How four twenty-somethings in Ohio built a neural network that outperforms offshore content mills.
Nick Shah, Paul DeMott, and Endel Maricq. Image courtesy of Paul DeMott
It's not only bud tenders and Uber drivers who are in danger of losing their jobs to automation these days. A small company in Ohio called AI Writer has created a neural net that can churn out the search engine optimization (SEO) filler currently being created by content mills in places like India and the Philippines. And they managed to do it straight out of college on a shoestring budget.
AI Writer is a neural net (an AI architecture modeled after the human brain) that is not only capable of teaching itself how to write its own unique internet marketing articles, but those articles are polished enough to fool companies interested in buying that content into thinking it was generated by a human.
AI Writer was co-founded by Paul DeMott, who had created a small internet marketing company as a summer job in college and Nick Shah, who had studied neurology for about a year and a half at medical school before dropping out to teach himself about artificial intelligence.
"We saw a demand in the internet marketing sphere for content and that was initially how we got into AI," DeMott told Motherboard. "So we said let's put five months in and see what can happen. At first we were super pumped when [the neural net] would just spell a word right. Then it started linking a few words together and at some point it started writing coherent paragraphs. We were just blown away."
When the AI Writer neural network was born, it knew absolutely nothing about the world or English grammar—all it was programmed to do was crawl the web and devour content. Yet after about four or five days of consuming hundreds of thousands of pieces of web content, the neural net had such a command over English that it was able to generate its own grammatically correct articles.
So far, the results have been promising. Each of the articles produced by the AI is checked against a database for uniqueness and according to DeMott, the net has yet to write the same sentence twice. Moreover, the AI Writer nets are creating upwards of 2000 internet marketing articles a week for clients, all of whom are unaware that the articles they're buying are being generated by an AI instead of a human.
It only gets tricky when one of their customers asks for the editor who worked on their article.
"We say we have a team of writers, they just don't know that this team of writers is a ton of computers," said DeMott. "All the computers have names, so we can say who wrote that particular article. We're not trying to hide the fact that we use AI because at the end of the day it's all about whether the customer is happy with the content."
When you read some of the samples produced by AI Writer, the coherence of the individual sentences is impressive, but taken as a whole, the articles read a little bit like a word salad. By way of example, here is a selection from an article that AI Writer wrote about the diet pill, Phentermine:
"It is recommended to take a regular amount of exercise before the initial stage of the weight loss process. The main goal of this exercise is to reduce the risk of heart attacks and stroke. It is also recommended that people with heart disease should consult a doctor before beginning any type of exercise program. Exercise is one of the most important things to do in the prevention of high blood pressure."
In the context of the whole article about Phentermine, this paragraph feels like a bit of a non-sequitur, albeit a grammatically and factually correct one. Yet for the purposes of AI Writer's clients, this is generally all that is needed.
Many of these clients are what DeMott refers to as "SEO-ers" which are essentially building fictitious websites on expired, once-popular domains that link to their 'true' website. The goal is to trick Google's search algorithms into thinking that a bunch of legitimate websites are pointing to the client's real website, which would suggest the real site is important or highly trafficked and should be higher up in the search results.
Once upon a time, gaming Google's search engine would just require pumping websites full of targeted keywords that would get picked up by the algorithm. Google has gotten increasingly savvy to these tricks however, and employs tricks like latent semantic indexing which essentially gauges how 'real' the content on a website seems by mathematically analyzing the relationships between the words on the site. If you just fill a website with the same keyword over and over, Google's algorithm will know it's a fake.
That means that fake SEO websites now need to pay up for realistic content to fill their pages. Most of this content would seem trivial or annoying to a human reader, but that's okay because these SEO-ers aren't making these fake websites for human eyes. As long as the content is realistic enough to trick Google, that's good enough.
This requires a massive amount of content, the production of which is generally outsourced to content farms in places like India or the Philippines, where labor is cheaper but there is still large English speaking populations. While AI Writer's articles might not be perfect, they are comparable to what is being generated by humans in these content factories meaning clients buying the articles wouldn't have a reason to suspect that the articles they are buying were generated by a machine.
"Our content is less likely to make grammatical and spelling errors, but outsourced work tends to be stronger conceptually," said DeMott. "In terms of actually threading together something logical and making a strong point, AI Writer is at a high school level."
The idea of a machine pumping the internet full of articles that were never intended for human eyes may sound nefarious, but for DeMott and Shah, this was simply the next step in the general trend toward automation.
"This is happening regardless of whether we build this network or not," Shah said. "There are people that are hiring thousands of Indians and Filipinos to get these articles done, so if we had not brought our product to market I don't think the situation would've changed that much."