October 21, 2020

Can AI and GPT-3 Replace Authors? A Conversation with Liam Porr on Writer on The Side

Check out the Writer on The Side Episode Page & Show Notes

“GPT-3 is a text-completion artificial intelligence. Specifically, it’s in the category of natural language processing artificial intelligence. This is a technology developed by OpenAI which is a machine learning AI company originally started by Elon Musk now run by Sam Altman.” – Hassan Osman
- “Basically what it does is it takes a piece of text that you give it and it tries to complete that text”
- The AI is based on writing from a small subset of the internet
After publishing his first GPT-3 generated post on Substack, Liam got 26,000 unique readers in 2 weeks
- Liam never revealed the text was written by an AI and the public liked the writing on its own accord
Liam believes that while GPT-3 is attention-grabbing, it currently doesn’t have much practical use
- A far more practical text-based AI would be one that can analyze information and summarize it for the public
“I think in the near future you are going to see a lot of writers who are using GPT-3 or similar models to basically accelerate the writing process” – Hassan Oman
Liam believes GPT-3 will make it easier to become a writer, but it will also raise the bar for the quality of content that people read

Liam Porr (@liamport9) is a computer science student at the University of California, Berkeley and an online blogger
Host: Hassan Osman (@HassanO)

“GPT-3 is a text completion artificial intelligence. Specifically, it’s in the category of natural language processing artificial intelligence. This is a technology developed by OpenAI which is a machine learning AI company originally started by Elon Musk now run by Sam Altman.” – Hassan Osman
- “Basically what it does is it takes a piece of text that you give it and it tries to complete that text”
- “If you give it a question the most logical response would be to answer that question. If you gave it half of a sentence it would try to complete that sentence.”
You can’t control the output yet
- You can give GPT-3 a word limit, but the AI will just suddenly cut off rather than finishing the story within the limit
The AI is based on writing from a small subset of the internet so it’s possible GPT-3 could even start to feed itself its own information

OpenAI is limiting the release of GPT-3 to the public so Liam had to first find someone who had access
After publishing his first post on Substack, Liam got 26,000 unique readers in 2 weeks
- Liam never revealed the text was written by an AI and the public loved the writing on its own accord
In the early comments, only two readers suspected the post was written by GPT-3
- This was after GPT-3 already received media hype in 2020
- The response to these comments was that it was rude and offensive to imply someone’s writing was similar to an AI’s

Liam only wrote the title and first couple sentences for every blog post; GPT-3 wrote the rest
Liam would then create about eight different AI-generated posts from his introduction and choose the best one
- The only edits he performed were when GPT-3 generated illogical or incoherent sentences, otherwise, it was straight copy and paste
Engineering the initial human input isn’t as simple as writing whatever you want. It takes some effort to make the AI craft the desired response rather than veering on a tangent.
When sharing his story with The Guardian, the newspaper told Liam that AI writing was easier to edit than human writing

US copyright law requires a bare minimum creative contribution in order to claim ownership
- By constructing the prompt you provide a creative contribution and thus you are the copyright owner
The caveat to this is that the output must adhere to OpenAI’s terms of use
- Otherwise, you can distribute the output as you wish

OpenAI is devoting much of their time addressing how GPT-3 might be used by extremist political groups to generate support
- This might be in training the AI to give charismatic speeches espousing extremist ideologies, creating fake information sources, or by creating fake voices in an astroturf movement
- Liam believes OpenAI should dedicate more time to the actual text digestion processes
Liam believes that while GPT-3 is attention-grabbing, it currently doesn’t have much practical use
- “There’s just so much abundant information I think that the actual utility is in figuring out better ways to automatically digest text and summarize it for people” – Hassan Oman
  - An AI that can analyze masses of information would greatly streamline document-based fields like law or government and especially reduce bureaucracy
“I think in the near future you are going to see a lot of writers who are using GPT-3 or similar models to basically accelerate the writing process” – Hassan Oman
- “Specifically when you’re looking for inspiration and you’re trying to start a block of text from nothing, GPT-3 is really good at giving you something to work with”
- GPT-3 could even help with research as it will formulate writing based on all published texts on a topic
Liam believes GPT-3 will make it easier to become a writer, but it will also raise the bar for the quality of content that people read

Writing nonfiction is harder for an AI than fiction and GPT-3 isn’t even currently capable of writing a full-length fiction book
- “With nonfiction, it’s a little more difficult because now you have to add in the aspect of fact-checking” – Hassan Oman
“I think that it’s quite possible that we can get this kind of technology in the next couple of years” – Hassan Oman