 The price of power in Texas. It’s up 100% amidst a massive heat wave »»

🔽 Grocery delivery options in Paris. Getir’s leaving France »»

💬 “I immediately binned the show after that.” The opening credits to Secret Invasion, a new Marvel tv show, are AI generated. Audiences noticed. And they hate it »»

🛫 Tokyo’s Bellustar Hotel combines minimalist interiors with great views in Shinjuku »» Also: Europe’s 25 most beautiful places »»

👗 It's Pharrell's world, we all just live in it »»

💎 Inside Qatar Airways secret private jet fleet »»


Nike’s new Calm Slide will debut in 5 monochromatic colorways, all suited for summertime (Nike / Hypebeast)

HBO shows are coming to Netflix »»

11 solo founders doing millions in revenue —with zero employees »»

The Bank of England is set to raise interest rates to 6%, a 23-year high »»

A wild new app called ChatDox promises to let you chat “with” any YouTube video you give it »»

Rampant groundwater pumping has changed the tilt of Earth’s axis »»

Get to know MediaLink, the ubiquitous marketing services consultancy owned by Hollywood talent agency UTA that helps its corporate clients review agencies, place execs, hire talent, parse data, employ new tech, market themselves, do DD, and more. They’ve hosted some 700 meetings at Cannes Lions this week »»

Bartender secrets to help you master a frozen Margarita »»

Hilton announced its first ever hotel in Ghana »»

A Tesla software hacker found a hidden Elon mode that allows Teslas with FSD (Full Self-Driving) to operate without any driver monitoring »»

Cars track drivers nowadays »»

Nike’s launching a chill new Summer slide »»

The world’s top 11 steakhouses »»

A six-month-old CBS News report on a Titanic tourism submersible is going viral. The report raised safety concerns about the now-missing vessel »»

Bitcoin broke US$30,000 for the second time this year. People are excited because traditional finance (TradFi) players are moving into the space »»

After shutting its massively popular London Fields food stall last year, Sonora Taqueria is finally opening up a restaurant, in Stoke Newington »»


The secret story behind the army of people —yes, people— that make AI feel as magic as it does

Generative AI needs continuous human input to avoid “model collapse.” So AI firms are paying human beings to chat with their bots about anything at all. (Dall-E)


Things move fast in AI.

Last issue, I shared an intriguing scientific theory called “model collapse.” To summarise it (read the full piece here), all of the AI-generated content flooding the Internet right now is poised to start poisoning the source data that future generations of ChatGPT (and other systems like it) will be trained on.

That’s important because if the source data’s polluted, ChatGPT and bots like it will stop returning valuable (or even comprehensible) text.

Bottom line: enjoy ChatGPT while you can. It’s answers are about to turn into goop.


Stay with me as I break down how exactly model collapse won't happen, why human written words are about to get even more valuable (though not in the way that you might think), and the secret story behind the army of people —yes, people— that make AI feel as magic as it does.


Okay, first things first, let's go back to how AI “learns.” Take image generating AI systems, like DALL-E or MidJourney, for starters.

Most people know that image generating AI systems are only able to create new pictures because that system has been trained on countless existing photographs.

So far, so good.

But, wait: did you know that AI firms actually employ human beings —thousands of them, in developing nations as well as in rich economies— to tag these photographs, and manually identify the elements therein, so that AI systems can distinguish between a door and a desk, or a knee and an elbow?

Yup, in order to train AI on those existing photographs, an army of human beings had to click through thousands of images, and literally circle and label every fish, bicycle, t-shirt, elbow, and much more, so the systems can know what’s what.

Without those humans doing that work, AI wouldn't be able to do its thing.


Identifying elements and adding annotations to images doesn’t appear to be as automatable as AI firms would like.

In other words, AI firms are having to employ human workers to ID and tag images for way longer than anyone thought.

Here's a concrete, yet intriguing, example of why that is.

Imagine a photograph of a shirt on a hanger, in front of a mirror. The pic shows a shirt, and its reflection. So far, so clear.

Well, you —clever human that you are— know the difference between the shirt on the hanger and the reflection of said shirt in the mirror.

But, now think about all of the things that AI needs to know in order to recognize that there's only one shirt, and not two, in this image:

The mirror has to be identified and distinguished from for example, a window. Then, the concept of reflection —something that we humans are able to get our heads around before we even have the power of speech— has to also somehow be imparted upon the AI system.

This simple image, that our brains can immediately get, actually presents a really complex challenge to AI.


Words aren’t necessarily easier. It can be hard for us humans to recognize emotion in the written word, as anybody who's stared way too long at a short text message from a friend trying to decipher what it means can attest.

Now take a moment to consider how much more difficult that is for a computational system.

Is this sentence sincere or sarcastic? They can’t answer that without the data.

They need the data.


This newsletter has written about how data is the business model for the next 100 years. Data is already more valuable than oil, diamonds, gold, and anything else you can think of.

Data is the most precious resource we have. And don't just take my word for it. The Economist magazine has been saying so since 2018.


Don’t forget, data take many forms.

We tend to think of data as numbers on spreadsheets, but anybody who works in market research knows that data can also be verbatims. (“Verbatims” is just fancy research speak for “direct quotes from people.”)

The most valuable verbatims for researchers are often very human expressions of frustration, appreciation, enjoyment. Guilt. Even fear.

That stuff is also data.

And that's the data that AI companies are desperate for.


Did you know that some AI firms employ human beings to literally just chat all day long about whatever’s on their mind with a chatbot?

Literally, that’s the whole job: just chat about whatever you can think of.

This is AI firms’ logical —maybe even predictable— answer to “model collapse.”


Humans will continue to feed data into the system by tagging photos, chatting with it behind the scenes by sharing the mundane, as well as spilling the serious.

Look for AI to not only avoid model collapse, but actually get closer and closer to that tantalising approximation of intelligence that many seek from it.

And human written words are going to be more valuable than we think, at least in the near and medium term, though not necessarily as original works of creation, to be enjoyed far and wide. Instead, human written words are going to find their value as as fresher and fresher “organic” data for AI firms who will use them to create an artificial future for all written communications.

What a time to be alive.

More: AI is a lot of work »»

Written by Jon Kallus. Any feedback? Simply reply. Like this? Share it!

Join the conversation

or to participate.