April 3, 2021

1326 words 7 mins read

Machine Writing

Machine Writing

Inferring the correlation between X and Y, or writing an article in a magazine. Between these two tasks, which one can only be done by a human, and which can be done by a machine? It would seem like the answer is fairly straightforward isn’t it? Inferring a correlation between two variables is math. Of course a human could do it. But that would be tedious and time consuming. It would be easier to just let the machine do that, because it can. But writing an article? That is an art. It takes creativity and imagination to write, surely only a human would be capable of doing something like that? Oh how far technology has come.

In May 2020 researchers unveiled the latest advancement in Natural Language Processing (NLP)1. Dubbing it Generative Pre-trained Transformer 3 (GPT-3). In technical terms GPT-3 is “an autoregressive language model with 175 billion parameters”. In kinder terms, it is a machine learning model that has been pre-trained using a large dataset to “learn” how to write like a human (autoregressive language model). And It has a lot of parameters that can be finely tuned by a user to speak in a very specific way (175 billion parameters). So once given a prompt, like a question, GPT-3 is able to return a response that would sound like it is coming from a human. It has become so proficient at this that authors, and multiple reviewers, have asserted that it is almost indistinguishable from a human response.

It is worth noting though that GPT-3 is able to accomplish this feat because it has analyzed the frequency of how certain words appear next to each other. As an example, GPT-3 may “know” that the word “grapes” is frequently used with words like eat, sweet, fruit, purple, and/or wine. But it doesn’t actually fully comprehend the fact that a grape is a sweet purple fruit that we can either eat or turn into wine.

But despite these faults GPT-3 is making great inroads. And a whole crop of different applications that use it as a foundation are cropping up. I was recently made aware of one such new application called Headlime . A company that is using GPT-3 to help companies produce articles for publication.

So I thought we could have a bit of fun with this new application. The way that Headlime works is that the user is meant to give the GPT-3 model a few keywords and a prompt to get started. And then using that as input, Headlime will then produce a few model-generated paragraphs as output. What would those output look like compared to a human writing from the same prompts?

To find out, I asked a friend of mine to give me a prompt. I will then pass this prompt to the GPT-3 model to write a few paragraphs on the topic; while I will also be doing the same. We can then compare the two and see which one is better.

In the interest of fairness I wrote my paragraphs immediately after seeing the prompt, and finished within 30 minutes. I’ll then post both outputs, but I won’t say which one is which. And then we can vote on which one is the better written paragraph.

So this is the prompt (c/o my friend Paul Sedille):

Keywords: AI safety.

Description: AI safety, once an obscure branch of computer science, has grown into an essential part of the AI research world. They have their own panels in all the major AI conferences, and the world’s leading universities and research institutes have all set up dedicated teams to explore the subject, from Open AI, a non-profit research laboratory founded by Elon Musk, to Stanford’s Center for Human Aligned AI.

What concerns and developments have fueled the field’s rise to prominence?

Output A:

AI safety has gone mainstream.

The field of AI safety has come a long way since its inception in the 1980s. At first, it was just a tiny branch of computer science that most experts didn’t take seriously. Nowadays, AI safety is mainstream and becoming an essential part of the world’s leading universities and research institutes as well as many industries such as health care.

What are some key concerns or developments that have fueled this rise to prominence? More importantly, can it meet the rising challenges surrounding AI today?

Only a few years ago, concerns over biased data, unaligned machines and other AI-powered catastrophes were at best the subject of idle conversation among AI researchers. On September 28, 2018, IEEE held a symposium on AI safety and ethics at New York University in the USA. In this event, they discussed what some of the challenges are that need to be overcome for future progress in AI research and development as well as how to integrate it into our daily lives. There are many different concerns about biased algorithms with unbiased data because these biases can negatively affect humans' decisions or even restrict their abilities.

At present, many people have a lot of concerns about bias-free algorithms with unbiased data because these biases can negatively affect humans’ decisions or even restrict their abilities.

Output B:

Whether we know it or not, Artificial Intelligence (AI) has become an integral part of our day-to-day lives. Whenever you open up Facebook, Twitter, or whatever other social media platform, you are being fed content that is curated for you by an AI. When you walk through an airport, more often than not, you will be scanned by a facial recognition program that is powered by AI. When you go to ask a loan from the bank, apply for a job, or even make a Google search, you are interacting with an AI in some way, shape, or form.

So it is unsurprising that there is now increasing concern on how AI works and behaves. Since its inception, the field of AI study has mostly been focused on getting the AI-concept to “work”, to spit out an answer. But now, we are seeing more and more focus being placed on getting AI to “work well”, to spit out a good answer. Getting to that point means having algorithms that can account for the messiness of the real world. Concepts such as racial and gender bias, insufficient data, or outlier events.

Progress is being made however. Researchers are no longer satisfied with the “black-box” nature of AI systems. At every major AI conference you will see panels on AI safety. Organizations such as OpenAI and Stanford’s Center for Human Aligned AI have been created to ensure the ethical and safe implementation of AI algorithms.

Which one was written better?

The algorithm wrote A. I wrote B.

GPT-3 is being heralded as the next great leap for NLP technology. But it is not without its risks. Like any tool NLP has the potential to be misused; so much so that the authors of the original paper that introduced GPT-3 spent a considerable section of their paper talking about possible misuse of their algorithm, as well as its biases that could negatively impact different demographics2.

Moving forward we need to be conscious of how new technology like GPT-3 can impact society. We need to be aware of how tools like these can be used to cause ill. But we also need to work towards making the tool itself more equitable for all.

Technology has the potential to be a great equalizer in society, but it also has the potential to further grow inequality with misuse. So let’s make sure that it is for the former, rather than the latter.

Banner image credit to mikemacmarketing

  1. The study of getting machines to understand and produce text like humans. ↩︎

  2. The NLP model was trained using data parsed from the internet. So its predictive text is indicative of that. ↩︎