Human or Machine: Who Wrote This?
By Aditi Desai
I started writing an email to my Chemistry professor. After typing ‘H,’ the word “Hello!” appeared next to my blinking cursor. I shook my head, realizing that I should probably start the email with “Dear” instead. I typed ‘D’ and took a small pause to see if the computer would once again prompt me with its guess. This time, the computer didn’t match the exact word I had in mind — it did more. In a faded print, adjacent to the ‘D’ I had typed, the sentence “Dear Professor X, Good Morning! I hope you had a great weekend!” was displayed on the screen. The computer had done more than just guess my next word; it had formed a considerate, compassionate sentence. At this point, hitting the Tab key to finish my thoughts and wrap my sentences up was almost an involuntary reaction. After all, the computer produced words which not only finished my sentence, but also mimicked the tone of my own writing.
For a few days, I had been testing my willpower to ignore my computer’s hints and compose my own sentences — anything to not succumb to the automated suggestions. These suggestions, which some refer to as short-cuts, are fairly new, put forth by Gmail’s Smart Compose and Smart Reply features in May 2018 (Wu, 2018). At their core, both serve to assist users when composing email responses. Smart Reply contains a list of already written messages — most of which are concise and generic — which users can choose fromwhen responding to an incoming email. The idea is that instead of having to compose a longer response from scratch, Gmail users can select a Google-generated response (“Sounds great”, “Very cool!, “Would love to talk later!”, “I completely understand.”), saving both time and effort. Contrary to this, Smart Compose prompts users with words, short sentences, and phrase suggestions as the user types in real time.For example, as you write “Thank you again,” the words “I truly appreciate your efforts” may show up next to your cursor. The assumptions made by Smart Compose are not entirely genetic, either. In fact, the assumptions are more personalized since the software aims to replicate users’ voice, tone, and language based on previous emails and messages that users have written.
What ties Smart Compose, Smart Reply, and other predictive language technologies together is their reliance on Artificial Intelligence (AI). Such platforms are trained to perform specific tasks, but are also flexible in nature, constantly adapting based on the specific needs and language of the user. Up until now, many individuals have relied on spell and grammar-check when texting, emailing, or typing long-form text. However, predictive language technologies are a leap forward from spell-check; Smart Compose and Smart Reply are not simply correcting words, they are composing the words and crafting the whole message, almost entirely on their own by utilizing the power of deep learning, a subset of machine learning.
Earlier this year, OpenAI, an artificial intelligence research lab based in San Francisco, built a software that is refreshing, spectacular, and daunting all at the same time (Metz, 2019). This software, known as GPT-3, is quite a few steps up from Google’s Smart Reply and Smart Compose in that it is the “most powerful” language model to be created (Metz, 2020). Although based on a similar technology, GPT-3 is trained on an overwhelmingly large database of text — ranging from articles and blog posts to complete novels. With enough text and enough time spent processing, GPT- 3 is able to dissect the connection between words, phrases, and sentences. Equipped with this knowledge, it does what most humans do when we read storybooks as children: learn to read and write. Such skills, though difficult to acquire accurately, can prove to be immensely beneficial. In fact, GPT-3 enables machines (think: digital assistants, automatic responding software, and video game characters) to understand, process, and respond to humans in a way that resonates with us. Instead of responding with otherwise abstract lines of code which only programmers are able to decipher, machines can convey phrases and complete messages in an accessible manner.
GPT-3’s artificial intelligence is based on what many researchers call a neural network, or an organized computerized system which is loosely based on the neurons in the brain (Ray, 2020). And, like neurons in humans,which learn to associate words and images to particular meanings, GPT-3 also creates its own “mind map” and mathematical representations of patterns from the sea of digital prose, articles, and texts which it is trained on. In some ways, GPT-3 is more programmatic and less spontaneous than humans can be. For example, GPT-3 can calculate how likely one word is to appear in a text given other words in the sentence. In the sentence: I wanted to make brownies, so I went to the pantry and took out some, the black space can be filled by a variety of words (think: sugar, flour, baking pan, oil spray). However, the word “flour” is more likely to score higher than “oil spray,” indicating to the software that one word choice may be a more likely or better fit than another. Strictly statistically speaking, the term “flour” is indeed assumed to fit a sentence concerning brownies; however, human communication is not always predictable. In other words, there does exist a chance that an individual is looking for “oil spray” rather than “flour.” This is precisely where GPT-3’s limitations emerge—GPT-3 is unable to predict the spontaneity or unpredictability of human discourse. Of course, such a barrier is difficult to overcome.
However, in order to strengthen predictability and accuracy of human communication, GPT-3 must continuously be fed samples of texts. Words within these texts are converted to numeric vectors and expressions through a sophisticated process known as compression. Then, the model will work on training and accuracy, making sure that predictions score high conditional probabilities. Ultimately, the more information, words, texts, and images GPT-3 is exposed to, the more closely it resembles specific dimensions of human thought processes—giving GPT-3 the ability to conduct creative activities like writing poetry or responding to human questions with emotional intelligence.
Technologies rooted in simulating the human brain are intensely fascinating; yet, their advancement and striking sophistication raise questions about how we will be able to tell machines apart from humans. If computers are able to write for humans, will we lose compassion and honesty? Or, will we train computers to incorporate all aspects of human communication into their language? If computers are trained using texts found online, do they inherently incorporate biases reflected in society? By “reading” texts that are written by humans and thus inherently racially, sexually, and culturally biased, GPT-3 learns to “write” like us. Tucked away in GPT-3 exist problematic algorithms and vocabulary which perpetuate toxic biases. Put simply, we teach artificial intelligence systems everything—including our very own biases. Cultural attitudes embedded within digital information systems are woven into new technologies, such as GPT-3. Such pressing ethical barriers lean into the complications that arise with machine learning and artificial intelligence. OpenAI has dodged complications by explicitly specifying whether a text was written by a computer and has also invited external researchers to devise ways to reduce bias. Moreover, researchers are trying to dissect algorithmic biases which occur when GPT-3 generates overtly sexist or racist texts. In areas where biases have been baked into artificial intelligence systems, OpenAI’s review team admits that “the company is not quite where it should be.” But, as noted by Sandhini Agarwal, an AI policy researcher at OpenAI, the company will not “broadly expand access to GPT-3 until it’s comfortable that it has a handle on such ethical issues. If [the company] opens it up to the world now, it could end really badly.” While OpenAI is certainly attempting to generate a “detoxified,” compassionate, and empathetic algorithmic system, it is not clear how the company will alleviate the risk of toxic language and bias. The current way of giving machines language—of training models to think and write like humans—must change. We must pay attention to who codes, how we code, and whether training sets reflect humanity’s diversity and sense of ethics.
About the Author
Aditi Desai is a first-year student at Princeton University studying Neuroscience.
References
Heaven, W. D. (2020, December 10). OpenAI's new language generator GPT-3 is shockingly good-and completely mindless. MIT Technology Review. Retrieved from https://www.technologyreview.com/2020/07/20/1005454/openai-machine-learning-language-
generator-gpt-3-nlp/
Metz, C. (2020, November 24). Meet GPT-3. It has learned to code (and blog and argue). The New York Times. Retrieved from https://www.nytimes.com/2020/11/24/science/artificial-intelligence-ai-gpt3.html
Metz, C. (2019, July 22). With $1 billion from Microsoft, an A.I. lab wants to mimic the brain. The New York Times. Retrieved from https://www.nytimes.com/2019/07/22/technology/open-ai-microsoft.html
Ray, T. (2020, August 25). What is GPT-3? Everything your business needs to know about OpenAI's
breakthrough AI language program. ZDNet. Retrieved from https://www.zdnet.com/article/what-is-gpt-3-everything-business-needs-to-know-about-openais- breakthrough-ai-language-program/#:~:text=is%20GPT%2D3%3F-,Everything%20your%20
business%20needs%20to%20know%20about%20OpenAI's%20breakthrou gh%20AI,like%20a%20person%20wrote%20them
Wu, Y. (2018, May 16). Smart compose: Using neural networks to help write emails. Google AI Blog. Retrieved from https://ai.googleblog.com/2018/05/smart-compose-using- neural-networks-to.html
For a few days, I had been testing my willpower to ignore my computer’s hints and compose my own sentences — anything to not succumb to the automated suggestions. These suggestions, which some refer to as short-cuts, are fairly new, put forth by Gmail’s Smart Compose and Smart Reply features in May 2018 (Wu, 2018). At their core, both serve to assist users when composing email responses. Smart Reply contains a list of already written messages — most of which are concise and generic — which users can choose fromwhen responding to an incoming email. The idea is that instead of having to compose a longer response from scratch, Gmail users can select a Google-generated response (“Sounds great”, “Very cool!, “Would love to talk later!”, “I completely understand.”), saving both time and effort. Contrary to this, Smart Compose prompts users with words, short sentences, and phrase suggestions as the user types in real time.For example, as you write “Thank you again,” the words “I truly appreciate your efforts” may show up next to your cursor. The assumptions made by Smart Compose are not entirely genetic, either. In fact, the assumptions are more personalized since the software aims to replicate users’ voice, tone, and language based on previous emails and messages that users have written.
What ties Smart Compose, Smart Reply, and other predictive language technologies together is their reliance on Artificial Intelligence (AI). Such platforms are trained to perform specific tasks, but are also flexible in nature, constantly adapting based on the specific needs and language of the user. Up until now, many individuals have relied on spell and grammar-check when texting, emailing, or typing long-form text. However, predictive language technologies are a leap forward from spell-check; Smart Compose and Smart Reply are not simply correcting words, they are composing the words and crafting the whole message, almost entirely on their own by utilizing the power of deep learning, a subset of machine learning.
Earlier this year, OpenAI, an artificial intelligence research lab based in San Francisco, built a software that is refreshing, spectacular, and daunting all at the same time (Metz, 2019). This software, known as GPT-3, is quite a few steps up from Google’s Smart Reply and Smart Compose in that it is the “most powerful” language model to be created (Metz, 2020). Although based on a similar technology, GPT-3 is trained on an overwhelmingly large database of text — ranging from articles and blog posts to complete novels. With enough text and enough time spent processing, GPT- 3 is able to dissect the connection between words, phrases, and sentences. Equipped with this knowledge, it does what most humans do when we read storybooks as children: learn to read and write. Such skills, though difficult to acquire accurately, can prove to be immensely beneficial. In fact, GPT-3 enables machines (think: digital assistants, automatic responding software, and video game characters) to understand, process, and respond to humans in a way that resonates with us. Instead of responding with otherwise abstract lines of code which only programmers are able to decipher, machines can convey phrases and complete messages in an accessible manner.
GPT-3’s artificial intelligence is based on what many researchers call a neural network, or an organized computerized system which is loosely based on the neurons in the brain (Ray, 2020). And, like neurons in humans,which learn to associate words and images to particular meanings, GPT-3 also creates its own “mind map” and mathematical representations of patterns from the sea of digital prose, articles, and texts which it is trained on. In some ways, GPT-3 is more programmatic and less spontaneous than humans can be. For example, GPT-3 can calculate how likely one word is to appear in a text given other words in the sentence. In the sentence: I wanted to make brownies, so I went to the pantry and took out some, the black space can be filled by a variety of words (think: sugar, flour, baking pan, oil spray). However, the word “flour” is more likely to score higher than “oil spray,” indicating to the software that one word choice may be a more likely or better fit than another. Strictly statistically speaking, the term “flour” is indeed assumed to fit a sentence concerning brownies; however, human communication is not always predictable. In other words, there does exist a chance that an individual is looking for “oil spray” rather than “flour.” This is precisely where GPT-3’s limitations emerge—GPT-3 is unable to predict the spontaneity or unpredictability of human discourse. Of course, such a barrier is difficult to overcome.
However, in order to strengthen predictability and accuracy of human communication, GPT-3 must continuously be fed samples of texts. Words within these texts are converted to numeric vectors and expressions through a sophisticated process known as compression. Then, the model will work on training and accuracy, making sure that predictions score high conditional probabilities. Ultimately, the more information, words, texts, and images GPT-3 is exposed to, the more closely it resembles specific dimensions of human thought processes—giving GPT-3 the ability to conduct creative activities like writing poetry or responding to human questions with emotional intelligence.
Technologies rooted in simulating the human brain are intensely fascinating; yet, their advancement and striking sophistication raise questions about how we will be able to tell machines apart from humans. If computers are able to write for humans, will we lose compassion and honesty? Or, will we train computers to incorporate all aspects of human communication into their language? If computers are trained using texts found online, do they inherently incorporate biases reflected in society? By “reading” texts that are written by humans and thus inherently racially, sexually, and culturally biased, GPT-3 learns to “write” like us. Tucked away in GPT-3 exist problematic algorithms and vocabulary which perpetuate toxic biases. Put simply, we teach artificial intelligence systems everything—including our very own biases. Cultural attitudes embedded within digital information systems are woven into new technologies, such as GPT-3. Such pressing ethical barriers lean into the complications that arise with machine learning and artificial intelligence. OpenAI has dodged complications by explicitly specifying whether a text was written by a computer and has also invited external researchers to devise ways to reduce bias. Moreover, researchers are trying to dissect algorithmic biases which occur when GPT-3 generates overtly sexist or racist texts. In areas where biases have been baked into artificial intelligence systems, OpenAI’s review team admits that “the company is not quite where it should be.” But, as noted by Sandhini Agarwal, an AI policy researcher at OpenAI, the company will not “broadly expand access to GPT-3 until it’s comfortable that it has a handle on such ethical issues. If [the company] opens it up to the world now, it could end really badly.” While OpenAI is certainly attempting to generate a “detoxified,” compassionate, and empathetic algorithmic system, it is not clear how the company will alleviate the risk of toxic language and bias. The current way of giving machines language—of training models to think and write like humans—must change. We must pay attention to who codes, how we code, and whether training sets reflect humanity’s diversity and sense of ethics.
About the Author
Aditi Desai is a first-year student at Princeton University studying Neuroscience.
References
Heaven, W. D. (2020, December 10). OpenAI's new language generator GPT-3 is shockingly good-and completely mindless. MIT Technology Review. Retrieved from https://www.technologyreview.com/2020/07/20/1005454/openai-machine-learning-language-
generator-gpt-3-nlp/
Metz, C. (2020, November 24). Meet GPT-3. It has learned to code (and blog and argue). The New York Times. Retrieved from https://www.nytimes.com/2020/11/24/science/artificial-intelligence-ai-gpt3.html
Metz, C. (2019, July 22). With $1 billion from Microsoft, an A.I. lab wants to mimic the brain. The New York Times. Retrieved from https://www.nytimes.com/2019/07/22/technology/open-ai-microsoft.html
Ray, T. (2020, August 25). What is GPT-3? Everything your business needs to know about OpenAI's
breakthrough AI language program. ZDNet. Retrieved from https://www.zdnet.com/article/what-is-gpt-3-everything-business-needs-to-know-about-openais- breakthrough-ai-language-program/#:~:text=is%20GPT%2D3%3F-,Everything%20your%20
business%20needs%20to%20know%20about%20OpenAI's%20breakthrou gh%20AI,like%20a%20person%20wrote%20them
Wu, Y. (2018, May 16). Smart compose: Using neural networks to help write emails. Google AI Blog. Retrieved from https://ai.googleblog.com/2018/05/smart-compose-using- neural-networks-to.html