Reputational Algorithm Aversion∗

Gregory Weitzner
arXiv:2402.15418v3 [[Link]] 31 Jul 2024

July 2024

Abstract

People are often reluctant to incorporate information produced by algorithms into

their decisions, a phenomenon called “algorithm aversion”. This paper shows how
algorithm aversion arises when following or overriding an algorithm conveys in-
formation about a human’s ability. In my model, workers make forecasts of an
uncertain outcome based on their private information and an algorithm’s signal.
Although it is efficient for low-skill workers to follow the algorithm, they sometimes
override it due to reputational concerns. The model provides a fully rational mi-
crofoundation for algorithm aversion that aligns with the broad concern that AI
systems will displace many types of workers.

∗
McGill University. Email: [Link]@[Link]. I thank Jeremy Bertomeu, Adolfo De Motta,
Vincent Glode and seminar participants at the Columbia/RFS AI in Finance Conference for the helpful
comments and discussions.

1
1 Introduction

In recent years, algorithms have become better at forecasting many types of outcomes

than humans, raising concerns that artificial intelligence will cause the mass displace-
ment of jobs.1 However, many argue that these concerns are overstated: instead of being
displaced by algorithms, humans will play a critical role in collaborating with and/or
supervising algorithms.2 Based on this idea, a rapidly growing literature across com-

puter science, economics and psychology explores how humans work alongside artificial
intelligence systems.
However, despite the clear value of algorithms, humans are often reluctant to incorpo-
rate algorithmic forecasts into their own decisions, a phenomenon which Dietvorst, Simmons, and Massey

(2015) coin “algorithm aversion”. Algorithm aversion has been documented empirically in
domains as disparate as health (e.g., Promberger and Baron (2006), Longoni, Bonezzi, and Morewedge
(2019), Shaffer et al. (2013) and Agarwal et al. (2023)), finance (e.g., Önkal et al. (2009),
Niszczota and Kaszás (2020) and Greig et al. (2023)) and judicial decision making (e.g.,

Angelova, Dobbie, and Yang (2023)). In some cases, even after being provided the algo-
rithm’s forecast, humans’ forecasts perform worse than the algorithm on its own. At first
blush, this behavior is surprising. If humans act as rational Bayesians, their forecasts
should always be at least as accurate as an algorithm’s. For this reason, the main ex-
planations for algorithm aversion center on behavioral biases and psychological reasons.3

Given the large potential productivity benefits of AI-human collaboration, it is critical

to understand why exactly humans are so reluctant to incorporate valuable information
from algorithms into their own decision making. Moreover, understanding the reasons for
algorithm aversion can also help in design more efficient systems in which humans work
1
I follow Ludwig and Mullainathan (2021) by using the phrase algorithm interchangably with “ar-
tificial intelligence”, “machine learning” or “deep learning”. An incomplete list of papers that show
algorithms outperform humans across many domains include: Yeomans et al. (2019), Lai et al. (2021),
Mullainathan and Obermeyer (2019), Liu (2022) and Agarwal et al. (2023).
2
E.g., Langlotz (2019), Agrawal, Gans, and Goldfarb (2019), Ludwig and Mullainathan (2021)
Acemoglu, Autor, and Johnson (2023). For example, Ludwig and Mullainathan (2021) argue that in
many contexts, humans can supplement algorithm’s because they have better context/can better assess
rare situations by using intuition.
3
See Burton, Stein, and Jensen (2020), Jussupow, Benbasat, and Heinzl (2020) and Mahmud et al.
(2022) for review articles discussing the various explanations for algorithm aversion.

2
alongside algorithms.
In this paper, I argue that a simple, yet overlooked, reason for algorithm aversion is
that a human’s choice to follow or override an algorithm’s forecast may convey information
about that human’s ability. To illustrate this idea, consider a situation in there is worker

forecasting a binary outcome with the help of an algorithm provided by the firm. After
observing the algorithm’s forecast, the worker makes his own forecast and the outcome
is realized. The worker has private information about his skill, where he can either be
high-skill, in which case he is better at predicting the outcome than the algorithm, or he
can be low-skill, in which case he is worse than the algorithm. Hence, it is efficient for

the low-skill worker to report the algorithm’s forecast, while the high-skill worker report
his own information. This paper shows that the efficient use of the algorithm can never
be achieved. Why? If the algorithm is used efficiently, only the high-skill worker ever
overrides the algorithm. However, if this were the case, the worker would always override

the algorithm to convince the firm they are high-skill. Hence, in equilibrium the low-skill
worker must at least sometimes inefficiently override the algorithm. Put in more simple
terms, in order for a worker to convince his boss that he is worth keeping around, he
needs to sometimes override the algorithm, even if doing so is the wrong decision for the

company.
Formally, I analyze a strategic communications, i.e., cheap talk, game in which a
worker is tasked with predicted an uncertain outcome. The worker receives a private
signal as well as a public signal produced by an algorithm. After receiving both signals,
the worker reports a forecast, after which the outcome is realized. Importantly, the worker

has private information about his own skill. Specifically, the worker can either be low-skill,
in which case his information is always inferior to the algorithm, or he can be high-skill
in which case his information is always superior to the algorithm. The firm prefers the
worker report an accurate forecast; hence, from the firm’s perspective, it is efficient for

the low-skill worker to report the algorithm’s forecast, while the high-skill worker always
reports his own signal. After receiving the worker’s forecast and the realization of the
outcome, the firm updates its beliefs regarding the worker’s skill. Importantly, the worker

3
faces reputational concerns (e.g., Holmström (1999)) and hence reports a forecast which
maximizes the firm’s perception that he is high-skill.
I first consider a benchmark case with no algorithm. In this benchmark, there always
exists an efficient equilibrium in which the worker honestly reports his signal. Intuitively,

the firm assesses the worker entirely on the accuracy of his forecast which incentivizes
the worker to report his signal truthfully in order for his forecast to be perceived as
accurate as possible. In the main model in which the worker has access to the algorithm’s
signal, this is no longer the case. Here it is efficient for the low-skill worker to report
the algorithm’s forecast, while the high-skill worker should always report his own signal.

However, whereas in the benchmark case the forecast on its own provided no information
regarding the worker’s skill, here the firm learns both from the accuracy of the worker’s
forecast and the forecast itself when compared to the algorithm’s signal. In fact, because
of this latter effect the first-best use of information can never be achieved. The reason

behind this result is simple: if the high-skill worker is the only type to ever override
the algorithm, overriding the algorithm will instantly reveal that the worker is high-skill.
Hence, under this set of beliefs, both worker types would always deviate by reporting the
opposite signal of the algorithm.

I next show that there exists an informative equilibrium in which the high-skill worker
always reports his signal and the low-skill worker reports his signal with some positive
probability when his signal differs from the algorithm. In this equilibrium, the worker
exhibits “algorithm aversion”, i.e., he overrides the algorithm even when his own signal is
less accurate than the algorithm. Hence, algorithm aversion causes the overall accuracy

of the worker’s forecast to be strictly lower than the first-best. In fact, in some cases
the accuracy of the worker’s forecast is worse than the algorithm alone and the worker
provides at best zero static benefits to the firm.
I next show that algorithm aversion increases in the uncertainty of the algorithm.

Intuitively, as the algorithm becomes more uncertain it becomes less costly to override
it because its signal is less accurate. However, there is also an interesting second effect.
Because the algorithm is less accurate, and the high-skill worker always reports his own

4
signal, the high-skill worker’s signal differs from the algorithm more often and hence he
defies the algorithm more frequently in equilibrium. This increases the signaling value
of overriding the algorithm for the low-skill worker. Hence, this indirect effect causes an
even larger increase in algorithm aversion.

Beyond generating a rational microfoundation for algorithm aversion, the model’s

predictions are consistent with substantial empirical evidence. First, humans’ forecasts
are often less accurate than an algorithm’s even after being given access to the algorithm’s
forecast (e.g., Angelova, Dobbie, and Yang (2023) and Agarwal et al. (2023)), which can
occur enogenously in my model for completely rational reputational reasons. Relatedly,

Angelova, Dobbie, and Yang (2023) analyze judges’ discretion over algorithm and find
that 90% of judges underperform an algorithm when they override it’s recommendation.
In my model, the worker overrides the algorithm too frequently and the accuracy of his
forecast in the case of an override is often lower than that of the algorithm. While part

of Angelova, Dobbie, and Yang (2023)’s results could be due to judges simply making
mistakes, my model shows how override decisions can also be driven by reputational
concerns. If a worker never override an algorithm, it can be difficult for them to justify
keeping their job.

The main mechanism of my model is that by relying on an algorithmic recommen-

dation, a human conveys negative information about their ability. Consistent with this
prediction, Shaffer et al. (2013) show that patients rate doctors using computerized rec-
ommendations lower than those that do not rely on computerized recommendations.4
While data about the use of algorithms within firms is limited, the model would predict

that the same type of dynamic would exist for workers relying on algorithmic forecasts
within firms.
To my knowledge this is the first model providing a mechanism for algorithm aver-
sion which is consistent with rational Bayesian behavior. A non-exhaustive list of other

explanations include i) the desire to understand forecasts (e.g., Yeomans et al. (2019)),
4
Leyer and Schneider (2019) show that humans are less likely to delegate decisions to an AI system
versus another human. Although there is no delegation decision in my model, if such a delegation decision
reflects negatively on the human, my model would also predict a reluctance to do so.

5
ii) valuing agency of decisions (e.g., Sunstein (2023)), iii) humans having incorrect priors
(e.g., Yeomans et al. (2019), Greig et al. (2023)) and iv) humans deriving utility from
interacting with another human and/or disutility from interacting with a machine (e.g.,
Dietvorst, Simmons, and Massey (2015) and Greig et al. (2023)).5 While I do not doubt

that psychological and behavioral reasons may partially explain this phenomenon, I argue
that a simple, yet missing part of the story is the reputational concerns of the people
using these algorithms. More broadly, by highlighting an impediment of humans working
alongside AI systems, this paper contributes to the literature on human-AI collaboration.6
In the model, algorithm aversion makes the worker’s forecast less accurate and in

some cases the worker provides zero value to the firm in the short-run. In practice, this
can put downward pressure on wages for humans. However, if firms are are impatient
and it is costly to retain/train workers, reputational algorithm aversion could result in
firms firing workers entirely. For these reasons, this paper relates to the literature an-

alyzing the use of AI within firms and in AI’s affect on labor market outcomes (e.g.,
Acemoglu and Restrepo (2018), Acemoglu and Restrepo (2020), Acemoglu et al. (2022),
Cao et al. (2023), Babina et al. (2024), Jha et al. (2024) and Eisfeldt, Schubert, and Zhang
(2023)).

Finally, from a theory perspective, the model follows the reputational cheap-talk
framework (e.g., Scharfstein and Stein (1990), Ottaviani and Sørensen (2006a) Ottaviani and Sørensen
(2006b), Ottaviani and Sørensen (2006c), Guembel and Rossetto (2009)) which combines
cheap talk (Crawford and Sobel, 1982) with career concerns (Holmström, 1999). Like
other reputational cheap talk models, agents make strategic forecasts in order to max-

imize the likelihood that they are perceived as skilled. However, there are two main
differences between my model and typical reputational cheap-talk models. First, in my
model the worker knows his type, whereas typically the sender’s type is unknown to him-
self. Second, the worker has access to both a private signal and a public signal (i.e., the

algorithm). These two ingredients cause the worker’s forecast itself to convey information
5
See Jussupow, Benbasat, and Heinzl (2020), Burton, Stein, and Jensen (2020) and Mahmud et al.
(2022) for review articles of algorithm aversion.
6
See Lai et al. (2021) for an excellent review of this literature.

6
about his ability beyond the accuracy of the forecast, which results in workers placing
too little weight on the public signal due to reputational concerns.

2 Model Setup

There is a worker who is tasked by his manager to predict an outcome regarding a state
of the world ω ∈ {ω0 , ω1 }. The worker is of type θ ∈ {θL , θH } (low-skill or high-skill) and

receives a private signal s ∈ {s0 , s1 } regarding ω, where both the worker’s type and signal
are only known by the worker. The worker is also given access to an algorithm developed
by the firm, which produces a signal a ∈ {a0 , a1 } regarding ω, where the algorithm’s
signal is observable to both the worker and the manager. I assume the algorithm is a

“black-box” which obtains information regarding ω by analyzing a set of training data

which I do not explicitly model.
The worker observes both signals then reports a message m ∈ {m0 , m1 }, i.e., forecast,
to the manager, where m = m0 means (in equilibrium) that the worker predicts that
ω = ω0 .7 After the worker reports his message the state ω is realized and observed by all.

Because the worker’s actions are costless, the model applies to any situation in which
the worker’s action serves as a forecast, potentially conveying information about his
skill. For example, it could be the worker is forecasting the firm’s earnings or cash
flows, but could also be be that the worker is making an investment decision (e.g.,

Scharfstein and Stein (1990)).8

2.1 Information Structure

1
The worker is equally likely to be low-skill or high-skill, i.e., P r(θH ) = P r(θL ) = 2
and
both states of the world are equally likely, P r(ω1) = P r(ω0 ) = 21 . When ω = ω1 , the
high-skill worker receives the signal s1 with probability υH , while the low-skill worker
7
Because this is a cheap-talk model, the messages only have meaning to the extent they convey
information to the manager in equilibrium.
8
In Scharfstein and Stein (1990) the outcome of the investment decision is observable whether or not
the firm invests. Hence, this interpretation would be most relevant to situations in which the the firm
can tell whether the action would have succeeded regardless of whether the action was taken, e.g., if the
investment outcome is mostly tied to market-wide conditions.

7
receives it with probability υL < υH . When ω = ω0 , the high-skill worker receives the
signal s1 with probability 1 − υH and the low-skill worker receives it with probability
1 − υL :

1
P r(s1 |ω1 , θL ) ≡ υL > ,
2
P r(s1 |ω1, θH ) ≡ υH > υL ,

P r(s1 |ω0 , θL ) ≡ 1 − υL ,

P r(s1 |ω0, θH ) ≡ 1 − υH .

These assumptions make it such that the high-skill worker always has more accurate
information than the low-skill worker and that the ex-ante distribution of signals is the
same for both low-skill and high-skill workers, i.e., P r(s1|θH ) = P r(s1|θL ) = 21 . The latter
simplifies the problem by ensuring that the signal itself does not convey any information

about the skill of the worker.9

When ω = ω1 , the algorithm produces the signal a1 from the algorithm with probabil-
ity α, while if ω = ω0 the algorithm produces signal a1 with probability 1−α. Importantly,
the algorithm’s recommendation does not depend on the worker’s skill. This assumption
is reasonable so long as the algorithm is developed by other employees of the firm or

outside the firm entirely. In practice, different firms may even use the same algorithm
developed by a third party.10
To make the problem interesting, I assume that the high-skill worker’s signal is more
informative than the algorithm, while the low-skill worker’s is less informative.

Assumption 1. the high-skill worker’s signal is more informative than the algorithm,
while the low-skill worker’s is less informative, i.e., α ∈ (υL , υH ).

Without this assumption, it would always be optimal for the worker to either always
report the algorithm’s signal or always report his own signal. In practice, the “ground

truth” in algorithms’ training data is often established based on “expert” assessments.

9
A similar assumption is made in Scharfstein and Stein (1990).
10
For example, firms often use the same large language models, e.g., ChatGPT.

8
Similarly, Angelova, Dobbie, and Yang (2023) find that 10% of judges outperform an
algorithm in the context of bail decisions. Hence, it is natural that at least some workers
can forecast better than the algorithm. Moreover, if the algorithm always provided the
better forecast, there would be no need to use the worker’s information at all.

2.2 Worker’s Objective and Equilibrium Concept

I assume that the manager always prefers that the worker’s forecast is correct but do not
explicitly model why this is valuable to the manager.11 For example, the firm may prefer
that its treasurer make more accurate cash flow estimates to help guide firms’ financial
policies. Relatedly, the firm would prefer to make investments when market conditions

turn out to be favorable for that investment.

The manager first observes the worker’s forecast m and the algorithm’s signal a, after
which the outcome ω is realized. The manager then forms a posterior belief θ̂(m, a, ω)
on the probability that the worker is high-skill.12
Through assessing the worker’s forecast and the accuracy of the forecast in predict-

ing the outcome, the manager can learn about the worker’s skill. As is typical in career
concerns models, there are no long-term contracts and hence, the worker attempts to max-
imize the manager’s perceived probability that the worker is high-skill. This is equivalent
to a risk-neutral worker maximizing his next period, competitive spot wage, where the

wage is linear in the manager’s posterior belief that the worker is high-skill.13 Hence,
the worker’s payoff following his own signal s, the algorithm’s signal s and the worker’s
message m as follows:

X
P r(ω|s, θ)θ̂(m, a, ω).
ω

The equilibrium concept I use throughout the analysis is perfect Bayesian Nash Equi-
11
This is standard in reputational cheap talk models (e.g., Scharfstein and Stein (1990),
Ottaviani and Sørensen (2006a), Ottaviani and Sørensen (2006b), Ottaviani and Sørensen (2006c).
12
We can equivalently think of the manager as anyone who observes both the worker and the algorithm’s
recommendation. If the worker and algorithm’s forecast are observable to other firms, then we think of
the manager as the entire labor market (e.g., Holmström (1999)).
13
This approach is common in reputational cheap talk models (e.g., Scharfstein and Stein (1990)).

9
librium, in which the worker’s equilibrium strategy maximizes his expected reputational
payoff and the manager’s conditional probabilistic belief is correct on equilibrium path,
i.e., θ̂(m, a, ω) = P r(θH |m, a, ω) for ω ∈ {ω0 , ω1 }, a ∈ {a0 , a1 } and any m that is used in
equilibrium.

3 Benchmark: No Algorithm

First, it will be useful to establish a benchmark in which the worker does not have access
to the algorithm’s signal to highlight the mechanism throughwhich algorithm aversion
arises. Because the manager always prefers a more accurate forecast, it is efficient for the
worker to report his signal truthfully. For this analysis, it will be useful to calculate the

worker’s posterior belief regarding the likelihood of state ω:

P r(s1|ω1 , θH )P r(ω1 |θH )

P r(ω1|s1 , θH ) = = υH ,
P r(s1|θH )
P r(s0|ω1 , θH )P r(ω1 |θH )
P r(ω1|s0 , θH ) = = 1 − υH ,
P r(s0|θH )
P r(s1|ω1 , θL )P r(ω1|θL )
P r(ω1 |s1 , θL ) = = υL ,
P r(s1 |θL )
P r(s0|ω1 , θL )P r(ω1|θL )
P r(ω1 |s0 , θL ) = = 1 − υL .
P r(s0 |θL )

It will also be useful to calculate the manager’s posterior beliefs assuming that the worker

engages in truth-telling:

υH
θ̂(ω1 , s1 ) = ,
υH + υL
1 − υH
θ̂(ω1 , s0 ) = ,
2 − υH − υL
1 − υH
θ̂(ω0 , s1 ) = ,
2 − υH − υL
υH
θ̂(ω0 , s0 ) = .
υH + υL

10
In order for the low-skill and high-skill workers to truthfully report m1 when s = s1 , the
following incentive compatibility conditions must hold:

X X
P r(ω|s1, θL )θ̂(s1 , ω) ≥ P r(ω|s1, θL )θ̂(s0 , ω) (1)
ω ω
X X
P r(ω|s1, θH )θ̂(s1 , ω) ≥ P r(ω|s1, θH )θ̂(s0 , ω).
ω ω

I next show that in this benchmark case, there always exists an equilibrium in which the

worker reports his signal as his forecast.

Proposition 1. In the benchmark case without the algorithm, a truth-telling equilibrium

always exists in which both the low-skill worker and high-skill worker truthfully report
their signal, i.e., m = m0 if s = s0 and m = m1 if s = s1 .

Proof. Because the problem is symmetric, the incentives to deviate will always be the
same regardless of the signal the worker receives. Hence, it is without loss of generality to
consider the incentives to deviate following signal s1 . Plugging in the posteriors calculated

above into (1) and simplifying we have:

(υH − υL ) (2υL − 1)
,
(2 − υH − υL ) (υH + υL )

which is positive given that υH > υL and υL > 21 . Hence, the low-skill worker will not
deviate. Plugging in the posteriors into (1) and simplifying we have:

(υH − υL ) (2υH − 1)
,
(2 − υH − υL ) (υH + υL )

which is also clearly positive. Hence, the high-skill worker will also not deviate from

truth-telling. 

Intuitively, because the worker’s forecast itself does not convey any information about
the worker’s ability, the worker can only convey his skill by making as accurate of a

forecast as possible. Hence, it is incentive compatible for the worker to report his signal
truthfully if the manager expects him too. However, once I introduce the algorithm, the

11
worker’s forecast will convey information about his skill because it can be compared to
the algorithm’s signal.
It is worth mentioning that as in all cheap talk models, there also exists a “babbling”
equilibrium, in which both high and low-skill workers report a forecast m with the same

probability regardless of their type and the signal they receive. In this case, the man-
ager cannot glean any information from the worker’s message and hence, the worker is
indifferent between any potential message he sends.

4 Algorithm

Next, I analyze the case in which the worker has access to the algorithm before making

his forecast. Because the problem is symmetric, throughout the analysis I focus on the
case in which the worker receives signal s1 ; however, the analysis is the same for the
case in which s = s0 . Again, it will be useful to calculate the worker’s posterior belief
regarding the likelihood of the state:

P r(s1 , a1 |ω1 , θL )P r(ω1|θL ) αυL

P r(ω1|s1 , a1 , θL ) = = ,
P r(s1 , a1 |θL ) αυL + (1 − α)(1 − υL )
P r(s1, a1 |ω1 , θH )P r(ω1|θH ) αυH
P r(ω1|s1 , a1 , θH ) = = ,
P r(s1, a1 |θH ) αυH + (1 − α)(1 − υH )
P r(s1 , a0 |ω1 , θL )P r(ω1|θL ) (1 − α)υL
P r(ω1|s1 , a0 , θL ) = = ,
P r(s1 , a0 |θL ) (1 − α)υL + α(1 − υL )
P r(s1, a0 |ω1 , θH )P r(ω1|θH ) (1 − α)υH
P r(ω1|s1 , a0 , θH ) = = .
P r(s1, a0 |θH ) (1 − α)υH + α(1 − υH )

Based on these posterior probabilities, the manager would always prefer that the low-skill
worker report the algorithm’s forecast rather than his own signal, while the high-skill

worker reports his own signal. To see this, first note that from Assumption 1, υH >
α > υL > 21 . Hence, when the worker’s signal coincides with that of the algorithm, the
posterior probability that the outcome coincides with the worker’s signal is always greater
1
than one half, i.e., P r(ω1 |s1 , a1 , θ) > 2
for θ ∈ {θL , θH }. However, when the worker’s

signal differs from the algorithm, the high-skill worker’s posterior probability that the
1
outcome coincides with the signal is greater than one-half, i.e., P r(ω1|s1 , a0 , θH ) > 2
,

12
while the low skill’s worker is less than one-half, i.e., P r(ω1 |s1 , a0 , θL ) < 21 . Hence, for
the forecast to be as accurate as possible, the manager always prefers that the high-skill
worker report his own signal, while the low-skill worker always report the algorithm’s
signal.

Henceforth, I refer to the case in which the worker reports the algorithm’s signal
when he is low-skill and his own signal when he is high-skill the “first-best”. I next
show that because of reputational concerns, the first-best can never be achieved as an
equilibrium.

Proposition 2. The first-best cannot be achieved as an equilibrium.

Proof. First, it is useful to calculate the following posterior beliefs given that the manager
expects the worker to follow the first-best use of information:

υH
θ̂F B (m1 , a1 , ω1 ) = ,
υH + 1
1 − υH
θ̂F B (m1 , a1 , ω0 ) = ,
1 − υH + 1
θ̂F B (m0 , a1 , ω1 ) = 1,

θ̂F B (m1 , a1 , ω) = 1.

After receiving signal s1 , incentive compatibility for the low-skill worker requires:

X X
P r(ω|s1, a1 , θL )θ̂F B (m1 , a1 , ω) ≥ P r(ω|s1, a1 , θL )θ̂F B (m0 , a1 , ω), (2)
ω ω
X X
P r(ω|s1, a0 , θL )θ̂F B (m0 , a0 , ω) ≥ P r(ω|s1, a0 , θL )θ̂F B (m1 , a0 , ω),
ω ω

where (2) says that when the low-skill worker receives the signal s1 he should report m1

when the algorithm’s signal is a1 and m0 when the algorithm’s signal is a0 . Plugging in

13
the first-best posterior beliefs we have:

X υH X
P r(ω|s1, a1 , θL ) ≥ P r(ω|s1, a1 , θL ), (3)
ω
υ H + 1 ω
X 1 − υH X
P r(ω|s1, a0 , θL ) ≥ P r(ω|s1, a0 , θL ).
ω
1 − υ H + 1 ω

υH 1−υH
Since, υH +1
and 1−υH +1
are both less than one (3) is violated. Hence, the first-best is not
incentive compatible for the low-type and cannot be sustained in equilibrium.14 

Intuitively, under the first-best only the high-skill worker ever overrides the algorithm.
Because of this, by overriding the algorithm the worker will be perceived as high-skill with
probability one. Hence, both the low-skill and high-skill workers would have incentives
to override the algorithm no matter what their signal is in order to be perceived as high-

skill. As shown below, there exists an equilibrium in which the low-skill worker sometimes
overrides the algorithm even though it is inefficient for him to do so. Before presenting
the proposition, I establish the following definition regarding a “informative equilibrium”,
in which the distribution of the worker’s message depends on the worker’s type:15

Definition 1. In a informative equilibrium, the manager’s posterior belief that the worker
is high-skill is always higher when he correctly forecasts the state, i.e., θ̂(m1 , a, ω1) >
θ̂(m0 , a, ω1 ) and θ̂(m0 , a, ω0) > θ̂(m1 , a, ω0 ) for a ∈ {a0 , a1 }.

This definition is useful because it rules out any babbling equilibria in which the the low

and high-skill worker always follow the algorithm, override it, or randomize independently
of their own signal.16
Based on this definition, I next prove the following lemma which says the high-skill
worker will always report his signal in any informative equilibrium.

Lemma 1. The high-skill worker always reports his signal in an informative equilibrium.

14
It can easily be shown the same is true for the high-type.
15
This definition follows closely the definition in Guembel and Rossetto (2009). The only difference is
that it assumes that the manager interprets m1 as ω1 is more likely and m0 as ω0 is more likely. This is
however, without loss of generality given that the messages have no intrinsic meaning to begin with.
16
Intuitively, the manager does not learn anything from the worker’s forecast and subsequent state
realization if the joint distribution of the worker’s forecast and realization is independent of his skill.

14
Proof. See Appendix. 

The intuition of Lemma 1 is straightforward. For a given set of the manager’s posterior
beliefs, it is always less costly for the high-skill worker to follow his own signal because
his signal is more informative about the state than the low-skill worker’s. Hence, in any
informative equilibrium, the high-skill worker will always report his own signal. It will

also be useful to show that the low skill worker never overrides the algorithm when he
receives the same signal as it.

Lemma 2. The low-skill worker always reports his own signal when his signal equals that
of the algorithm in an informative equilibrium, i.e., m = m0 when a = a0 , s = s0 and
m = m1 when a = a1 , s = s1 .

Proof. See Appendix. 

The intuition for (2) is as follows: if the low-skill worker is going to override the

algorithm, it is always less costly to do so when his signal is different than that of the
algorithm. Indeed, as shown below, the low-skill worker will sometimes override the
algorithm when his signal differs from it.
Following Lemmas 1 and 2, the following proposition characterizes the unique infor-

mative equilibrium.

15
Proposition 3 (Algorithm Aversion). There exists a unique, informative equilibrium in

which the high-skill worker always reports his own signal and the low-skill worker reports
the algorithm’s signal with probability γ ∈ (0, 1) and his own signal with probability 1 − γ,
where γ is the unique solution to the following equation:

X X
P r(ω|s1, a0 , θL )θ̂(m0 , a0 , ω) − P r(ω|s1, a0 , θL )θ̂(m1 , a0 , ω) = 0,
ω ω

where

υH
θ̂(m1 , a0 , ω1 ) = ,
υH + (1 − γ)υL
1 − υH
θ̂(m1 , a0 , ω0 ) = ,
1 − υH + (1 − γ) (1 − υL )
1 − υH
θ̂(m0 , a0 , ω1 ) = ,
(1 − υH ) + (1 − υL ) + γυL
υH
θ̂(m0 , a0 , ω0 ) = .
υH + υL + γ (1 − υL )

Proof. Because of symmetry, we can again restrict focus to the case when the worker
receives signal s1 . First consider the case when a = a1 . From Lemmas 1 and 2 both low
and high-skill workers will report m = m1 . Now consider, the case in which a = a0 . The
low-skill worker has the option to report his own signal or the algorithm’s. Let γ denote

the probability that the low-skill worker reports the algorithm’s signal, i.e., m = m0 . We
can calculate the manager’s posterior beliefs as follows:

υH
θ̂(m1 , a0 , ω1 ) = ,
υH + (1 − γ)υL
1 − υH
θ̂(m1 , a0 , ω0 ) = ,
1 − υH + (1 − γ) (1 − υL )
1 − υH
θ̂(m0 , a0 , ω1 ) = ,
(1 − υH ) + (1 − υL ) + γυL
υH
θ̂(m0 , a0 , ω0 ) = .
υH + υL + γ (1 − υL )

where the posteriors do not depend on a because the algorithm’s signal is independent of

16
the worker’s type θ and independent of the worker’s signal s conditional on the state ω.
First, it can easily be confirmed that these posterior beliefs satisfy the definition of an in-
formative equilibrium, i.e., θ̂(m1 , a0 , ω1) > θ̂(m0 , a0 , ω1 ) and θ̂(m0 , a0 , ω0 ) > θ̂(m1 , a0 , ω0 ).
Next, define G(γ) as the low-skill worker’s difference in payoffs from reporting m1 versus

m0 when he receives signal s1 and the algorithm’s signal is a0 :

X X
G(γ) ≡ P r(ω|s1, a0 , θL )θ̂(m0 , a0 , ω) − P r(ω|s1, a0 , θL )θ̂(m1 , a0 , ω).
ω ω

First note from Proposition 2, when γ = 1, we are in the first-best use of information
and G(γ) is negative, implying that both the low-skill and high-skill worker will always
report m1 . Hence γ must be less than 1 in any informative equilibrium. Differentiating
G(γ) w.r.t to γ we have:

2 2
(1−α)υH υL α(1−υH )(1−υL )2 αυH (1−υL )2 (1−α)(1−υH )υL
(υH +(1−γ)υL )2
+ (2−γ−υH −(1−γ)υL )2
+ (γ+υH +(1−γ)υL )2
+ (2−υH −(1−γ)υL )2
G′ (γ) = − , (4)
α − (2α − 1)υL

which is negative because both the numerator and denominator are positive, implying

that G(γ) is decreasing in γ. When γ = 0 we have:

(α − υL ) (υH − υL )
G(0) = ,
(2 − υH − υL ) (υH + υL ) (α − (2α − 1)υL )

which is positive. Hence, there exists a unique γ such that G(γ) = 0. Note that whenever
a = a0 , the low-skill worker must be indifferent between reporting m0 and m1 . Hence,
this is the unique informative equilibrium. 

Proposition 3 shows that, despite the algorithm providing a more informative signal

than the low-skill worker’s, the low-skill worker sometimes overrides it when his own signal
is different than that of the algorithm’s. Intuitively, the low-skill worker cannot always
follow the algorithm, otherwise overriding the algorithm signals with probability one that
he is high-skill. In other words, the worker must override the algorithm occasionally in

order to convince their manager they are not dispensable. Hence, Proposition 3 shows how
completely rational reputational concerns can endogenously create algorithm aversion.

17
 Agarwal et al. (2023) show that radiologists are less likely to follow an algorithm when
they believe that the algorithm’s forecast is more uncertain. Motivated by this empirical
finding, I next show that the more uncertain the algorithm’s signal is, the more likely the
low-skill worker overrides the algorithm in the unique informative equilibrium.

Proposition 4. Algorithm aversion increases in the uncertainty of the algorithm, i.e., γ

is increasing in α in the informative equilibrium.

Proof. Applying the implicit function theorem we have:

dγ ∂G(γ)/∂α
=− ,
dα ∂G(γ)/∂γ

where the denominator is negative from (4) and the numerator is equal to:

X ∂P r(ω|s1, a0 , θL ) X ∂P r(ω|s1, a0 , θL )
θ̂(m0 , a0 , ω) − θ̂(m1 , a0 , ω)
ω
∂α ω
∂α
(1 − υL ) υL  
= θ̂(m0 , a0 , ω 0 ) − θ̂(m0 , a0 , ω 1 ) + θ̂(m1 , a0 , ω 1 ) − θ̂(m1 , a0 , ω 0 ) ,
(α + (2α − 1)υL ) 2

where first term is positive and the second term is also positive based on the definition
dγ
of an informative equilibrium. Hence, dα
is positive and γ is increasing in α in the
informative equilibrium.


This result may at first glance seem obvious. If the algorithm is less accurate, the low-skill

worker should be less likely to follow it. However, from the manager’s perspective so long
as Assumption 1 is still satisfied, it is always efficient for the low-skill worker to report
the algorithm’s signal. Moreover, there are two effects here. On the one hand, it is less
costly to override the algorithm given that the algorithm’s signal is uninformative. On

the other hand, the high-skill worker also becomes more likely to override the algorithm
which increases the signaling value of overriding the algorithm for the low-skill worker.
To see this, we can calculate the probability the high-skill worker receives a different

18
signal than the algorithm:

P r(s1, a0 |θH ) = Pr(s1 , a0 |ω1 , θH ) Pr(ω1 |θH ) + Pr(s1 , a0 |ω0, θH ) Pr(ω0 |θH )
1
= (Pr(s1 , a0 |ω1 , θH ) + Pr(s1 , a0 |ω0 , θH ))
2
1
= ((1 − α)υH + α(1 − υH )) . (5)
2

Differentiating the RHS of (5) with respect to α we have 21 −υH , which is negative. Hence,
P r(s1, a0 |θH ) is decreasing in α, which implies that the less accurate the algorithm’s signal
is, the more likely the high-skill worker receives the opposite signal of the algorithm.

Since in equilibrium the high-skill worker always overrides the algorithm whenever he
receives the opposite signal of the algorithm, this result implies that the high-skill worker
more frequently overrides the algorithm. In turn, this increases the signaling content of
overriding the algorithm in equilibrium which further incentivizes the low-skill worker to

override the algorithm.

5 Labor Market Implications

In this section I explore some of the labor market implications of reputational algorithm
aversion.
Several studies find that even when being able to incorporate an algorithm’s forecast,

humans’ forecasts often perform worse than the algorithm on its own (e.g., Agarwal et al.
(2023) and Angelova, Dobbie, and Yang (2023)). The next proposition shows that this
phenomenon can also occur in the model.

Proposition 5. When the average accuracy of the worker’s signal is less than that of the

algorithm’s signal, i.e., 12 (υL + υH ) > α, the expected accuracy of the worker’s forecast
can be lower than that of the algorithm’s.

Proof. Again, because of symmetry we can restrict focus to the case in which the worker

reports m1 and examine how likely ω = ω1 as compared to when the algorithm produces

19
signal a1 .17 For the worker to be more accurate than the algorithm, the following condition
must hold:

P r(ω1 |m1 ) ≥ P r(ω1|a1 ), =⇒ P r(ω1|m1 ) ≥ α.

We can rewrite the P r(ω1 |m1 ) as follows:

= P r(m1 |ω1 , a0 )P r(a0 |ω1 ) + P r(m1 |ω1 , a1 )P r(a1 |ω1)

1
= (αγ + υH + (1 − γ)υL ) . (6)
2

From this expression notice that if the average accuracy of the worker is higher than

that of the algorithm, i.e., 21 (υL + υH ) > α, then the worker’s forecast is always more
accurate than the algorithm. However, when 21 (υL + υH ) < α this may no longer be
the case. For example suppose that υL = 0.55, α = 0.60 and υH = 0.62. Solving the
informative equilibrium numerically, we have that γ = 0.0148. Plugging these values into

the expression for the worker’s forecast accuracy, (6), we have 0.58537 which is clearly less
than the algorithm alone (0.60). Hence, there are cases in which the algorithm performs
better than the worker’s forecast even though the worker has access to the algorithm’s
signal. 

This result suggests that despite the potential of humans improving the forecast of
the algorithm, their reputational concerns can make it such that their forecast is worse
than the algorithm’s on its own. Hence, to the extent wages are determined by workers’
ability to improve the algorithm’s forecast, reputational algorithm aversion should reduce
workers’ wages relative to the wages they would receive if they used information efficiently.

Moreover, if there are costs of training or maintaining workers, this phenomenon could
cause more firms to avoid hiring them at all.
Next I explore how algorithm aversion affects the relationship between the worker’s
wage and the algorithm’s accuracy. I do not explicitly model the wage setting process,

but instead simply assume wages are increasing in the difference in the expected accuracy
17
Because of symmetry m1 and a1 are always realized half the time.

20
of the worker’s forecast and that of the algorithm.
First, it will be useful to subtract the algorithm’s accuracy, α, from the human’s
forecast accuracy given in (6):

1
(αγ + υH + (1 − γ)υL ) − α. (7)
2

Differentiating (7) with respect to α we have:

 
1 dγ
γ + (α − υL ) −2 .
2 dα

Notice that increasing the accuracy of the algorithm has two effects on the worker’s
marginal contribution to the forecast’s accuracy. The obvious direct effect is that a more
accurate algorithm makes it such that the human becomes less valuable for forecasting
dγ
the outcome. However, from Proposition 4, dα
is positive, which creates a counteracting
attenuating effect due to the worker endogenously following the algorithm more frequently

when the algorithm becomes more accurate. In other words, the human’s added value
increases due to a reduction in algorithm aversion as the algorithm becomes more accu-
rate. Hence, while overall wages should be overall lower, algorithm aversion also flattens
the relationship between wages and algorithm accuracy. Hence, relative to a benchmark

in which workers use information efficiently, this result suggests that algorithms’ effect
on workers’ wages should be more concentrated in the extensive margin of algorithm
adoption rather than the intensive margin of improvements in algorithm accuracy.18
It is worth also briefly discussing how algorithm aversion affects the adoption of algo-

rithms. Consider a situation in which the manager has a worker in place and can pay a
fixed-cost to adopt the algorithm. Here it will be useful to consider the difference between
in accuracy between the worker’s forecast with the algorithm and that without it:

1 1 1
(αγ + υH + (1 − γ)υL ) − (υL + υH ) = (α − υL ) γ (8)
2 2 2
18
Of course, this assumes algorithms are sufficiently accurate to be useful at all to begin with.

21
Intuitively, (8) says that the value of the algorithm is equal to the difference in signal
accuracy of the low-skill worker and the algorithm, times the probability that the worker
is low-skill and follows the algorithm. Naturally, the less often the worker follows the
algorithm, i.e., the lower γ, the less valuable the algorithm is to the firm, whereas if

γ = 1 then we are in the first-best use of information. Hence, to the extent that workers
cannot simply be replaced by algorithms the reputational concerns of the worker can also
deter the firm from adopting the algorithm to begin with.

6 Model Discussion and Applications

In this section I discuss some of the key assumptions in the model as well as its applica-

tions.
First, one may wonder whether the problem can be resolved if the manager simply asks
the worker to report his own forecast after which the manager makes his own forecast.
However, this ultimately collapses to the original problem. The low-skill worker would
still have incentives to at least sometimes report the his own signal rather than the

algorithm’s. Relatedly, in some contexts it may not be possible for the worker to convey
his signal to the manager, but must instead make a forecast or take an action himself
that indirectly conveys that information.19
One potential solution is to make the algorithm’s signal only observable to the worker.

In this case, the worker’s forecast cannot be compared to the algorithm’s and hence the
worker’s forecast cannot convey information to the manager on its own. There are two
practical problems with this. First, the algorithm is likely developed by other people in
the firm (or outside the firm) with expertise in artificial intelligence.20 For example, data

scientists within the firm can use firm-level inputs to forecast the same outcomes that
other employees are tasked with forecasting using their own skill or intuition. The data
scientists maintaining the algorithm will inevitably want to see the information produced
19
This could occur if the information cannot be substantiated or if the worker is making a decision,
such as an investment, that must be implemented by himself, not the manager.
20
This is the main motivation for making the algorithm’s output observable to both the worker and the
manager - the fact that those who develop algorithms are likely to be different than those who implement
decisions using them.

22
by it in order to evaluate and continue developing it. Moreover, even if the firm could
make the algorithm’s signal only observable to the worker, it may not be optimal for the
firm because it prevents them from learning about the skill of that worker.
The manager could also consider hiding or delaying the availability of the algorithm’s

signal to the worker until after the worker reports his forecast. However, here the worker’s
problem would collapse to the benchmark case without the algorithm, which can easily be
shown has a lower forecast accuracy than the informative equilibrium with the algorithm.
Moreover, even if the manager ultimately takes responsibility for the forecast or action,
the manager cannot distinguish the skill of the worker given that both low and high-skill

workers will always report their information. Hence, the manager either will always follow
the worker or the algorithm, depending on which has a higher average accuracy.
In the informative equilibrium the low-skill worker plays a mixed strategy in which
he is indifferent between reporting the algorithm’s signal or his own. While this may

not seem realistic at first, mixed strategies can also be interpreted as pure strategies
with random disturbances (e.g., Harsanyi (1973)). Equivalently, I could assume that the
low-skill worker’s signal precision is a continuous random variable with mean υL in which
the worker reports his own signal whenever the signal precision is above some threshold.

This simple alteration would lead to essentially the same equilibrium in pure strategies.
It is important to emphasize that the signaling mechanism in this model is distinct
from that in costly signaling settings (e.g., Spence (1978)). In costly signaling, the signal
is exogenously costly (e.g., going to college), while in this model there is no intrinsic
cost to overriding the algorithm.21 Rather, the forecast conveys information about the

worker’s skill purely because the high-skill worker has more confidence to override the
algorithm given that he receives a more precise signal than the low-skill worker.
One may also wonder whether the main mechanism of the model is an artifact of the
binary structure of signals and outcomes. If outcomes are continuous then the problem is

essentially the same because the worker still has only two potential pieces of information
21
This is why equilibrium refinements have less bite in cheap-talk models as compared to signaling
models (e.g., Chen, Kartik, and Sobel (2008)) because it is costless for senders to randomize over actions
on the equilibrium path.

23
he can report. If the signal is continuous the mechanism would still be present; however,
rather than the override decision being a signal, the distance of the worker’s forecast
from the algorithm would be a signal of the worker’s quality.22 Intuitively, the high-skill
worker would have more confidence to report a forecast farther away from the algorithm

given that his signal is more informative than the low-skill worker.
As is typical in career concerns models (e.g., Holmström (1999)), I assume there
are no long-term contracts. However, it can easily be shown that if the firm can write
contracts, then the first-best use of information can be achieved. Specifically, the firm
could provide a long-term wage that is invariant of the worker’s performance. In this case,

the worker would have no reputational concerns and hence, would be willing to report in
accordance with the first-best use of information.23 This is unlikely in practice for several
reasons. First, long-term contracts may not be enforceable.24 Second, although I do not
explicitly model the hiring process, offering these types of long-term contracts could lead

to a lemons problem in which only low-skill workers join the firm. Finally, and perhaps
largely due to these aforementioned issues, these types of long-term contracts are not
often seen in practice.
In terms of applications, it is important that the decisions/forecasts the worker makes

are consequential enough and not too frequent. Otherwise, the firm can learn quickly the
worker’s skill and the inefficiency will be small. For example, a natural application is an
employee in a firm’s treasury department projecting the future cash flows of the firm or
making an investment decision.
It is also important to emphasize that the model applies to situations in which fol-

lowing an algorithm has a reputational cost for a human. In practice, there are some
instances in which a human is presented with a forecast either produced by a human
or an algorithm and the human is reluctant to follow the algorithm even though there
is no obvious reputational cost of doing so (e.g., see Greig et al. (2023) in the case of
22
This logic applies whether or not the outcome is continuous or not.
23
This case is knife-edge because the worker is indifferent between the two messages; however, while
I do not directly model firm-value, the firm could also provide the worker with some equity to further
incentivize them.
24
In particular due to at-will employment the high-skill worker can potentially break the contract after
one period for a higher offer outside the firm.

24
robo-advising). The model best applies to situations in which an expert has to make
some type of forecast or decision based on his own information and an algorithm’s and
he is ultimately evaluated by others. For example, the model applies to any situation
in which an employee of a firm must incorporate information from an algorithm into

his decisions (e.g., cash flow forecasting, investment decisions and financial policies) and
those decisions are ultimately evaluated by his manager or the firm as a whole.
Finally, it is worth mentioning that while I assume the public signal is generated
by an algorithm, the model would also apply to any public information observable to
both the manager and the worker. Nonetheless, algorithms are a particularly relevant

and important application given that they produce information that is 1) in many cases
becoming more accurate than humans and 2) observable within and across firms and
institutions.25

7 Conclusion

Rapid advances in artificial intelligence have raised many questions about the future role

of human workers in society. While for certain tasks, algorithms may displace humans
entirely, for many others tasks, humans will have to work alongside them. This paper
shows how humans’ concerns for their own reputations can impede human-AI collabora-
tion. Completely rational humans may engage in “algorithm aversion” and override the

algorithm even if it is inefficient to do so. The intuition for this idea is straightforward:
humans have to sometimes override algorithms in order to prove they are not replaceable.
Practically, this can lead to one of two problems: either firms become slower to adopt
algorithms, or firms forego using humans for a wider variety of tasks than would be

necessary in the absence of humans’ reputational concerns.

The solution to reputational algorithm aversion is not obvious. As mentioned earlier,
reputational concerns can be alleviated if the decision-makers are also those put in charge
of developing the algorithm (e.g., Dietvorst, Simmons, and Massey (2018)). However, in
25
As discussed above, very often the employees who develop algorithms are different than those asked
to use them, e.g., the treasurer of a firm. Hence, it is unlikely that the user of the algorithm is the only
one who can observe its output.

25
many cases, these decision-makers do not have expertise in AI systems. For example,
most treasurers of firms likely have little understanding of how a deep learning algorithm
works. Similarly, only allowing the decision-maker to see the algorithm’s signal would
also mitigate reputational concerns. However, this would obviously be difficult if the

algorithm is developed by other employees of the firm. Moreover, after promising to keep
the algorithm’s forecast private, firms may be tempted ex-post to observe the algorithm’s
signal in an attempt to learn about workers’ skills.
Taken together, this paper identifies an entirely rational, yet realistic force that can
cause humans to be averse to incorporating information generated by algorithms into their

own decisions. Moreover, as AI systems become more and more integrated into firms, I
argue that analyzing human-AI interactions through the lens of strategic communication
games will be extremely fruitful method to understand these interactions.

26
References

Acemoglu, Daron, David Autor, Jonathon Hazell, and Pascual Restrepo, 2022, Artificial

intelligence and jobs: evidence from online vacancies, Journal of Labor Economics 40,
S293–S340.

Acemoglu, Daron, David Autor, and Simon Johnson, 2023, Can we have pro-worker ai?

Acemoglu, Daron and Pascual Restrepo, Artificial intelligence, automation, and work,

The economics of artificial intelligence: An agenda, 197–236 (University of Chicago

Press 2018).

Acemoglu, Daron and Pascual Restrepo, 2020, The wrong kind of ai? artificial intelligence
and the future of labour demand, Cambridge Journal of Regions, Economy and Society

13, 25–35.

Agarwal, Nikhil, Alex Moehring, Pranav Rajpurkar, and Tobias Salz, 2023, Combining
human expertise with artificial intelligence: Experimental evidence from radiology,
Working paper, National Bureau of Economic Research.

Agrawal, Ajay, Joshua S Gans, and Avi Goldfarb, 2019, Artificial intelligence: the am-
biguous labor market impact of automating prediction, Journal of Economic Perspec-
tives 33, 31–50.

Angelova, Victoria, Will S Dobbie, and Crystal Yang, 2023, Algorithmic recommenda-

tions and human discretion, Working paper, National Bureau of Economic Research.

Babina, Tania, Anastassia Fedyk, Alex He, and James Hodson, 2024, Artificial intel-
ligence, firm growth, and product innovation, Journal of Financial Economics 151,
103745.

Burton, Jason W, Mari-Klara Stein, and Tina Blegind Jensen, 2020, A systematic review
of algorithm aversion in augmented decision making, Journal of behavioral decision
making 33, 220–239.

27
Cao, Sean, Wei Jiang, Baozhong Yang, and Alan L Zhang, 2023, How to talk when a
machine is listening: Corporate disclosure in the age of ai, The Review of Financial
Studies 36, 3603–3642.

Chen, Ying, Navin Kartik, and Joel Sobel, 2008, Selecting cheap-talk equilibria, Econo-

metrica 76, 117–136.

Crawford, Vincent P and Joel Sobel, 1982, Strategic information transmission, Econo-
metrica: Journal of the Econometric Society 1431–1451.

Dietvorst, Berkeley J, Joseph P Simmons, and Cade Massey, 2015, Algorithm aversion:

people erroneously avoid algorithms after seeing them err., Journal of Experimental
Psychology: General 144, 114.

Dietvorst, Berkeley J, Joseph P Simmons, and Cade Massey, 2018, Overcoming algorithm
aversion: People will use imperfect algorithms if they can (even slightly) modify them,
Management science 64, 1155–1170.

Eisfeldt, Andrea L, Gregor Schubert, and Miao Ben Zhang, 2023, Generative ai and firm
values, Working paper, National Bureau of Economic Research.

Greig, Fiona, Tarun Ramadorai, Alberto G Rossi, Stephen Utkus, and Ansgar Walther,
2023, Algorithm aversion: Theory and evidence from robo-advice .

Guembel, Alexander and Silvia Rossetto, 2009, Reputational cheap talk with misunder-
standing, Games and Economic Behavior 67, 736–744.

Harsanyi, John C, 1973, Games with randomly disturbed payoffs: A new rationale for
mixed-strategy equilibrium points, International journal of game theory 2, 1–23.

Holmström, Bengt, 1999, Managerial incentive problems: A dynamic perspective, The

Review of Economic Studies 66, 169–182.

Jha, Manish, Jialin Qian, Michael Weber, and Baozhong Yang, 2024, Chatgpt and cor-
porate policies, Working paper, National Bureau of Economic Research.

28
Jussupow, Ekaterina, Izak Benbasat, and Armin Heinzl, 2020, Why are we averse towards
algorithms? a comprehensive literature review on algorithm aversion .

Lai, Vivian, Chacha Chen, Q Vera Liao, Alison Smith-Renner, and Chenhao Tan, 2021,
Towards a science of human-ai decision making: a survey of empirical studies, arXiv

preprint arXiv:2112.11471 .

Langlotz, Curtis P, 2019, Will artificial intelligence replace radiologists?

Leyer, Michael and Sabrina Schneider, 2019, Me, you or ai? how do we feel about
delegation .

Liu, Miao, 2022, Assessing human information processing in lending decisions: A machine
learning approach, Journal of Accounting Research 60, 607–651.

Longoni, Chiara, Andrea Bonezzi, and Carey K Morewedge, 2019, Resistance to medical
artificial intelligence, Journal of Consumer Research 46, 629–650.

Ludwig, Jens and Sendhil Mullainathan, 2021, Fragile algorithms and fallible decision-
makers: lessons from the justice system, Journal of Economic Perspectives 35, 71–96.

Mahmud, Hasan, AKM Najmul Islam, Syed Ishtiaque Ahmed, and Kari Smolander, 2022,
What influences algorithmic decision-making? a systematic literature review on algo-
rithm aversion, Technological Forecasting and Social Change 175, 121390.

Mullainathan, Sendhil and Ziad Obermeyer, A machine learning approach to low-value

health care: wasted tests, missed heart attacks and mis-predictions (National Bureau
of Economic Research 2019).

Niszczota, Pawel and Dániel Kaszás, 2020, Robo-investment aversion, Plos one 15,

e0239277.

Önkal, Dilek, Paul Goodwin, Mary Thomson, Sinan Gönül, and Andrew Pollock, 2009,
The relative influence of advice from human experts and statistical methods on forecast
adjustments, Journal of Behavioral Decision Making 22, 390–409.

29
Ottaviani, Marco and Peter Norman Sørensen, 2006a, Professional advice, Journal of
Economic Theory 126, 120–142.

Ottaviani, Marco and Peter Norman Sørensen, 2006b, Reputational cheap talk, The Rand
Journal of Economics 37, 155–175.

Ottaviani, Marco and Peter Norman Sørensen, 2006c, The strategy of professional fore-
casting, Journal of Financial Economics 81, 441–466.

Promberger, Marianne and Jonathan Baron, 2006, Do patients trust computers?, Journal
of Behavioral Decision Making 19, 455–468.

Scharfstein, David S and Jeremy C Stein, 1990, Herd behavior and investment, The
American Economic Review 465–479.

Shaffer, Victoria A, C Adam Probst, Edgar C Merkle, Hal R Arkes, and Mitchell A
Medow, 2013, Why do patients derogate physicians who use a computer-based diag-

nostic support system?, Medical Decision Making 33, 108–118.

Spence, Michael, Job market signaling, Uncertainty in economics, 281–306 (Elsevier

1978).

Sunstein, Cass R, 2023, The use of algorithms in society, The Review of Austrian Eco-
nomics 1–22.

Yeomans, Michael, Anuj Shah, Sendhil Mullainathan, and Jon Kleinberg, 2019, Making
sense of recommendations, Journal of Behavioral Decision Making 32, 403–414.

30
A Additional Proofs

Proof of Lemma 1. First, it will be useful to prove the following claim that the high-

skill worker never mixes in any informative equilibrium.

Claim 1. The high-skill worker plays a pure strategy in any informative equilibrium.

Proof. Because of symmetry, it is without loss of generality to consider the case in which
the algorithm’s signal is a1 . First consider the case when s = s1 . Suppose to the contrary
that the high-skill worker reports m1 with some probability p ∈ (0, 1). For this to be an
equilibrium, he must be indifferent between sending m0 and m1 :

 
P r(ω1|s1 , a1 , θH ) θ̂(m1 , a1 , ω1 ) − θ̂(m0 , a1 , ω1 ) (A1)
 
= P r(ω0|s1 , a1 , θH ) θ̂(m0 , a1 , ω0 ) − θ̂(m1 , a1 , ω0 ) .

If the high-skill worker mixes when s = s1 , he cannot also mix when s = s0 . To see
this, we solve for θ̂(m1 , a1 , ω1) in (A1) and plug the solution into the high-skill worker’s
indifference condition when s = s0 :

(1 − α) (2υH − 1) (θ̂(m1 , a1 , ω0 ) − θ̂(m0 , a1 , ω0))

= 0. (A2)
υH ((2α − 1)υH − α)

However, for (A2) to hold it requires that the θ̂(m1 , a1 , ω0 ) = θ̂(m0 , a1 , ω0 ), which con-
tradicts the equilibrium being informative.
Now consider the case in which the high-skill worker mixes when we he receives signal
s1 and plays a pure strategy when he receives signal s0 . For the high-skill worker to
report m1 when he receives signal s0 , the LHS of (A2) must be positive. However, this

can only be the case when θ̂(m1 , a1 , ω0) − θ̂(m0 , a1 , ω0 ), which contradicts the equilibrium
being informative. Hence, the only other possibility is that the high-skill worker reports
m0 when he receives signal s0 . If we plug in the solution for θ̂(m1 , a1 , ω1 ) obtained from
(A1) into the low-skill worker’s difference in payoff from reporting m1 versus m0 after

31
receiving signal s0 we have:
 
(1 − α)(υH − υL ) θ̂(m1 , a1 , ω0 ) − θ̂(m0 , a1 , ω0 )
. (A3)
υH (1 − α + (2α − 1)υL )

For the equilibrium to be informative it must be that θ̂(m1 , a1 , ω0 ) < θ̂(m0 , a1 , ω0 ), which

implies (A3) is negative, which then implies the low-skill worker would also report m0
when s = s0 . The indifference condition of the high-skill worker, (A1), implies that
the low-skill worker also always reports m0 when s = s1 because P r(ω1|s1 , a1 , θH ) >
P r(ω1|s1 , a1 , θL ) and P r(ω0|s1 , a1 , θL ) > P r(ω0|s1 , a1 , θH ). Hence, under this scenario

the low-skill worker always reports m0 while the high-skill worker sometimes reports m1 .
However, based on the same logic as in Proposition 2, this can never be an equilibrium
because the worker will instantly be identified as high-skill if he reports m1 when a = a1 .
This can easily be confirmed by computing the posteriors and plugging them into either

the low or high-skill worker’s incentive compatibility constraints.

Since the signals are symmetric this also proves that the high-skill worker mixing
when he receives s0 can also never be an equilibrium. The same logic steps apply when
the algorithm’s signal is a0 instead of a1 . Hence, the high-skill worker always plays a pure
strategy in any informative equilibrium. 

Given that the high-skill worker does not mix, there are four possible cases to consider:

Case 1: The high-skill worker always reports the opposite of his signal

Case 2: The high-skill worker always reports the algorithm’s signal

Case 3: The high-skill worker always reports the opposite of the algorithm’s signal

Case 4: The high-skill worker always reports his signal

Given the symmetry of the problem, it is without loss of generality to assume a = a1

for each case.
First consider Case 1 in which the high-skill worker always reports the opposite of his

32
signal. This requires the following inequalities to hold:

X X
P r(ω|s1, a1 , θH )θ̂(m1 , a1 , ω) ≤ P r(ω|s1, a1 , θH )θ̂(m0 , a1 , ω)
ω ω
X X
P r(ω|s0, a1 , θH )θ̂(m0 , a1 , ω) ≤ P r(ω|s0, a1 , θH )θ̂(m1 , a1 , ω).
ω ω

The first inequality can be written as:

 
P r(ω1|s1 , a1 , θH ) θ̂(m1 , a1 , ω1 ) − θ̂(m0 , a1 , ω1) ≤
 
P r(ω0|s1 , a1 , θH ) θ̂(m0 , a1 , ω0 ) − θ̂(m1 , a1 , ω0 ) ,

and the second to:

 
P r(ω0|s0 , a1 , θH ) θ̂(m0 , a1 , ω0 ) − θ̂(m1 , a1 , ω0) ≤
 
P r(ω1|s0 , a1 , θH ) θ̂(m1 , a1 , ω1 ) − θ̂(m0 , a1 , ω1 ) .

Both sides of these inequalities must be positive in an informative equilibrium. More-

over, they imply that the low-skill worker would also report the opposite of his own signal
because P (ω0 |s1 , a1 , θL ) > P (ω0|s1 , a1 , θH ) and P (ω1|s0 , a1 , θL ) > P (ω0 |s0 , a1 , θH ). How-
ever, if the low-skill worker also reports the opposite of his signal, after calculating the
corresponding posterior beliefs of the manager, it immediately becomes clear that the

equilibrium is not informative:

1 − υH υH
θ̂(m1 , a, ω1 ) = < = θ̂(m0 , a, ω1) a ∈ {a0 , a1 }.
2 − υH − υL υH + υL

Hence, the high-skill worker cannot always report the opposite of his signal.
Cases 2 and 3 can easily be ruled out based on the following logic. In order for Case 2 or
3 to be an equilibrium, the low-skill worker must also always either report the algorithm’s
signal (Case 2) or the opposite of the algorithm’s signal (Case 3), otherwise the low-skill

worker would be identified with probability one as being low-skill. However, in either
of these cases the equilibrium would not be informative, which leads to a contradiction.

33
Hence, the only remaining possible case is Case 4 in which the high-skill worker reports
his signal.


Proof of Lemma 2. Because of the symmetry of the problem, it is again without loss
of generality to consider the case in which the algorithm’s signal is a1 . When s = s1

suppose that the low-skill worker reports m1 with some probability p ∈ (0, 1). Then the
following condition must hold:

 
P r(ω1|s1 , a1 , θL ) θ̂(m1 , a1 , ω1 ) − θ̂(m0 , a1 , ω1 ) (A4)
 
= P r(ω0 |s1 , a1 , θL ) θ̂(m0 , a1 , ω0) − θ̂(m1 , a1 , ω0 ) .

If we solve for θ̂(m1 , a1 , ω1 ) in (A4) and plug this into the difference in the low-skill

worker’s payoff from reporting m1 versus m0 when s = s0 we have:

 
(1 − α) (2υL − 1) θ̂(m1 , a1 , ω0 ) − θ̂(m0 , a1 , ω0 )
. (A5)
υL (α − (2α − 1)υL )

If the equilibrium is informative, i.e., θ̂(m1 , a1 , ω0 ) < θ̂(m0 , a1 , ω0 ), then (A5) is negative,

which implies that the low-skill worker always reports m0 when he his signal is s0 . We
can then calculate the posteriors and plug them into (A4):

 
υH 1 − υH
P r(ω1|s1 , a1 ) −
υH + pυL 2 − υH − pυL
 
υH 1 − υH
= P r(ω0 |s1 , a1 ) − .
1 + υH − p (1 − υL ) 1 − υH − p (1 − υL )

Further simplifying we have:

(1−α)(1−υH −p(1−υL ))(1−υL ) αυL (υH −pυL )

(1−υH +p(1−υL ))(1+υH −p(1−υL ))
+ (2−υH −pυL )(υH +pυL )
= 0. (A6)
1 − α + (2α − 1)υL

However, the LHS of (A6) is strictly positive which means it is not incentive compatible
for the low-skill worker to mix when s = s1 .

34
 Now consider the low-skill worker reporting m0 when s = s1 . Let p denote the
probability the low-skill worker reports m1 when s = s0 . We can then calculate the
posteriors and plug them into the low-skill worker’s incentive compatibility constraint
when s = s1 :

 
υH 1 − υH
P r(ω1|s1 , a1 ) −
υH + p (1 − υL ) 2 − υH − p (1 − υL )
 
υH 1 − υH
≤ P r(ω0|s1 , a1 ) − .
1 + υH − pυL 1 − υH + pυL

Further simplifying we have:

αυH υL α(1−υH )υL (1−α)(1−υH )(1−υL ) (1−α)υH (1−υL )

υH +p(1−υL )
− 2−υH −p(1−υL )
+ 1−υH +pυL
− 1+υH −pυL
≤ 0. (A7)
1 − α + (2α − 1)υL

However, the LHS of (A7) is strictly positive which contradicts the low-skill worker re-
porting m0 when s = s1 . To see that (A7) is positive, we can differentiate the LHS of
(A7) with respect to p and we have:

 
αυH α(1−υH ) (1−α)υH (1−α)(1−υH )
(1 − υL ) υL (υH +p(1−υL ))2
+ (2−υH −p(1−υL ))2
+ (1+υH −pυL )2
+ (1−υH +pυL )2
− <0
1 − α + (2α − 1)υL

Hence, (A7) is decreasing in p. However, when p = 1 (A7) simplifes to:

(α + υL − 1) (υH + υL − 1)
,
(1 − (υH − υL ) 2 ) (1 − α + (2α − 1)υL )

which is strictly positive. Hence, (A7) is violated. Therefore, the low-skill worker reports
m1 when s = s1 and a = a1 . Because of symmetry, the same steps can be used to show
the low-skill worker reports m0 when s = s0 and a = a0 . 

35