Journal of Experimental Psychology: General
Mistaking Minds and Machines: How Speech Affects
Dehumanization and Anthropomorphism
Juliana Schroeder and Nicholas Epley
Online First Publication, August 11, 2016. http://dx.doi.org/10.1037/xge0000214
CITATION
Schroeder, J., & Epley, N. (2016, August 11). Mistaking Minds and Machines: How Speech Affects
Dehumanization and Anthropomorphism. Journal of Experimental Psychology: General. Advance
online publication. http://dx.doi.org/10.1037/xge0000214
Mistaking Minds and Machines: How Speech Affects Dehumanization
and Anthropomorphism
Juliana Schroeder
University of California, Berkeley
Nicholas Epley
University of Chicago
Treating a human mind like a machine is an essential component of dehumanization, whereas attributing
a humanlike mind to a machine is an essential component of anthropomorphism. Here we tested how a
cue closely connected to a person’s actual mental experience—a humanlike voice—affects the likelihood
of mistaking a person for a machine, or a machine for a person. We predicted that paralinguistic cues in
speech are particularly likely to convey the presence of a humanlike mind, such that removing voice from
communication (leaving only text) would increase the likelihood of mistaking the text’s creator for a
machine. Conversely, adding voice to a computer-generated script (resulting in speech) would increase
the likelihood of mistaking the text’s creator for a human. Four experiments confirmed these hypotheses,
demonstrating that people are more likely to infer a human (vs. computer) creator when they hear a voice
expressing thoughts than when they read the same thoughts in text. Adding human visual cues to text
(i.e., seeing a person perform a script in a subtitled video clip), did not increase the likelihood of inferring
a human creator compared with only reading text, suggesting that defining features of personhood may
be conveyed more clearly in speech (Experiments 1 and 2). Removing the naturalistic paralinguistic cues
that convey humanlike capacity for thinking and feeling, such as varied pace and intonation, eliminates
the humanizing effect of speech (Experiment 4). We discuss implications for dehumanizing others
through text-based media, and for anthropomorphizing machines through speech-based media.
Keywords: communication, voice, mind perception, dehumanization, anthropomorphism
Alan Turing (1950) created a famous benchmark for determin-
ing whether a computer can “think:” when it can convince a
majority of people that they are interacting with a person instead of
a machine. Turing’s link between thinking and personhood is no
accident (Farah & Heberlein, 2007). Boethus, writing in the 6th
century, defined personhood as “an individual substance of ratio-
nal nature” (Singer, 1994). Centuries later, John Locke (1841/
1997) echoed this definition of a person as “an intelligent being
that has reason and reflection.” Immanuel Kant (1785/1993) used
this definition of personhood as the guiding light of morality,
noting that, “rational beings are called persons inasmuch as their
nature already marks them out as ends in themselves.” Recent
surveys of laypeople likewise identify the capacity for thinking as
a unique feature of humanity (Leyens et al., 2000, 2001). A person
has a mind capable of thinking but a computer does not.
As clear as this reality may be, it may not be so clear psycho-
logically. People sometimes recognize a thoughtful mind in their
cars, computers, or other mindless gadgets (Guthrie, 1995; Naas,
2010). A robot that moves at a humanlike pace seems more
thoughtful than a relatively sluggish or frantic robot (Morewedge,
Preston, & Wegner, 2007). An autonomous automobile that inter-
acts with you using a human voice while driving itself seems
“smarter,” and therefore, more trustworthy, than a noninteractive
vehicle (Waytz, Heafner, & Epley, 2014). Attributing humanlike
mental capacities of thinking and feeling to nonhuman agents is
the essence of anthropomorphism (Epley, Waytz, & Cacioppo,
2007).
Inversely, people sometimes fail to recognize a thoughtful mind
in other human beings, treating them instead like relatively mind-
less animals or objects (Haslam, 2006). Failing to attribute a
humanlike mind to another person is the essence of dehumaniza-
tion. These twin phenomena of anthropomorphism and dehuman-
ization raise a fundamental question in social life that goes far
beyond Turing’s test: how does an agent convey the fundamentally
humanlike capacity to think or feel? Answering this question will
predict when machines might be treated like people, and when
people might be treated like machines. It also predicts when
machines might be mistaken for people, and people mistaken for
machines.
Existing theory predicts that anthropomorphism and dehuman-
ization are determined by features of the agent being perceived
(e.g., morphology, motion, and observed behavior) and by features
Juliana Schroeder, Haas School of Business, University of California,
Berkeley; Nicholas Epley, Booth School of Business, University of Chicago.
We thank the Neubauer Family Faculty Fellowship for financial support
of this research, and Daniel Gilbert for suggesting the main dependent
variable used in these experiments. We also thank Jasmine Kwong, Michal
Dzitko, Annette Felton, Shreya Kalva, Alex Kristal, Paul Lou, Adam
Picker, Megan Porter, Sunni Rogers, Jenna Rozelle, Max Snyder, and
Sherry Tseng for assistance conducting these experiments. Portions of this
research were presented at the 2013 Annual Meeting of the Society for
Personality and Social Psychology.
Correspondence concerning this article should be addressed to Juli-
ana Schroeder, Haas School of Business, University of California,
Berkeley, 2220 Piedmont Avenue, Berkeley, CA 94720. E-mail:
jschroeder@haas.berkeley.edu
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Journal of Experimental Psychology: General © 2016 American Psychological Association
2016, Vol. 145, No. 9, 000 0096-3445/16/$12.00 http://dx.doi.org/10.1037/xge0000214
1
of the perceiving agent (e.g., group affiliation, motivation, and
social connection; Epley, Waytz, & Cacioppo, 2007). Here we test
what we believe is a particularly important feature of the agent
being perceived: a humanlike voice. Beyond the semantic content
in speech, a humanlike voice also conveys paralinguistic informa-
tion (e.g., volume, tone, and rate) that provides additional insight
into one’s thoughts and feelings. Indeed, voice evolved in large
part as a tool to communicate an agent’s mind to others through
speech (Pinker & Bloom, 1990), and people can more accurately
estimate others’ mental states when they hear someone speak than
when they read the same words in text (Hall & Schmid Mast, 2007;
Kruger, Epley, Parker, & Ng, 2005). Therefore, we predicted that
communicating with a humanlike voice would make an agent seem
more like a person (vs. machine) than communicating the same
content through other communication media (e.g., reading text,
observing body language, or speaking with a voice that lacks
critical paralinguistic cues). Adding a humanlike voice to a ma-
chine might make it seem more like a person (i.e., anthropomor-
phism). Likewise, removing voice from an actual person by com-
municating through text might make a person seem more like a
machine (i.e., dehumanization).
Several existing experiments suggest that adding a humanlike
voice to computerized agents increases anthropomorphism (Nass
& Brave, 2005; Takayama & Nass, 2008; Waytz, Heafner, &
Epley, 2014). Our experiments go beyond these results by provid-
ing a more precise understanding of the interpersonal conse-
quences of hearing a humanlike voice compared with observing
other inferential cues (e.g., visual cues such as seeing a human), by
identifying which cues in voice convey personhood, and by testing
the inverse possibility that removing voice from human interac-
tions might lead to dehumanized perceptions of a speaker. This
latter hypothesis is especially important as technology makes
text-based interactions increasingly common in everyday life.
We test how a humanlike speaking voice affects dehumaniza-
tion and anthropomorphism in four experiments. For each exper-
iment, we report how we determined our sample size, all data
exclusions, all manipulations, and all measures. Data for all ex-
periments can be retrieved at http://faculty.haas.berkeley.edu/
jschroeder/data.html. Experiments 1 and 3 remove voice and mea-
sure dehumanization (mistaking a human for a computer) whereas
Experiment 2 adds voice and measures anthropomorphism (mis-
taking a computer for a human). All experiments test the effect of
hearing speech versus reading the same words in text. To increase
generalizability and remove confounds, we generated text using
different methods in each experiment: transcriptions of human
speech (Exp. 1), computer-generated text (Experiment 2), or writ-
ten essays (Experiments 3 and 4). We added a third channel of
human visual cues in Experiments 1–3 to determine whether
speech is uniquely humanizing, and to test an alternative explana-
tion that any humanlike cue will reduce dehumanization and
increase anthropomorphism. We predicted a speaker’s voice would
be uniquely humanizing because it contains paralinguistic cues
that reveal active mental experiences of thinking and feeling, and
that these cues uniquely reveal the presence of a humanlike mind
(Schroeder & Epley, 2015).
Our final experiment tests why speech is humanizing. We sug-
gest that paralinguistic cues in a person’s voice can convey the
presence of a lively, thoughtful, and active mind, in a way that is
analogous to how visual cues convey the presence of biological
life. A person can tell whether another agent is alive or dead
because of variance in motion. A living person’s body moves in
naturalistic ways. A dead person’s body is still, with no variance in
motion. Likewise, a person’s voice also contains naturalistic vari-
ance through paralinguistic cues that may analogously convey the
presence of a lively and active mind. A speaker’s pitch rises and
falls over time, yielding variance in tone (intonation). A speaker’s
pace quickens and slows, producing variance in speech rate. The
presence of a lively and active mind—a humanlike mind— could
be reflected through these cues. A rising pitch, for instance, may
reflect confidence in judgment whereas a falling pitch may reflect
more careful deliberation. Consistent with this possibility, speak-
ers communicate mental states such as the valence of emotional
experience or intentions even when speech lacks any meaningful
semantic content (McAleer, Todorov, & Belin, 2014; Scherer,
Banse, & Wallbott, 2001; Weisbuch, Pauker, & Ambady, 2009).
This predicts that a person with a speaking voice lacking natural-
istic variance in paralinguistic cues would be evaluated similar to
a person being evaluated over text or another communication
media devoid of paralinguistic cues (such as silent video). A voice
lacking variance in paralinguistic cues might make another per-
son’s mind seem relatively dead or dull, more like a mindless
machine than like a mindful human being.
Experiment 1: Dehumanization
We examined our hypotheses by using the inverse of Alan
Turing’s (1950) famous test: are observers more likely to mistak-
enly believe genuine human speech was created by a computer
when they hear the speech than when they read the very same
words? To rule out the artifact that any human cue might lead
observers to think the script was created by a human, we also
added visual cues: seeing a human. If a person’s voice uniquely
conveys humanlike mental capacities, as we predict, then we
should find an effect of voice, and no unique effect of visual cues,
on judgments of a speech’s creator. To test our prediction, we
created a novel paradigm in which individuals evaluate whether
the creator of the content that they hear or read was a computer or
a human. Note that participants are not evaluating whether a voice
is that of a computer or a human, but rather whether the creator of
the content was a human or a machine. This is essential because
computerized voices could obviously sound different than a real
human voice. Our hypotheses are not testing how people evaluate
voices in communication, but rather about the inferences people
make about the agents who are actually communicating.”
Method
Participants. We decided to collect at least 20 speakers to
obtain a more ecologically valid range of stimuli than is common
in most psychology experiments (Fiedler, 2011; Kenny, 1985;
Wells & Windschitl, 1999). This stimulus sampling is an important
feature of all of our experiments because it enables us to assess
how people evaluate a range of naturalistic speakers that might be
encountered in everyday life (Brunswik, 1947). It also eliminates
concerns that any result might be produced by some idiosyncratic
feature of a single person’s voice, or by some artifact introduced
by creating artificial speech content. Naturalistic stimulus sam-
pling enables stronger inferences about the strength of a phenom-
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
2
SCHROEDER AND EPLEY
enon in the midst of the perceptually rich and chaotic environment
of everyday life.
We initially recruited 33 people (M
age
⫽ 20.0, SD
age
⫽ 2.3,
58% female) from a Chicago research laboratory to serve as
speakers in exchange for $2.00. Each speaker created two videos:
one talking about a positive emotional experience and the other
talking about a negative emotional experience (in counterbalanced
order). We had no a priori prediction about the effect of emotional
valence; we manipulated valence only as a robustness check for
the magnitude of our predicted effect of speech (vs. text). To
obtain our target of 20 speakers, two independent raters subse-
quently coded the videos based on the extent to which the speaker
followed our instructions. Specifically, they evaluated “the extent
to which participants talked about a very emotional experience and
described all of their emotions in detail” on a scale of 0 (not at all),
1(somewhat), or 2 (very much), r ⫽ .73. We included the 20
speakers who followed our instructions the best on this measure
across their two speeches, giving us a final sample of 40 total
speeches (20 positive experiences and 20 negative experiences).
The 20 speakers included in the study did not differ significantly
from the 13 speakers not included based on their gender,
2
(1,
33) ⫽ 0.12, p ⬎ .10, or their age, t(31) ⫽ 1.23, p ⬎ .10.
We decided to collect at least 640 observers so that at least four
would watch each type of stimulus for each of the 40 videos in the
four experimental conditions (160 conditions total). In total, 652
observers (M
age
⫽ 32.0, SD
age
⫽ 10.6, 42% female, 7 missing
gender) from Amazon.com’s Mechanical Turk participated. These
observers completed the experiment in exchange for $0.30. We
removed five observers who did not complete our primary depen-
dent variable from analysis.
Speaker procedure. Participants described both a positive
and negative emotional experience in their life (in counterbalanced
order) on videotape. We selected this speech topic because it is
generally humanizing, to the extent that most people believe hu-
mans have more emotional experience than do machines. For the
positive emotional experience, the experimenter asked participants
to:
Think back on an important positive emotional experience that you
had. Talk about the entire experience and all of the positive emotions
that you felt during the experience. Describe your emotions from
beginning to end. This should be an experience that led you to feel
very emotional, such as feeling deep happiness. Be sure while telling
your story to explain all of your feelings and emotions throughout the
entire experience.
Instructions for the negative emotional experience were modi-
fied to refer to experiences of “deep sadness” instead of “deep
happiness.” The experimenter asked the speakers to repeat the
instructions to ensure that they understood. Speakers then sat in a
chair facing a video camera. To make the speeches as natural as
possible, the experimenter stood behind the camera and told speak-
ers to look at the camera and pretend like they were talking directly
to the experimenter. The speaker talked until he or she was
finished telling the story. Speech lengths varied from 1 to 3 min.
The experimenter then stopped the video camera and read the next
set of instructions for the negative emotional experience video
(order counterbalanced). Again, speakers repeated the instructions,
sat facing the video camera, and gave their speech. Finally, the
experimenter debriefed speakers.
One research assistant transcribed the speeches after the verba-
tim transcript procedure used in United States courtroom deposi-
tions. A second assistant checked the transcriptions for accuracy.
Observer procedure. We randomly assigned observers to one
of four experimental conditions: listening to a speaker (audio
condition), watching and listening to a speaker (audiovisual con-
dition), reading the speech (text condition), or reading and watch-
ing a speaker (subtitled video condition, with no sound included).
To make the presentation of stimuli as consistent as possible across
conditions, we presented all stimuli to observers as videos. Ob-
servers in the audio and text conditions, therefore, saw a black
video screen and either heard the words or read them on the screen,
respectively. This paradigm allows us to keep constant the amount
of time each observer spent on each stimulus.
Observers in the text condition read the following information
before observing the stimuli: “As you may know, computer tech-
nology is now attempting to mimic real human speech. Some
computer programs are good enough that they can convince some
observers that they are real people, whereas others are not as good.
For the next few minutes, you will read the text of a script. Your
job is to figure out whether the content of this script was originally
created by a computer trying to create a script that sounded like a
real human being or whether it was created by an actual human.”
Observers in the audio version read the same instructions, except
the third sentence read, “For the next few minutes, you will listen
to an actor (or actress) reciting a script.” Observers in the audio-
visual and subtitled video condition likewise read the same in-
structions, except the third sentence read, “For the next few min-
utes, you will watch an actor (or actress) reciting a memorized
script.”
To make sure that observers understood their task, those in the
audio condition received further clarification: “To be absolutely
clear, your job is not to determine whether the voice is of a real
person or not. You will hear a real human actor reading a script.
Your job is to determine whether the script itself was originally
written by a computer or an actual human.” Observers in the
audiovisual and subtitled video condition, in contrast, received this
clarification: “To be absolutely clear, your job is not to determine
whether the actor or actress is a real person or not. You will watch
a real human actor reciting a memorized script. Your job is to
determine whether the script itself was originally written by a
computer or an actual human.”
After reporting whether the script was originally created by a
human or computer, participants also reported how confident they
were that their answer was correct on an 11-point response scale
(0 ⫽ not at all confident,5⫽ moderately confident,10⫽ abso-
lutely confident), and then explained “why they made the choice
they did” in a free-response box. We did not analyze the free
responses. Participants reported their confidence in this same way
in each of the following experiments as well (Experiments 2– 4),
but confidence did not vary reliably by condition in any experi-
ment reported in the manuscript and we, therefore, do not discuss
it further.
Results. Whether the speech was about a negative or positive
emotional experience did not affect the human versus computer
judgment, F(1, 639) ⫽ 0.01, nor did it interact with communica-
tion medium on this judgment, F(3, 639) ⫽ 0.82. Therefore, we
collapsed across this factor in the following analyses.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
3
MINDS AND MACHINES
Observers’ judgments of the script’s creator varied by experi-
mental condition, F(3, 643) ⫽ 10.76, p ⬍ .01,
2
⫽ .05. Because
we used a nested experimental design (multiple observers for each
speaker), we analyzed the effect of each condition (fixed factors)
in a hierarchical regression controlling for the effect of speaker
(random factor). As shown in Figure 1, removing voice was
dehumanizing: observers who read the speeches were less likely to
believe it was created by a human (text condition; M ⫽ 53.6%,
SD ⫽ 50.0%) than observers who listened to them (M ⫽ 80.8%,
SD ⫽ 39.5%), t(627) ⫽ 5.29, p ⬍ .01, d ⫽ 0.42. Adding individ-
uating visual cues to the voice in the audiovisual condition did not
increase the percentage who guessed the script was created by a
human (audiovisual condition; M ⫽ 71.6%, SD ⫽ 45.2%) com-
pared with audio alone, t(628) ⫽⫺1.79, p ⫽ .07, d ⫽ 0.14. In
contrast, stripping away the person’s voice while retaining text in
the subtitled video condition reduced the percentage who guessed
the script was created by a human (subtitled video condition; M ⫽
60.7%, SD ⫽ 49.0%) compared with the audio condition alone,
t(628) ⫽ 3.70, p ⬍ .01, d ⫽ 0.29. The text and subtitled video
conditions did not differ significantly from each other, t(628) ⫽
1.43, p ⫽ .15, d ⫽ 0.11. An observer was most likely to believe a
script was created by a human when they heard the speaker’s
voice.
Discussion
Participants who heard a speech were more likely to believe its
content was created by a human (vs. computer) than participants
who read the very same speech. The cues in voice seem uniquely
humanizing, as visual cues— being able to see speakers in addition
to hearing their voices— did not increase the percentage of eval-
uators who believed the script was created by a human. Being able
to see a speaker without hearing his or her voice, however, sig-
nificantly decreased the percentage who believed the script was
created by a human. The presence of a mindful human creator was
most clearly conveyed through a person’s voice rather than by the
overall amount of individuating (humanlike) cues available.
Experiment 2: Anthropomorphism
The results of Experiment 1 suggest that removing a voice from
human-generated speech can lead people to believe its content was
created by a mindless machine. Experiment 2 tests the inverse: can
adding a human voice to computer-generated speech lead people to
believe its content was created by a mindful human being? This
experiment is therefore closer to an actual Turing test, examining
when human observers might mistake a mindless machine for a
mindful person.
Method
Participants. We predetermined a sample size of at least 10
speakers and 240 observers. These sample sizes were smaller than
Experiment 1 because speakers all read the same computer-
generated text and so we expected less variability in evaluations
across speakers. Ten people (50% female) from a Chicago research
laboratory served as speakers in exchange for $3.00. Subsequently,
243 people (M
age
⫽ 30.9, SD
age
⫽ 10.0, 38% female) from
Amazon Mechanical Turk served as observers in exchange for
$0.30.
Speaker procedure. We created the essay using a “Postmod-
ernism Generator” (http://www.elsewhere.org/pomo/) that uses a
computer system to generate random text from recursive grammars
(Bulhak, 1996). The full text was:
“Society is elitist,” says Derrida. It could be said that Porter suggests that
we have to choose between material nihilism and neocultural theory.
“Truth is intrinsically dead,” says Debord; however, according to
Reicher, it is not so much truth that is intrinsically dead, but rather the
paradigm, and subsequent stasis, of truth. Sartre promotes the use of
dialectic subconceptualist theory to deconstruct the status quo. How-
ever, Baudrillard uses the term ‘cultural feminism’ to denote the
common ground between sexual identity and class. Marx suggests the
use of material nihilism to analyze sexual identity. But Bataille uses
the term ‘cultural feminism’ to denote a self-supporting totality. The
example of material nihilism intrinsic to Gibson’s Count Zero
emerges again in Virtual Light, although in a more mythopoetical
sense. Therefore, Baudrillard promotes the use of subdeconstructivist
theory to challenge hierarchy. Debord’s model of semantic Marxism
states that culture is part of the futility of consciousness. It could be
said that the main theme of the works of Gibson is the failure, and
hence the dialectic, of postcultural society. Cultural feminism holds
that context comes from the masses. But Bataille suggests the use of
Lyotardist narrative to read and analyze art.
Just before giving their speech, speakers read the following
instructions that were intended to maintain the natural paralinguis-
tic cues present in actual human speech:
When you perform, please imagine that you are the person who wrote
the essay. Imbue your words with all of the thoughts, emotions, and
20%
40%
60%
80%
100%
Audio Audiovisual Text Subtitled Video
Likelihood of Choosing "Human"
(for Human-Generated Speech)
20%
40%
60%
80%
100%
Audio Audiovisual Text Subtitled Video
Likelihood of Choosing "Human"
(for Computer-Generated Text)
Figure 1. Percentage of observers in Experiment 1 (Panel 1; n ⫽ 652;
stimuli created from human-generated speech) and Experiment 2 (Panel 2;
n ⫽ 243; stimuli created from computer-generated text) who believed a
script had been created by a human (vs. computer) in the audio, audiovi-
sual, text, and subtitled video conditions. Errors bars represent the SEM.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
4
SCHROEDER AND EPLEY