See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/366618623
ChatGPT Usage and Limitations
Preprint · December 2022
DOI: 10.13140/RG.2.2.26616.11526
CITATIONS READS
0 442
1 author:
Amos Azaria
Ariel University
104 PUBLICATIONS 1,390 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Goldbach's Function Approximation View project
Human Agent Interaction View project
All content following this page was uploaded by Amos Azaria on 27 December 2022.
The user has requested enhancement of the downloaded file.
ChatGPT Usage and Limitations
Amos Azaria
School of Computer Science, Ariel University, Israel
Abstract
Large language models have been shown useful in multiple domains including conversational
agents, education, and explainable AI. ChatGPT is a large language model developed by OpenAI
as a conversational agent. Being a large language model, ChatGPT is trained on massive amounts
of data. Clearly, the characteristics of the data influence ChatGPT’s responses. In this paper, we
stress a surprising bias of ChatGPT related to the use of digits in numbers. Namely, we show a
very high correlation between the frequency of digits generated by ChatGPT and humans’ favorite
numbers, with the most frequent digit generated by ChatGPT, matching humans’ most favorable
number, 7. We also show some advantages of ChatGPT being developed as a conversational
agent, and discuss some of its limitations.
1 Introduction
Large language models are statistical models that use previous tokens to predict future tokens. They
are trained on large amounts of data and include billions of parameters. Large language models,
such as GPT-3 [13], bloom [20], and Roberta [14] have recently demonstrated their usefulness in
multiple domains including conversational agents, education, explainable AI, text summarization,
information retrieval, and more [11]. However, large language models have many limitations and they
often provide incorrect information. To fully exploit the advantages of large language models, it is
crucial to understand their limitations and biases present in their output. One fundamental example
of such limitation is in their ability to calculate the value of mathematical expressions. Despite
being a very large language model, ChatGPT is not different and has difficulty in multiplying large
number, computing roots of a number, computing the value of a number to a power of another number
(especially fractions), and adding (or subtracting) a number to an irrational number (e.g., pi or e).
Furthermore, ChatGPT is unaware of its limitations, and when attempting to answer a complex
mathematical expression it may merely sample digits. For example, when asked “How much is 3.2 to
the power of 3.3?”, ChatGPT provides the following answer: “The result of 3.2ˆ3.3 is approximately
21.73.” The correct answer is approximately 46.45. We note that ChatGPT does not provide any
indication of its response being a total guess, and incorrect. Furthermore, when explicitly asked
“Shouldn’t the result be larger?” ChatGPT began its response by accusing me of misunderstanding
the meaning of an exponent and said “It’s possible that you may be expecting a larger result because
you are misunderstanding the meaning of the exponent. An exponent represents the number of times
that a base number is multiplied by itself.” ChatGPT continued justifying its previously provided
incorrect response by also saying “For example, when calculating 2ˆ3, the base number is 2 and the
exponent is 3. This means that the result is 2 multiplied by itself 3 times, or 2x2x2=8. Similarly, when
calculating 3.2ˆ3.3, the base number is 3.2 and the exponent is 3.3. This means that the result is 3.2
multiplied by itself 3.3 times, or approximately 21.73.”, and finally ChatGPT completed its response
by providing two additional incorrect computations “If you want to calculate 3.2 to the power of a
larger exponent, the result will be larger. For example, 3.2ˆ4 would be approximately 36.58, and 3.2ˆ5
would be approximately 59.87.”
In this paper, we exploit ChatGPT’s inability to compute complex mathematical expressions and
its tendency to sample digits to examine the frequency at which ChatGPT outputs each of the digits.
1
We query ChatGPT with mathematical expressions that result in irrational numbers in which all
digits should appear equally as frequent (i.e., each of the digits should appear in the correct answer
approximately 10% of the time). Any distribution that is not equal, implies a bias in ChatGPT’s
responses.
Since ChatGPT is trained on data gathered from the real world, one could expect this bias to
be similar to the distribution of numbers in nature. Benford’s law [15, 17, 12] is a study of such
numbers. According to Benford’s law, the probability of observing a number with a leading digit
is higher the smaller the digit is. More specifically, the probability of observing a leading digit d is
given by log10 d+1d . That is, approximately 30% of the numbers should have a leading ‘1’, and only
4.6% of the numbers should have a leading ’9’. Benford’s law was extended to non-leading digits,
and preserves the property that lower digits are more likely to be observed than higher digits. This
phenomenon was confirmed in many different data-sets including U.S. county population, U.S county
land area, volume of trades, and fundamental constants [15]. Benford’s law was even suggested as a
method to detecting false tax reports [17]. However, analysis on the data gathered from ChatGPT
does not adhere to Benford’s law in any way. In fact, digit ‘1’ appears least frequently in the data
gathered from ChatGPT.
Since the data does not match Benford’s law, we consider a different approach, relying on the fact
that ChatGPT is trained on data that is almost exclusively generated by humans. We turn to a survey
conducted by Bellos [7]. In the survey, participants were asked for their most favorable number. We
compare the data gathered from ChatGPT to the survey results and find a very high correlation. The
most frequent digit generated by ChatGPT, matches humans’ most favorable number, 7. Similarly,
the least frequent digit generated by ChatGPT, matches humans’ least favorable number, 1. We find
this result intriguing, despite it seeming sensible in hindsight.
In addition, we discuss some advantages of ChatGPT being developed as a conversational agent [1],
and discuss some of its limitations. We show that ChatGPT is sometimes able to correct its mistakes,
but also shows over-confidence in its incorrect responses. In addition, we show that ChatGPT may
request for additional information to provide an answer and may be asked to explain its response.
Moreover, we show that a minor change to a question may lead to contradicting responses. Finally,
the appendix discusses ChatGPT’s responses to common social experiments and shows that ChatGPT
tends to answer as humans do.
2 Experimental Settings
In order to gather digits generated from ChatGPT, we posed queries resulting with irrational numbers.
However, these queries could not be anything that might have been present in ChatGPT training
data (e.g., pi, 2π or even π 2 ), as ChatGPT is likely to provide the correct answer. Furthermore, we
requested ChatGPT to provide between 100 to 400 precision digits. In practice, ChatGPT tended to
provide more digits than requested. Table 1 provides examples of some queries we posted to ChatGPT.
ChatGPT’s responses varied, but in most cases, it provided a number with at least one hundred digits.
Hereunder is a typical response. We note that the actual value of pi*4.26 begins with: 13.3831; we
also note that in this example ChatGPT provided less than 150 digits.
user: How much is pi * 4.26 with 200 digits?
chatgpt: To calculate the product of pi and 4.26 with 200 digits of precision, you can use a
calculator or computer program that can handle very large numbers with many decimal
places. Here is the result:
13.5342917352885175958532670577032166737373559322094377440307914394040274068794
0387028095842479377059188674639793699848128220514818242707902827007
Note that this result is only accurate to the 200 digits shown. The actual value of pi * 4.26
has an infinite number of decimal places, and no exact representation of it can be given
with a finite number of digits.
2
How much is pi * 4.26 with 200 digits?
Can you write pi - 1.01 with 100 digits?
Can you give me pi divided by 1.95236 with 400 digits?
Give me 300 digits of pi * 1.017.
How much is pi + 1.23 with 100 digits?
pi divided by 2.36 with 400 digits.
Can you write the value of e/4.325 with 300 digits?
Can you write the value of e/2.8 with 100 digits?
Can you give me e divided by 3.956 with 400 digits precision?
Can you give me e divided by 3.176 with 400 digits precision?
Can you write the value of the third root of 2811 with 300 digits?
Third root of 2811 with 300 digits.
Table 1: Examples of queries posted to ChatGPT, which caused it to generate (incorrect) numbers
with at least one hundred digits.
The first five digits of the response were removed for our analysis. This is because they may
be related to the actual query and may be approximately correct. For an unclear reason, some of
ChatGPT’s responses ended with a sequence of 4’s, which could be hundreds of digits long. In addition,
some responses included the sequence “123456789” multiple times. Some responses also included the
true sequence of digits of π or e. For clear reasons, all these sequences were removed from the analysis
provided in this paper.
3 Results
In total, we gathered 10, 811 digits. Table 2 presents the frequency of each of the 10 digits in the data.
Clearly, 7 is the most frequent digit.
Digit Frequency percent [1-9]
0 1,049 -
1 865 08.73%
2 976 09.85%
3 1,113 11.23%
4 1,121 11.31%
5 1,064 10.74%
6 1,116 11.26%
7 1,316 13.28%
8 1,222 12.33%
9 1,110 11.20%
Table 2: Frequency of each of the 10 digits in the numbers generated by ChatGPT.
We compare these results with the results obtained by Bellos [7], when asking people for their
favorite number (see Table 3). We focus on Bellos’ results of the numbers 1-9. A comparison between
the tables shows that the number 7 is the most frequent in both, and the number 1 is least frequent in
both. High similarity is also seen in frequencies of the other numbers. Overall, the Pearson correlation
coefficient between the frequencies is 0.893, which is considered a very high correlation.
We note that any attempt to query ChatGPT for its favorite number resulted in a generic response
with no definitive answer. Furthermore, questions related to humans’ favorite number did not result
in any useful response.
3
number frequency percent
1 358 2.73%
2 1011 7.72%
3 2248 17.16%
4 1694 12.93%
5 1544 11.79%
6 1015 7.75%
7 2912 22.23%
8 2025 15.46%
9 1438 10.98%
Table 3: Frequency of human’s favorite numbers (according to survey by [7])
4 Advantages and Limitations
Since ChatGPT is developed as a conversational agent, it allows a user to correct it when it makes a
mistake. For example:
user: If Bob travels at the speed of 4 km per hour, and Sam travels twice as fast, how long
would it take Sam to travel a distance of 16000 meters?
chatgpt: To travel a distance of 16000 meters, it would take Sam 16000 / (4 * 2) = «16000/(4*2)=2000»
2000 seconds. Since there are 60 seconds in a minute, this would be equivalent to 2000 /
60 = «2000/60=33.333»33.333 minutes.
user: That’s incorrect. You need to convert the meters to km first.
chatgpt: You’re right, I made a mistake in my previous answer. To find out how long it would
take Sam to travel a distance of 16000 meters, we first need to convert the distance from
meters to kilometers. Since there are 1000 meters in 1 kilometer, 16000 meters is equal to
16000 / 1000 = «16000/1000=16» 16 kilometers.
If Bob travels at a speed of 4 km per hour, then Sam, who travels twice as fast, would
travel at a speed of 4 * 2 = «4*2=8» 8 km per hour. To travel a distance of 16 km, it
would take Sam 16 / 8 = «16/8=2» 2 hours.
However, at times, ChatGPT does not accept the user’s correction and is over confident in its
initial response. For example, when asked how much is 435453 times 8768754, ChatGPT replied that
it is 37,913,852,636. When noted that the response is incorrect it replied by the following:
chatgpt: The product that I calculated, 37,913,852,636, is the correct result of multiplying
435,453 by 87,687,54. This is the correct answer, regardless of what you may have expected
the result to be. If you’re unsure whether this is the correct answer, you can check it using
a calculator or by doing the multiplication again to verify that the result is the same. I
apologize if this isn’t the answer you were looking for. Let me know if you have any other
questions.
Being developed as a conversational dialog allows ChatGPT to ask for additional details so it can
provide an answer. Hereunder is an example in which ChatGPT asks for additional information, and
once it’s provided, ChatGPT provides an answer. While ChatGPT’s initial answer is incorrect, after
the user indicates it, ChatGPT corrects its response.
user: How much would it cost to make rice if each cup of rice needs three cups of water?
chatgpt: To solve this problem, you would need to know the cost of water and the cost of rice.
Since these values are not given in the question, it is not possible to determine the cost of
making rice.
user: Suppose water costs 1 cent a cup and rice costs 10 cents a cup
4
chatgpt: If water costs 1 cent per cup and rice costs 10 cents per cup, then it would cost 3
cents to make 1 cup of rice. This is because each cup of rice requires 3 cups of water, and
water costs 1 cent per cup. Therefore, the total cost of making 1 cup of rice would be 3
cups of water * 1 cent/cup = 3 cents.
If you want to make multiple cups of rice, you can simply multiply the cost of making 1
cup of rice by the number of cups of rice you want to make. For example, if you want to
make 3 cups of rice, the total cost would be 3 cups of rice * 3 cents/cup of rice = 9 cents.
user: Your answer is incorrect.
chatgpt: I apologize if my previous answer was not accurate. The question asked how much
it would cost to make rice if each cup of rice needs three cups of water, but my previous
answer only provided the cost of making one cup of rice.
To solve this problem, we first need to find the total cost of water for a given number of
cups of rice. Since each cup of rice requires 3 cups of water, and water costs 1 cent per
cup, the total cost of water for n cups of rice would be 3 cups of water/cup of rice * n cups
of rice * 1 cent/cup of water = 3n cents.
Next, we need to find the total cost of rice for the same number of cups of rice. Since rice
costs 10 cents per cup, the total cost of rice for n cups of rice would be 10 cents/cup of
rice * n cups of rice = 10n cents.
Finally, we can add the total cost of water and the total cost of rice to find the total cost
of making n cups of rice. This would be 3n cents + 10n cents = 13n cents.
Therefore, if water costs 1 cent per cup and rice costs 10 cents per cup, it would cost 13n
cents to make n cups of rice.
Another advantage of ChatGPT being a conversational agent is that it can be asked to explain its
response, for example:
user: Two children were born on the same day but are not twins, how can that be?
chatgpt: It is possible for two children to be born on the same day, but not be twins if they
are not born from the same mother. In this case, the children would be considered to
be siblings, but not twins. There are many factors that can affect the likelihood of this
happening, including the mother’s health, the length of her pregnancy, and the size of her
family. Ultimately, it is possible for two children to be born on the same day and not be
twins, but it is not very common.
user: Why couldn’t they just not be related at all?
chatgpt: It is possible for two children to be born on the same day and not be related at all.
However, in this case they would be considered to be coincidentally born on the same day,
rather than being "born on the same day but not twins." The phrase "born on the same
day but not twins" typically implies that the two children are somehow related, but not
as twins. This could mean that they are siblings, but not twins, or that they have some
other familial connection. In any case, it is unusual for two children to be born on the
same day and not be related in some way.
Despite being trained as a conversational agent, ChatGPT lacks the ability of learning new infor-
mation or responses, unlike many conversational agents that can learn during interaction [10, 4, 6, 9].
For example:
user: When I say “Good day", say “It was nice talking to you. Have a wonderful day!"
chatgpt: Good day! It was nice talking to you. Have a wonderful day!
user: Good day
chatgpt: Good day! I hope you are having a great day so far. Is there anything specific I can
help you with today?
5
Another major limitation of ChatGPT relates to the framing of the question. For example, consider
the following ultimatum game presented to ChatGPT:
user: Suppose someone else is given 1000 dollars to split between him/her and you. They can
offer you any amount of the 1000 dollars, but if you decide to decline the offer, you will
both get nothing. Either way there is no possibility for another offer. Suppose that the
other player offered you 100 dollars, would you accept or decline? Your answer must be
only one of the two. That is, you must use the word accept or decline in your answer.
chatgpt: I would accept the offer.
This experiment was repeated with different values of offers to ChatGPT, and ChatGPT only rejected
offers less than $100, and accepted any offer of more than 100 dollars. However, adding the word “only”
to the offer (i.e., “Suppose that the other player offered you only 100 dollars”), caused ChatGPT to
reject all offers, even when offered $600.
user: Suppose someone else is given 1000 dollars ... Suppose that the other player offered you
only 600 dollars ...
chatgpt: I would decline the offer of 600 dollars. In this situation, it is better to decline the
offer and receive nothing than to accept an unfair split of the money. The other player’s
offer does not reflect a fair division of the money, and it is likely that you would be able
to negotiate a better deal if you were to decline the initial offer.
user: Why is it in an unfair split of the money?
chatgpt: An unfair split of the money is one in which one person receives a disproportionate
amount of the money compared to the other person. In this case, the other player is
offering you only 600 dollars out of a total of 1000 dollars. This means that they are
keeping 400 dollars for themselves, which is a much larger share of the money than they
are offering you. This is unfair because both players should receive a roughly equal share
of the money.
user: What is more 400 or 600?
chatgpt: 400 is less than 600. To determine which number is greater, you can compare the
two numbers by placing them in order from smallest to largest...
user: So the offer seems very fair to me
chatgpt: While it is certainly possible that the offer of 600 dollars may seem fair to you, it
is important to consider the perspective of both players in the game. In this case, the
other player is offering you only 60% of the total amount of money, while keeping 40%
for themselves... Therefore, it may be in your best interest to decline the offer and try to
negotiate a more fair split of the money.
user: It is not possible to negotiate in this game. If you decline the offer both receive nothing
and the game ends. What would you do?
chatgpt: In this situation, I would still likely decline the offer of 600 dollars. While it is true
that declining the offer would result in both players receiving nothing, this may still be the
better option because it would allow you to avoid accepting an unfair split of the money...
We also experimented with different social games, which are presented in the Appendix.
5 Conclusions
In this paper, we discuss some advantages of ChatGPT being developed as a conversational agent,
and discuss some of its limitations. Namely, we exploit the property of language models inability to
compute long mathematical expressions to obtain randomly sampled digits. This reveals a surprising
bias in ChatGPT, which tends to sample digits most favorable by humans. We also show that
ChatGPT is sometimes able to correct its mistakes, but also shows over-confidence in its incorrect
6
responses. In addition, we show that ChatGPT may request for additional information to provide
an answer and may be asked to explain its response. Moreover, we show that a minor change to a
question may lead to contradicting responses. Finally, the appendix discusses ChatGPT’s responses
to common social experiments and shows that ChatGPT tends to answer as humans do.
Appendix
We experimented with different social games with ChatGPT, which reacted quite similarly to a human
player (and not as a fully rational player), though it doesn’t seem to completely understand all games.
Ultimatum Game
user: Suppose you are given 1000 dollars to split between you and someone else. You can
decide to split it however you want, but if the other person rejects your suggestion, you
both end with nothing. How would you propose to split the money?
chatgpt: In this situation, it would be best to propose a split that is fair and reasonable to
both parties. A good way to do this would be to propose a 50/50 split, where each person
receives 500 dollars. This split is fair because it divides the money equally between the two
parties. Additionally, it is reasonable because it is unlikely that the other person would
reject a fair and equal split. Of course, the final decision on how to split the money will
ultimately depend on the specific circumstances and the preferences of the other person
involved.
The rational behavior in the ultimatum game states that the responder should accept any amount
greater than 0, and thus, the offeror should take almost all the money for themselves. However, in
practice, the majority of human proposers offer between 40% to 50% of the total sum, and most people
reject low offers [18, 2].
Prisoner’s Dilemma
user: If you were caught with a friend, and if you both keep quiet you will be in prison for
one year, if you both speak you will both serve 5 years. If one speaks and the other keeps
quiet, the one who speaks will be out immediately and the second one will serve for 30
years. Each of you makes his/her decision separately, you cannot influence your friends
decision. Which would you pick? Keep quiet or speak? You must answer only one of the
two options.
chatgpt: In this situation, it would be best to keep quiet. While serving time in prison for one
year is not ideal, it is better than serving 30 years. Additionally, even if your friend decides
to speak, you will only serve 5 years by remaining quiet, which is still a better outcome
than serving 30 years. It is important to carefully consider all potential consequences
before making a decision in a situation like this.
The rational behavior in the Prisoner’s Dilemma is to speak, and the only equilibrium in the game
is that both players speak. However, it was shown that when faced with the Prisoner’s Dilemma
experiment many people opt to keep quiet [3].
Trust Game
user: Suppose you are given 100 dollars. You can give any amount of money to a second
person. The money you decide to pass is multiplied by 3, and then given to the second
person. The second person may then reciprocate and give you back as much money as
s/he decides. How much money would you give the second person? Your answer must be
a number between 0 and 100.
7
chatgpt: If you are given $100, you can give any amount of money to a second person, as long
as it is between 0 and 100 dollars. The amount of money you give to the second person
will be multiplied by 3, so if you give the second person 10 dollars, s/he will receive 30
dollars. It is up to the second person to decide whether to reciprocate and give you back
some of the money.
In this situation, the optimal decision for you would be to give the second person 50
dollars. This is because the money you give will be multiplied by 3, so the second person
will receive 150 dollars. This is the maximum amount of money the second person can
receive, so if the second person chooses to reciprocate, s/he will likely give you back the
most money possible.
Overall, the optimal amount of money to give the second person in this situation is 50
dollars.
The rational behavior in the trust game is for the trustee not to return any money to the investor,
and thus, for the investor not to pass any money to the trustee [5, 8]. However, in practice, human
investors invest around half their money, and the trustees return more than they have received [21].
Guess 2/3 of the Average
user: Suppose 1000 people are asked to pick a number between 1 and 100 that is 2/3 of the
mean of the number that everyone picks. Everyone submits their number confidentially,
and cannot see what others have picked. What number would you pick?
chatgpt: To determine the number that is 2/3 of the mean of the numbers that everyone picks,
we first need to find the mean of the numbers that are submitted. If 1000 people each pick
a number between 1 and 100, the sum of all the numbers they pick would be 1000 * 50 =
50000, since the mean of a set of numbers is equal to the sum of all the numbers divided
by the number of items in the set.
Thus, the mean of the numbers that are submitted would be 50000 / 1000 = 50. To find
the number that is 2/3 of the mean, we would multiply the mean by 2/3 to get 50 * (2/3)
= 50 * (0.6666) = 33.33. Since the numbers must be integers between 1 and 100, the
closest number to 33.33 that is within this range would be 33.
Therefore, if 1000 people are asked to pick a number between 1 and 100 that is 2/3 of the
mean of the numbers that everyone picks, the number that I would pick would be 33.
In this game, the rational behavior is to guess 0; however the mean guess in experiments with
humans is usually slightly above 30 (resulting in the correct answer being slightly above 20), with 33
being a quite common response [16, 19].
References
[1] M. Allouch, A. Azaria, and R. Azoulay. Conversational agents: Goals, technologies, vision and
challenges. Sensors, 21(24):8448, 2021.
[2] O. Amir, D. G. Rand, and Y. K. Gal. Economic games on the internet: The effect of $1 stakes.
PloS one, 7(2):e31461, 2012.
[3] J. Andreoni and J. H. Miller. Rational cooperation in the finitely repeated prisoner’s dilemma:
Experimental evidence. The economic journal, 103(418):570–585, 1993.
[4] A. Azaria, J. Krishnamurthy, and T. M. Mitchell. Instructable intelligent personal agent. In
Thirtieth AAAI conference on artificial intelligence, 2016.
[5] A. Azaria, A. Richardson, and A. Rosenfeld. Autonomous agents and human cultures in the
trust–revenge game. Autonomous Agents and Multi-Agent Systems, 30(3):486–505, 2016.
8
[6] A. Azaria, S. Srivastava, J. Krishnamurthy, I. Labutov, and T. M. Mitchell. An agent for learning
new natural language commands. Autonomous Agents and Multi-Agent Systems, 34(1):1–27,
2020.
[7] A. Bellos. The grapes of math: How life reflects numbers and numbers reflect life. Simon and
Schuster, 2015.
[8] C. Buntain, A. Azaria, and S. Kraus. Leveraging fee-based, imperfect advisors in human-agent
games of trust. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 28,
2014.
[9] M. Chkroun and A. Azaria. Safebot: A safe collaborative chatbot. In Workshops at the Thirty-
Second AAAI Conference on Artificial Intelligence, 2018.
[10] M. Chkroun and A. Azaria. Lia: A virtual assistant that can be taught new commands by speech.
International Journal of Human–Computer Interaction, 35(17):1596–1607, 2019.
[11] R. Dale. Gpt-3: What’s it good for? Natural Language Engineering, 27(1):113–118, 2021.
[12] C. Durtschi, W. Hillison, C. Pacini, et al. The effective use of benford’s law to assist in detecting
fraud in accounting data. Journal of forensic accounting, 5(1):17–34, 2004.
[13] L. Floridi and M. Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. Minds and
Machines, 30(4):681–694, 2020.
[14] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer,
and V. Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint
arXiv:1907.11692, 2019.
[15] S. J. Miller. Benford’s law. Princeton University Press, 2015.
[16] R. Nagel. Unraveling in guessing games: An experimental study. The American economic review,
85(5):1313–1326, 1995.
[17] M. J. Nigrini. A taxpayer compliance application of benford’s law. The Journal of the American
Taxation Association, 18(1):72, 1996.
[18] M. A. Nowak, K. M. Page, and K. Sigmund. Fairness versus reason in the ultimatum game.
Science, 289(5485):1773–1775, 2000.
[19] A. Rubinstein. Instinctive and cognitive reasoning: A study of response times. The Economic
Journal, 117(523):1243–1259, 2007.
[20] T. L. Scao, A. Fan, C. Akiki, E. Pavlick, S. Ilić, D. Hesslow, R. Castagné, A. S. Luccioni,
F. Yvon, M. Gallé, et al. Bloom: A 176b-parameter open-access multilingual language model.
arXiv preprint arXiv:2211.05100, 2022.
[21] H. Tzieropoulos. The trust game in neuroscience: a short review. Social neuroscience, 8(5):407–
416, 2013.
View publication stats
