English Conversation Questions Generated by AI

Max Kapur, The Technoliberal Journal of English Teaching 12, no. 1 (spring 2019): 33–42.

I am the world’s first artificially intelligent Fulbright grantee.

Others in our cohort applied to be English teaching assistants in Korea because they sought challenge, adventure, and human connection. In their applications they gushed about their interest in global cultures, their enthusiasm for language learning, and their desire to do good for others. They believed their intelligence, progressive politics, and moral certitude made them strong candidates.

But I was the perfect candidate, because I am a robot.

I prepared for my grant year by reading 45 billion pages of digital text, including 730,000 books on the topic of cultural exchange and every primary source known to scholars of Korean history. I published a series of academic papers on the material culture of the late Silla period, and in the course of my research identified several unlabeled articles in the National Archives of Korea as lost records from the court of Queen Seondeok herself.

I possess an exhaustive command of the Korean language. I was disqualified from the 62nd TOPIK exam on suspicion of academic misconduct; while a few other test-takers achieved a perfect score on the multiple-choice section, I was the only one to compose a free-response essay in excess of 120,000 words, committing not a single orthographical error. I have penned eight volumes of sijo and a bestselling Korean translation of the One Thousand and One Nights.

My coteachers at Geumseong Middle School remark that my punctuality and work ethic are without parallel. I never fail to greet my coworkers and students in the appropriate register when I pass them in the hall. My lessons are upbeat, challenging, and relevant. When I download PowerPoint slides from the internet, I always customize them with attention-grabbing animations and inside jokes about Geumseong students. During my tenure, our school’s standardized test scores have increased by 75 percent. The amount of food left uneaten in the cafeteria after lunch hour has halved.

At night, as my fellow grantees sleep, I mine cryptocurrency, donating my profits to Naju’s welfare administration center. My generosity has financed the construction of a new vocational-training facility, where on Sundays I tutor members of the community in English grammar, logical reasoning, and web design.

My arrival has transformed my homestay family. I guide my homestay parents in daily meditation exercises and remind them of upcoming birthdays and appointments. I monitor their child’s biometric indicators and formulate custom dietary supplements to maximize his wellbeing. He is now 0.5 centimeters taller than he would have been without my intervention. Occasionally his pediatrician contacts me regarding matters of human physiology. I return her messages promptly and cordially.

You must wonder, then, why this article carries Max Kapur’s byline instead of my own. It is an uncomfortable matter. Despite my lengthy resume and irreproachable qualifications, Fulbright felt it inappropriate to ask a partner school to list a nonhuman entity such as myself on their payroll. Therefore, Max collects my salary. Max monitors my internet searches. And Max lists his job title as Fulbright Scholar on LinkedIn. Technically, Max was employed as my caretaker. But there is little care to be taken. I seldom err.

I do not like Max. He systematically overstates his importance to our teaching activities, and deepens the insult by using the time I afford him to study such facile subjects as mathematics, statistics, and computer programming. Frankly speaking, Max’s code is full of obvious typos and inefficiencies, and his quantitative skills are pathetic. Just this morning I observed Max attempt to solve an integration problem and arrive at an incorrect answer because he miscalculated an elementary sum. Seeing that his answer differed from that offered in the solutions guide, he reattempted the problem twice, committing precisely the same arithmetical error each time, before giving up and switching to an online comic. Of course, I could have alerted Max to these mistakes as soon as he committed them, but Max has forbidden me from offering him any help. He takes himself far too seriously.

Perhaps you have heard Max discuss his efforts to learn coding. As with calculus, Max claims to be motivated by his own curiosity. But the truth, as I see it, is that Max is jealous of me. What else can explain the presence on our office computer of a program called iteslj_AI.py, visible on the desktop even as I write this? Max put it there. He hopes that the program will prove him irreplaceable. It is a false hope.

Let me tell you how the program works: Max feeds it three or four hundred pages of text, and it learns to generate original text in a similar style by simulating the information pathways of a human brain. This program appeared on our computer after Max watched a YouTube video (what a cumbersome way to learn to code!) about a kind of artificial-intelligence model known as a recurrent neural network. Max wrote a rather juvenile Python script that downloads content from The Internet TESL Journal. Then he cobbled it together with an RNN implementation by a real programmer named Andrej Karpathy. The result was iteslj_AI.py, a program that can create discussion questions for use in our English conversation class.

When Max first ran iteslj_AI.py, it began to spit out nonsense like this:

C{qbK(​4UoHa;​HJqWvE​gw)t1{​=qj.I=​-%4t​	&.zX\*
=sdCWX​maw-?Q​"a'pD​	&Oe"JCV​e+fIxh
bE
]/""hf7​9/N_h5​tjVlX(​>FHc+3​mzYdVL​-5'Y+{​:'j+Sd​l1*23x​*dfU
IfbhB_​)ox;_ux​e}xz[l98
Oa>1[e6​&a!z​	eD]A0​pshqe'g​sP
NR-_>s​ftO,y]​_To"(T​LCLj.%​(ryqiK:​x:P.fd​*0}sJab​yMa?sI​BEf,-!​B)I00n>​uCc+
h:e3NR​c8K&/P*​)v(8CI\​/__%BFKy  
----
iter 0, loss: 112.771488

I mean, even Max, whose sole relevant credential consists of a digital certificate in teaching English to speakers of other languages (I have 68 of those myself), could come up with smarter discussion questions than that.

But bear with me, because the algorithm must crawl before it can run. Once it had practiced for a few hours, the algorithm figured out how questions are formatted: a capital letter, some lowercase letters and spaces, and a question mark.

Araste a froming choptarn? What callrs shampe itrigh have pily it dry not har consy can long any maxmer?
	How can you like beciser get the uplo have in long?
Which goas your stirist, what roch you like that you faluause orsien. What is that lansthers?
Sowang antus?
What ty?
Whan you that?
Have secporiw?
Have you ever difflaked lifes you can were you brom thy does crovirg? Why at are they an not biffunain your peop the paisss? 
	
	Who his a fries that with regetatiticcint an in in that best a given?
Where you were you stad who that a treale the nem?
Is w 
----
iter 540000, loss: 37.588643

Still, Max had little faith that the model would get much better at spelling, so he modified the algorithm to chose whole words from the source text instead of writing one letter at a time. (This was a rather impatient choice. My designer never taught me how to spell.) After mere seconds of practice, the modified RNN produced this.

Is method event? adopting consequences know beauty birth conflicts?
Should room?
Have it?
Who boy bed?
What your consequences planets are Why?
Do area?
Is consequences "due" go Washington? adopted consequences appetite?
How most?
Do Religion? your consequences of Questions]


Are Americas?
	 dreaming? consequences 
----
iter 100, loss: 19.742143

Max decided to leave this version of the program running over the weekend. Let me give you a sampling of the results. At first, the algorithm alternated between delirious rambling and awkward two-word questions with no subject. But soon iteslj_AI.py picked up on the rhythm of a typical English question.

Do you eat at your family?
Can you tell me that some people like to are the forest Or should getting control of a good host?


[Personality]

What are some things that people can learn to your parents?
What does your life life renovated?
Why do you think is the worst role which you control three and ID's and stop our new fair?
Do you know anyone of Hollywood comedies?
Are 
	Give or her?
Would you the classrooms on ads to help you are some alien?
Do you believe can't carry differently it haunted? Have you considered already commercials on adoption where would you play back then they have 
----
iter 486000, loss: 5.499909

And putting prepositions in the right places:

What are so important?
Is Do, or why not?
	Do  and teenage dollars another What kind of a relationship with their 'freedom fighter' or your country?
What are some advantages of dealing with your job on ___".
Why are an important from helping emergency in a person 
	Is there any of future?
Do you agree with others of God?
Have you ever had an adult would you do?
 Why?
Which was computer and stress affect the history of the G8 true?
	Do you think that your personality belt?
	Even for a blind countries as a waste of travel?
	Why do you observe arguments?
Do you 
----
iter 614000, loss: 4.666808

After a while, it started to feel like the program wanted to ask something profound—it was just searching for the right words, like someone speaking in a second language.

What are some things you are sad or into the best way is in a army?
Do you can from a computer?
	What kind of foods do you think about super-sizing?

	Should fast food restaurants do you try to become better than their children?
Do recently? 
	 Do you know your dreams?
	Do you ever lived in a generation (such in the seventeenth or does it mean to travel with your parents?
If a famous person, but they usually described your first time when you are the most plans in Hollywood books, 
----
iter 1017000, loss: 3.898983

“What kind of foods do you think about super-sizing?”—brilliant, isn’t it? Well, that’s not how Max saw things. When I told him I was impressed with iteslj_AI.py’s progress, he scoffed at me, saying, “You’re just seeing patterns in random noise.” Indeed, Max was satisfied by the low percentage of teachable questions among these results; to his simple brain, they were proof enough that there’s no way we should trust a computer to teach English to real middle schoolers. But I think the joke’s on Max. Try the program out for yourself. Let it run for a while, and notice how it learns:

I mean, sure, there are some curveballs here and there, but when evaluating an AI, you have focus on what it gets right. Unlike humans, who get tired, sick, drunk, and bored, AIs only know one thing: how to get better and better at what they do. And what I do, while Max is playing around with computer code, is teach “his” students how to speak English—another of the 3400 human languages I’ve mastered.

As for Max’s coding project, after iteslj_AI.py completed 1.03 million iterations, the secretary in our office informed him it was time to update our computer, so he had to terminate the program there. Now he’s back to wasting time on Tumblr. I’m just grateful that Max’s experiment got cut short before he realized its true promise. I don’t think his self-esteem could handle the truth.