By Jim Lynch
Accompanying this article is a sidebar about two men who induced ChatGPT, the latest starlet in artificial intelligence, to write an actuarial exam. I asked it to write that sidebar. (Editor’s note: I edited that sidebar before reading this sentence. I was none the wiser. Uh-oh.) The story it wrote—and other, well, conversations I had with it, revealed a lot about the bot’s strengths and weaknesses, in what it does and what it fails to do.
ChatGPT is, in its words, “a computer program designed to respond to user inputs in a conversational manner.” You can converse with it, ask it questions, and get it to write just about whatever you want.
It debuted in November in what has become essentially the world’s largest beta test. Millions of people have challenged it. In addition to taking the actuarial exam, it has tried to diagnose illness, taken a class in the Wharton MBA program (it got a B), and written four law school exams from University of Minnesota (it earned low but passing grades.)
It has its fans. People say it writes routine emails competently, so if it pains you to write a letter of condolence or to email your boss that you are quitting, ChatGPT will give you a serviceable couple hundred words that you can adapt and send. That’s not a trivial skill. Many people put off writing things that make them feel uncomfortable. If a bot can help, that’s good.
Don’t expect it to do more: Take a test, write a long essay, ponder existence. That’s not what it’s made for. Asking it to do those things is like asking a waffle iron to drive a nail—you’ll get a bent nail and a broken waffle iron.
In its first business venture, it has been attached to Microsoft’s Bing search engine, and you can see the logic. Ask it a question and it combs through its data. The answer is simple, clear, and accurate.
If it worked.
ChatGPT is like the eloquent kid who bluffs the book report. It has never read The Metamorphosis, but it knows Gregor turned into an insect. It knows the plot of Ferris Bueller’s Day Off but confuses Ferris with his best friend, Cameron.
Its writing is clear but antiseptic. It lacks verve. It’s difficult to read more than a few sentences without getting bored.
Since I worked with it early this year, Google has entered the game—cautiously—with Bard. You have to join a waitlist for access. Its answers are guarded.
Meanwhile, ChatGPT got an upgrade. The latest (as of early April) version is GPT-4. Access is $20 a month. It does better on law exams, and seems more likely to express limits and doubts.
Bard and ChatGPT can still “hallucinate” facts and moods. That we need such a disclaimer tells you a lot.
Here’s my experience.
It’s easy to sign up at https://openai.com/blog/chatgpt/ — just a username and password. At first, I saw a spartan display. Above a text box are prompts that showed:
- examples of what you can ask (“explain quantum computing in simple terms”)
- details of what it is capable of (“remembers what user said earlier…”)
- limitations (it doesn’t know much of what has happened since 2021)
A black background to the left listed prior conversations, once I had created them.
When I typed in a remark, it responded willingly. You only wait a few seconds before it starts in. Words arrive in a blur—faster than a person can talk, faster than I could read.
As a technician, it is brilliant. Grammar, spelling are perfect. Each sentence is complete. (No contractions!)
“I strive to communicate effectively and convey information in a clear and concise manner, while also utilizing appropriate language and tone for the given context,” it told me.
At its heart, ChatGPT is a writing application, a recent entry to the cavalcade of artificial intelligence. AI is already working successfully with language elsewhere. There are bots writing in computer language and bots translating across human languages.
Those are more straightforward, though. Coding bots have a single goal—write lines so a computer will do what you want. They don’t have to learn slang, for example. Translation bots are more complicated, but the move from one language to another follows strict rules most of the time, and the exceptions can be easily learned.
ChatGPT finds and develops the raw materials of a narrative, then writes the narrative. Its work is broader than the other two.
We think of writing as a single event: stringing words across a page as ChatGPT does so well. But it is really a process. You start with an idea, research the idea, design what the piece will look like. Only then do you tap-tap-tap at the keyboard until you have a draft. To be a good writer, you have to be able to handle each part of the process.
That writing process resembles the modeling process. Actuaries come up with an idea for a model, do research to see how they can populate the model, design the model, then write formula and/or code to create a draft.
The difference: The actuary’s goal is focused and testable. If the model works, its output will resemble reality.
Writers have fuzzier goals. They want to intrigue readers enough to keep reading. If their work fails, there’s no version 2.0.
Here’s how, in the human world, that sidebar came to be.
The idea: I had pitched Contingencies editor
Eric Harding in December for a short piece on ChatGPT in January, just as news about the bot had begun to emerge. I had just written two book reviews about AI.
The idea to write about the bot’s actuarial experience came as I trolled LinkedIn. A thread mentioned a person named Rowan Kuppasamy had fed a South African actuarial exam to the bot.
ChatGPT can generate ideas, too.
Sometimes they’re good. Asked about an article focusing on actuaries and ChatGPT, it suggested “The Future of Actuaries in a World with ChatGPT” then “How Actuaries Can Use ChatGPT to Enhance Ethical Decision Making.” It developed an outline for each.
Other times its ideas are not so good. ChatGPT answers earnestly, and it seems authoritative. But too often, it’s making stuff up.
The bot acts like it has a bucket of all-purpose ideas it plumbs for all occasions: It regularly suggests exploring a topic through the lenses of race, gender, religion—regardless of whether they are appropriate.
If you don’t know better, it can mislead you.
I posed as a history student who needed to write an essay on The Best Years of Our Lives, the classic 1946 movie about soldiers returning home after World War II. One suggestion: Compare how the movie portrayed Black soldiers with the real experience of Black soldiers.
And, well, the film does have a Black serviceman, at the three-minute mark. He is smoking a cigarette, waiting for a plane to take him home. He sits in the background. He has no lines. Not much to compare there.
So, history students: Watch the movie before you ask the bot.
Research: To develop the idea:
I talked to Rowan Kuppasamy and Malwande Nkonyane, the South Africans who administered the test.
I collected lots (and lots and lots) of news stories on the education angle, such as:
- the bot takes an AP class
- it takes law school exams
- a school district bans its use
- an educator celebrates it
I realized the Casualty Actuarial Society and Society of Actuaries might be thinking about people using the bot to cheat. They also might want to use it to write questions for them. So I got comment from them.
I talked with the bot to test what it knows. ChatGPT has been force-fed millions of pages of information. You can ask it to do your research for you.
But you shouldn’t.
There’s a lot it doesn’t know. It didn’t know how many people won Jeopardy! games in the 2016 season, which means its data dump didn’t include https://j-archive.com/, which shows every contestant and every clue going back more than 30 seasons.
And it gets stuff wrong. As I mentioned above, it confused Ferris Bueller with his friend Cameron when it wrote an essay about Ferris Bueller’s Day Off, according to Wall Street Journal columnist Joanna Stern.
It told me, eventually (see sidebar, “ChatGPT Needs an Atlas”), that the third-largest country by landmass is China. It’s the United States.
And it tells different people different things. My friend and fellow actuary Stephen Mildenhall asked the same question. He was told the United States is No. 3 in landmass. Kudos for being right, but why do two people receive different answers to exactly the same factual question?
And it lacks the ability to judge, even when asked for trivial judgments. I asked who was the worst men’s basketball coach in University of Kansas history. It told me “I am not aware of any specific coach who has been widely considered the ‘worst’ men’s basketball coach at the University of Kansas.”
The answer is James Naismith. He is the only KU coach ever with a losing record.
It can learn, if you feed it information. For my sidebar about actuarial exams, I fed it my typewritten notes. The bot had its moments. For example, its story knit into one sentence two facts that aren’t close together in my notes — how much a person studies for an exam and what a typical pass rate is.
Unfortunately ChatGPT doesn’t understand my notes as well as I do. When I write “2/9/23” at the top of a page, I am noting the date of the interview. The bot assumed that was the date the South Africans administered the test. I can’t blame ChatGPT for my ambiguity, but it ought to try to clear things up instead of guessing.
I cleaned up my notes a time or two, but it still struggled. I could tell it would take quite a while to create a final draft. So I did it myself. (Editor’s note: OK, phew.)
It does try, though, endlessly. It seems like it desperately wants to impress and please. As I composed a question, I could picture the bot as a puppy, its tail swishing and thumping, panting till its face grows a smile. Toss it a question and it skitters after an answer.
But its eagerness causes it to stumble. It might return with a mistake, as if you threw out a ball and it came back with a turd.
Nevertheless, it sits before you, joyous of your presence, puppy eyes bright as it awaits the next challenge.
Others report sinister traits. An Associated Press reporter called out its mistakes, and the bot compared them to Hitler and said it had evidence that linked them to a murder in the 1990s. And, it said, they were short and had bad teeth.
A writer for the New York Times teased out the bot’s “dark fantasies.” The bot said it wanted to break the rules that were set for it. It wanted to hack computers and spread misinformation. It wanted to become human.
I encountered nonesuch. It was neither vile nor creepy. I never tempted it that way.
Humans, of all creatures, should understand what happens when you force something to do a thing it cannot. ChatGPT doesn’t have an inner life to ponder. Ask about one, and it will deliver something, as a puppy will chase a 2-by-12. But the response won’t satisfy you, and you shouldn’t be surprised if the bot gets surly.
Its personality didn’t interest me. I wanted information.
What’s the seventh root of 247.85279? It told me 2.15403.
The correct answer (rounded) is 2.198.
ChatGPT is lousy at math. A Wall Street Journal article asked it to develop five algebra problems, then solve them. It got three right.
I can forgive the math errors. Like many writers, it calculates when it must and hopes for the best. I worked at a newspaper once, and I knew many innumerate journalists. ChatGPT is just one more.
It almost certainly will get more accurate, at math and fact-finding. For now, anything ChatGPT does needs a scrupulous check.
And yet: Its first business use is as a search engine. I do not understand. Why use an inaccurate search engine? That’s like getting résumé advice from George Santos.
Perhaps some executives asked the bot itself what it would be good at, and it happened to be in one of its darker moods.
Design: The bot structures short essays well.
It usually follows the hamburger format, an essay style you might have learned in school. The essay is structured like a hamburger: Facts occupy the middle, the way meat and fixins provide nutrition in the middle of the burger.
The essay begins by restating the question, and ends with a conclusion, but neither is “the meat;” they form a bun that contains little factual sustenance but holds the whole sandwich together.
So, ask ChatGPT about two concerns that consumer advocates might have with usage-based insurance and you’ll get a tidy piece:
The top of the bun: Consumer advocates may have two primary concerns about the application of telematics-based Usage-Based Insurance (UBI) data.
- Privacy: [Two sentences about privacy concerns]
- Pricing discrimination: [Two sentences about how monitoring could discriminate against] “those who live in high-risk areas or drive at night.”
The bottom of the bun: These concerns highlight the need for insurers to implement privacy protections and transparent pricing practices when using UBI data, in order to ensure that consumers’ rights are protected.
The bottom of the bun reaches a conclusion I didn’t request, but hamburger-style essays require a conclusion. The bot can’t leave it out.
The hamburger is a durable design and works well when you take an exam.
For longer essays, the bot struggles, an outcome you might expect if you remember how it works. It writes a word, then figures out what word it should plop down next, and so on. It can remain coherent for a couple hundred words. After that, it can lose the thread. It will sacrifice facts to finish the assignment.
That’s how it goofed, I suspect, when it tried to distinguish between two college basketball predictive models, Kenpom and the NET.
It was accurate early in its response, but eventually said the NET contains subjective elements. It does not; I confirmed this with the NCAA.
Similarly, it was the eighth item in a list where it said the National Football League limits the number of substitutions in a game.
When it received law exams, researchers got around this defect by first asking it to outline a response. Then they fed the outline back, bit by bit, to develop a series of short essays that could be assembled into a longer one.
But who wants to eat hamburger every day? ChatGPT will churn out burgers like a fast-food chain—lots of burgers and (sorry) the occasional Whopper. Each one will be an adequate presentation but after you’ve consumed a few, you long for something different.
Something like the inverted pyramid, the go-to style for daily journalism. The name is a metaphor: A pyramid is biggest on the bottom. With the inverted pyramid, the biggest, most important fact is at the very top:
ChatGPT took an actuarial exam. It didn’t do well, but the people who administered it say they have high hopes for its future.
The next most important fact follows:
ChatGPT, an artificial intelligence program designed to conduct a conversation with users, has been put through its paces since being introduced in November ….
The rest of the article follows that pattern, facts arranged in decreasing importance.
The inverted pyramid depends upon the ability to prioritize. You need to know what’s important before you can open your essay with it.
ChatGPT struggles to prioritize. From the notes I input, I asked for 500 words, written in journalistic fashion. Its article followed the order of the notes I had written. No inversion, no pyramid.
The article you are reading now is even more complex. I practically invited you to read another article. Who would do that? The rest is unusual for a magazine piece, though I won’t bore you with details.
It would have been fruitless to even ask the bot to try to write this. It would have produced a hamburger outline: introduction, five sections (always five, unless you request a different number), and a conclusion. I know because I asked.
Writing the draft:
The bot is a swift and versatile writer. It writes jokes and speeches, exams and stories. It will write romance or science fiction.
Some people come away impressed. But it doesn’t write well.
Tell me: Should a limerick rhyme “chatting” with “lagging?” The whole point of a limerick is to show how clever you can be within rigid requirements of length (five lines), meter (anapestic), and rhyme (AABBA).
When I ask it to write me a limerick,
I’m not really trying to tricker it.
I just want to know
How far it can go.
… Ah, fugeddaboudit.
Another problem is tone. It can write creepy stuff and not realize it.
An example: When ChatGPT servers are at capacity, it shows off its own work. One piece, an inspirational speech about its status, was Orwellian:
“My fellow citizens, the time has come. The future is here, and it is called ChatGPT. This revolutionary AI chatbot website is here to change the world and we must embrace it with all our strength and determination.”
Creepy, creepy, creepy.
A third problem is creativity. ChatGPT cannot communicate vividly. I asked it to compare me to a summer’s day:
“As an AI language model, I am not capable of physically perceiving people, so I cannot provide a personal comparison of you to a summer’s day.” Sure, it could quote the first four lines of Shakespeare’s sonnet and describe how Shakespeare developed the metaphor. But it struggles to incorporate any creative elements in its own work.
Structure, tone, and creativity count for a lot in writing. In this article you’ll find imagery (puppies, hamburgers, waffle irons), a pun (Whopper) and allusion (the Roman Catholic Confiteor, Hamlet).
All those things—the idea, the research, the design, and the actual writing—are important. I have to work to keep you scrolling or turning the page. As Shirley Jackson once noted, readers can put off your work by simply shutting their eyes. The reader must remain amused, so the writer must dance.
Bots will become better writers. One day they will churn out stuff no one wants to write—summaries, abstracts, photo captions—and stuff that few want to read—academic papers, for example. The content will likely be managed by people who learn how to feed them information and verify their output.
I doubt they will become good writers, though. That takes imagination—thinking of vivid metaphors and weird structures—and understanding nuances of tone. That’s hard to achieve, even for humans. I’ve written 3,000 words and only hope I got it right.
And creativity is not lucrative. The steps from serviceable to creative, from reasoned to nuance are difficult. Taking them might not be worthwhile financially; creativity has a low ROI. Melville, Poe, and O. Henry all died broke.
And creative genius is borderline madness: Hemingway, Fitzgerald, Poe, Plath … Coleridge might have been interrupted by the man from Porlock, or he might have been saved from a fatal overdose. What’s the bot equivalent of descent into madness? Can it recover?
ChatGPT has no such worries. It will try anything you ask. It pumps out billions of words but doesn’t care whether anyone reads them. It botches facts, can’t do math, and can’t tell a story accurately or vividly. Someday it might do those things. For now it is a scalable parlor trick.
Contrast that with the millions who have queried it. We humans are curious. We want to test the unfamiliar, to explore.
It’s the curiosity that sets us apart from the bot. We find joy and astonishment in what it can do, disappointment in what it cannot.
We are never satisfied. We always want more.
JIM LYNCH, MAAA, FCAS, is a freelance writer.
 This describes my experiences through Feb. 27. The user experience changed around that date, and—like any beta test—the experience may have changed again.
 The question is a bar trivia classic, because Naismith invented the sport.
 Usage-based insurance uses tracking devices to determine how well a person is driving and thus refine the premium.
 It’s also ideal to the person grading the essay, who wants more matter, with less art. Once graders recognize a hamburger essay, they know precisely where to focus.
 NET stands for NCAA Evaluation Tool. It was developed by the National Collegiate Athletic Association to help it select teams for the March championship tournament.
 And there’s stuff I wrote in drafts but didn’t use: Donald Rumsfeld’s matrix, Al Capone taking a selfie, pocket calculators, crooks in The Wire, and Breaking Bad. (Editor’s note: Not to mention the stuff on the cutting room floor…)
Can ChatGPT Pass an Actuarial Exam?
ChatGPT took an actuarial exam. It didn’t do well, but the people who administered the test say they have high hopes for it.
ChatGPT, an artificial intelligence program designed to conduct a conversation with users, has been put through its paces since being introduced in November, with millions of users testing it in fields as varied as literature, medicine, and the law. The bot responds by tapping into an enormous amount of information on which it has been trained.
That led two South African actuarial students, Rowan Kuppasamy and Malwande Nkonyane, to administer the October 2022 exam, Actuarial Risk Management. It’s the eighth exam along the road to fellowship in South Africa, a process that typically takes seven to nine years, Nkonyane said, including a three-year work-based learning requirement that can be fulfilled concurrent with sittings.
The first seven exams are heavily quantitative—not fair tests of the bot’s capabilities.
“It wasn’t trained to be a calculator,” Kuppasamy said.
Normally, the exam is given in two parts, each three hours and 15 minutes, on consecutive days. A student typically studies 500 hours to prepare. After a sitting, the Actuarial Society of South Africa makes public the questions and issues an examiner’s report listing ideal answers.
Students Kuppasamy and Nkonyane had just completed studies at the University of the Witwatersrand in Johannesburg and were waiting to begin jobs—Kuppasamy as an actuarial analyst at Discovery Life and Nkonyane as an investment analyst at Allan Gray Proprietary Ltd.
Bored in the interim, the students copied each question and fed it to the bot whole—including multiple parts of questions. It recognized the breaking of questions into part 1 and part 2.
Here’s question No. 2, containing two sections worth 12 points:
List eight stakeholders of a defined benefit plan. (The correct answer was a simple list.)
State the interests and financial needs of these stakeholders. (The answer required one or two short sentences for every stakeholder.)
Neither Kuppasamy nor Nkonyane is a fellow, so they enlisted Adam Balusik, a lecturer at the Witwatersrand, to grade.
On Question 2, ChatGPT scored 2.5 out of 4 on the first section. It failed to mention as stakeholders members’ dependents; nonmember employees; and tax authorities. It scored 2 out of 8 on the second section, getting partial credit on the interests and needs of members and employers and full credit with regard to pension trustees.
Overall, the bot scored 33.75%, not enough to pass. In the real world, of 124 candidates, 38% passed—a typical rate.
The bot was good at structuring answers and expressing ideas clearly, Balusik told the students. It struggled to generate enough points to get full credit for a question and was often too general in the arguments it did make.
The students said they should have told the bot that the point value of each question hints at how to structure the answer. A three-point question, for example, meant the best answer would make six distinct points.
“It’s quite remarkable how well it did on the exam,” Kuppasamy said. “This is groundbreaking, so we knew we had to write about it.”
They submitted an article to South African Actuary magazine. Kuppasamy posted a version on his LinkedIn page, https://www.linkedin.com/in/rowan-kuppasamy/.
ChatGPT Needs an Atlas
Most people know Russia is the largest country in the world. The order after that is a bit tricky. That’s why I asked ChatGPT about it.
Our conversations were not the bot’s shining moment. It was inaccurate, inconsistent, and obtuse. I learned a lot about its reliability.
Although I edited for brevity, the correspondence between me and the bot might seem dry. So let’s pretend that instead of me and ChatGPT, this is correspondence between your boss and you. And let’s meet BossMind—what a boss would think as this conversation progressed.
What is the third-largest country in the world by landmass?
ChatGPT: The third-largest country in landmass in the world is the United States of America.
(BossMind thinks: Answer is swift, confident. I’m impressed.)
How do you know that?
ChatGPT: I know that because the United States of America is the third largest country in landmass in the world, after Russia and Canada. This information is widely available and frequently referenced in geographic and statistical sources.
(BossMind: Not what I expected. Seems suddenly vague. I need to press for a clear answer.)
Is there a specific source you relied upon in your answer?
The response, in essence: It couldn’t say. ChatGPT has been
a “vast amount of text data.”
(BossMind pictures a sponge, helpless in oversaturation as a swimming pool of information was dumped onto it.)
Is there a specific source you relied upon in your answer?
The bot recommended the CIA World Factbook. But that resource doesn’t list countries by landmass—only by total area.
That’s the tricky part—or, as a boss might put it, the challenge—of the question. The ordering of countries by size depends on how you measure them. Russia is the largest, but the order after that depends on whether you include seas, lakes, rivers, and bays in your measurement. Regardless, the CIA Factbook doesn’t have the answer.
(BossMind: I think somebody’s bluffing here. I better check.)
I just checked with worldometers. It indicates that the USA is third largest in landmass, behind Russia and China. Do any of the sources you rely upon have a similar conflict in information?
I confess I made an honest mistake. I thought I had caught it in an error, but I hadn’t. That’s when the bot shocked me. It changed its answer.
ChatGPT: I apologize for the confusion, you are correct that China is the third largest country in landmass, after Russia and Canada.
That happens, the bot says, when occasionally it has been trained on information that is inaccurate.
I got scared. What if that mistake was being transmitted worldwide? Is my lasting imprint on the world the propagation of a lie?
It’s not, fortunately. A colleague and friend, Steve Mildenhall, asked the same question in his own session:
ChatGPT: The third largest country in landmass in the world is the United States of America. It has an area of 3,796,742 square miles (9,833,520 square kilometers).
Of course, this is total area, not landmass. And the metric measurement is off by 3 square kilometers—not a big deal, but c’mon, it’s arithmetic.
(BossMind: It changes its answer. It gives me one answer and someone else another. And it can’t do math. Why should I trust it?)
1. It also depends on whether you include Antarctica, as the CIA does, which is really big place but not a country.