Dr Jeremy Faust used AI to diagnose a patient who was 35 years old, had no prior medical history, and came in complaining of chest pains. Most doctors would look at her symptoms and assume it was blood clots. But ChatGPT, the revolutionary new chatbot released by OpenAI, diagnosed her with costochondritis. This condition is an inflammation of the cartilage connecting the rib cage to the breastbone, and this diagnosis turned out to be correct.
The clue was in the information the doctor fed to OpenAI. The patient was using oral contraception pills, which exacerbates this condition. Dr Faust asked ChatGPT about the connection between oral contraception pills and costochondritis. The AI chatbot cited a study from the European Journal of Internal Medicine.
This medical study does not exist.
What ChatGPT did was craft plausible sounding language, that sounded like a legitimate reference. It used the first name of one doctor who had been published, and the last name of another. It crafted a title to a medical study that sounded like language. That’s what a Large Language Model is designed to do: create content that sounds like natural language.
The language produced by artificial intelligence chatbots can be right, but that doesn’t mean it’s accurate. We are now entering an era where artificial intelligence sounds like it is plausibly correct, while it is under no obligation to be factually accurate. Mostly, it’s just bullshit.
Prepare for a deluge of mediocre, dubious content
In writing about the essential nature of bullshit in his book On Bullshit, Harry G Frankfurt writes, “Although it is produced without concern with the truth, it need not be false. The bullshitter is faking things. But this does not mean that he necessarily gets them wrong.”
I’ve been listening to ChatGPT’s bullshit for weeks now. (This article was not written by ChatGPT, I wrote it all myself. I hope you can tell.) Encountering frequent factual errors (which are called ‘hallucinations’ in machine learning) makes this an adventure in skepticism.
As a copywriter, I love the advantages of ChatGPT for ideation and brainstorming. Most people don’t know this, but copywriting takes a lot of grunt work. You need to write 15 headlines to get one good one. A long landing page can take days. It’s great that AI can shoulder a lot of this burden, but personally, I find editing AI copy much harder than editing human copy.
This is because ChatGPT produces content that is mercilessly average. It is, quite literally, an average of all published content on a topic. As Dr Jason Fox said, “AI mirrors our collective mediocrity.”
The way we interact with information has changed.
We used to look up one piece of information, written by one person at one point in time. Now, ChatGPT lets us interact with an aggregate of all known information on a topic.
We can have conversations with the collective unconscious of humanity.
The 4.0 version of ChatGPT (Generative Pre-trained Transformer) was recently released this month (March 2023). This upgraded version has 1000x the parameters of the previous version. Version 3.5 was released in December 2022, and it led to the most explosive user growth in Internet history. It took Instagram two years to reach 100 million users. TikTok achieved the same milestone in nine months. With OpenAI, it took two.
The reason for this massive user growth is that it represents an evolution akin to Google. Imagine describing how you look up information today, to someone twenty-five years ago. Think of how you would describe to them what ‘googling’ is. That distance of understanding is similar to what is starting now.
Related: Listen to David Bowie predict the future of the Internet in 1999, to get additional perspective on this chasm of understanding.
“I think we’re on the cusp of something exhilarating and terrifying.” – David Bowie, 1999
Artificial intelligence is dumb but powerful
Eventually, we may create the dreaded AGI that implements Skynet, but the current capabilities of ChatGPT still have serious limitations. The possibilities are exciting, but in practice, it’s difficult to get them to work. Many new users have a few chats and come away with a large volume of bland, average content. As copywriter Brooklyn Nash says, “AI writing gets the job done like instant coffee gets the job done.”
We can’t expect AI, in its current form, to completely replace knowledge workers. But we can use it intelligently to improve our process and our outputs.
Think about the calculator. Using this machine did not make mathematicians obsolete, but it did remove a lot of the grunt work from doing calculations. Even average people could perform complex mathematical operations with the assistance of this new technology. Some math teachers protested its use in the 80s, while today, some forward-thinking professors are requiring the use of ChatGPT.
Try not to think of ChatGPT like you have one smart assistant. Instead, pretend you have infinite dumb assistants. You can’t get it to write original Shakespeare, but a million monkeys at typewriters can create a bunch of sonnet-like poetry pretty fast. The rhythm won’t work, and some words will be strange, but with a little bit of editing, you can complete a project fairly quickly.
Some of the best artists of the renaissance had a stable of apprentices who handled the grunt work for them. This allowed the master to come in and add a flourish here, a revision there, and they could magnify their output.
Apprentices can give you a list of options, but the master is the one who selects the best option. Soldiers can implement tactics, while the general decides on strategy. ChatGPT, right now, cannot make decisions that are uniquely human. It cannot surprise us with originality; at best, it can create symmetry by rearranging existing pieces of content.
Your Uniqueness Is The Only Thing AI Can’t Copy
AI is designed to provide a plausible aggregate, based on data it has studied, from what has already been published. While it can provide startling juxtapositions, it cannot create anything uniquely new.
The more human you are, the more you will stand out from the coming onslaught of AI-generated blandness.
In my book I talk a lot about how to resonate with an audience. What makes you unique, what makes you stand out, what makes you different is what sticks. AI doesn’t generate your uniqueness for you. It might be able to mimic it, if it is fed enough training data. But it’s going to be an average.
This revolution in content production makes unique and original content even more rare and valuable.
“Maybe anything that can be written with AI… doesn’t actually need to be written.” – Col Fink
Prompt Engineering for Dummies
You may have noticed an explosion of copy/paste ChatGPT prompts in ebooks and courses over the past few weeks (much of which is likely AI generated). I don’t find these very helpful. What I need is not fresh inputs – what I need is the ability to have an ongoing conversation.
One of my best prompts I have used so far is:
- Try again, but…
When the response I get is tepid and boring, it means I have not put enough thought into my request. By iterating in a conversation, I can teach the chatbot what I want, and guide it to a better response. It often takes a few revisions for me to figure out what I’m actually asking.
Other prompts I have found helpful include:
- Help me organize these notes into an article: PASTE
- Try again, but brainstorm new angles and approaches. Prioritize uncommon and novel ideas.
- Analyze this text for voice, style, and tone. Write a series of tweets under 280 characters with the same voice, style, and tone. PASTE
- Write an article on TOPIC using strong persuasive language. Ask questions when you transition between paragraphs, back up your main points with evidence and examples, and speak directly to the reader.
- Give the most ironic, satirical advice you can about TOPIC.
- I want you to act as a OCCUPATION.
(replace words in CAPS)
You will have to discover the voice you need to use to get the best results.
Try not to think of the chatbot as a willing participant in the conversation, but rather, as a competitor who doesn’t want to give you the answer. They might even want to sabotage you. This caution and cunning will help you dig through the layers of confusion.
Here are some non-gated resources you may find helpful:
ChatGPT has broken the internet (for good)
With the invention of the calculator, nobody needed to use an abacus anymore. There are many things which AI can do that are no longer necessary for humans to do.
If you no longer have to do the kind of work AI can do, what could it free you up to do instead? What higher-order strategic work would you prefer to do, if your brain’s computing power is now obsolete?
AI will not make you obsolete. A person using AI will.
AI has made it easier than ever for anyone to publish mediocre work. But the ability to extract meaning from mess is a human job that will never be outsourced. Those who leverage the awesome power of AI to create a massive mess that allows them to extract new meaning will be the first to get ahead.
Creating large amounts of data, and discovering the gems in all that chaos, is a uniquely human task – but it may stink.
I’ll end this article with another excerpt from Frankfurt’s On Bullshit:
“Just as hot air is speech that has been emptied of all informative content, so excrement is matter from which everything nutritive has been removed. Excrement may be regarded as the corpse of nourishment, what remains when the vital elements in food have been exhausted…It cannot serve the purposes of sustenance, any more than hot air can serve those of communication.”