How Scientists Think: On-Line

Creativity and Conceptual

Change in Science
Kevin Dunbar

T his chapter reports an investigation of "On-line Creativity." I present

a new account of the cognitive and social mechanisms underlying
complex thinking of creative scientists as they work on significant prob-
lems in contemporary science. I lay out an innovative methodology that
I have developed for investigating creative and complex thinking in a real-
world context. Using this method, I have discovered that there are a num-
ber of strategies that are used in contemporary science that increase sci-
entists' likelihood of making discoveries. The findings reported in this
chapter provide new insights into complex scientific thinking and will dis-
pel many of the myths surrounding the generation of new concepts and
scientific discoveries.

This research was made possible by a grant from the Spencer Foundation. The research was also supported
by Grant OGP0037356 from the National Sciences and Engineering Council of Canada and a leave of ab-
sence given by the Department of Psychology at McGill University. I thank Laura Ann Petitto for her ex-
tensive comments on a draft of the manuscript. I also thank Tom Ward, Keith Holyoak, and Maude A. St.
Laurent for their comments on a draft of this chapter. Finally, I thank all of the scientists who opened up
their laboratories and their minds for this research.

461
 KEVIN DUNBAR

IN VIVO COGNITION: A NEW WAY OF

INVESTIGATING COGNITION
There is an extensive background in cognitive research on thinking, rea-
soning, and problem solving processes that form the foundation for creative
cognition (see Dunbar, in press; Holyoak, 1995, for recent reviews). How-
ever, to a large extent, research on reasoning has demonstrated that partic-
ipants in psychology experiments make vast numbers of thinking and rea-
soning errors even in the most simple problems. How is creative thought
even possible if people make so many reasoning errors? One problem with
research on reasoning is that the concepts and stimuli that the research par-
ticipants are asked to use are often arbitrary and involve no background
knowledge (cf. Dunbar, 1995; Klahr & Dunbar, 1988). I have proposed that
one way of determining which reasoning errors are specific and which are
general is to investigate cognition in the cognitive laboratory and the real
world (Dunbar, 1995). Psychologists should conduct both in vitro and in
vivo research to understand thinking. In vitro research is the standard psy-
chological experiment where individuals are brought into the laboratory and
controlled experiments are conducted. As can be seen from the research re-
ported in this volume, this approach yields many insights into the psycho-
logical mechanisms underlying complex thinking. The use of an in vivo
methodology in which on-line thinking and reasoning are investigated in a
real-world context yields fundamental insights into the basic cognitive
mechanisms underlying complex cognition and creativity. The results of in
vivo cognitive research can then be used as a basis for further in vitro work
in which controlled experiments are conducted. In this chapter, I outline
some of the results of my ongoing in vivo research on creative scientific
thinking. I relate this research to more common in vitro research to show
that the in vivo method generates new basic models of cognitive processes
and opens up avenues for new in vitro research.

On-Line Scientific Thinking

Scientific thinking is an ideal domain in which to develop theories of cre-
ative cognition and complex thinking (see Klahr, 1994, for a recent review

462
 HOW SCIENTISTS THINK

of this literature). First, scientists are constantly adding to knowledge and,

less frequently, developing new concepts and theories. Second, scientists
already have a rich background of knowledge in their domain that they
use as a foundation for their thought. Third, much creativity occurs in
groups rather than individuals alone. Contemporary science, which in-
cludes psychology, entails an experimental context that involves a group.
No longer is the lone scientist under the lightbulb the norm for science.
Rather, groups containing members with different levels of experience and
different scientific backgrounds form the basis of contemporary science.
Little is known about the way in which groups reason. Thus, scientific
groups are a very important source of creative thinking and reasoning. In
sum, by investigating science as it is practiced it is possible to address key
questions about the nature of thinking and creativity, uncover funda-
mental processes that underlie complex thinking, and suggest strategies
for enhancing creative thought.

Method
The research program that I have developed centers on understanding the
cognitive and social mechanisms involved in current day science. I have
selected molecular biology as a scientific domain to investigate because
this domain is of central importance to contemporary science. Many of
the brightest and most creative minds in science are attracted to this field,
and molecular biology receives a very significant proportion of funding
in science and medicine. Furthermore, the field of molecular biology is
undergoing an immense period of scientific discovery and breakthrough,
making it an ideal domain in which to investigate creative thinking.
Having identified molecular biology as a scientific domain I then
sought to identify leading laboratories in the United States that I could in-
vestigate. My goal was to investigate the thinking and reasoning strategies
that leading scientists use while conducting their research. After consult-
ing with a number of scientists, including one Nobel Prize winner, and
extensively reviewing the literature, I identified six world-renowned sci-
entists at a major U.S. university. All scientists were internationally known
for conducting innovative research that frequently stretched the bound-

463
 KEVIN DUNBAR

aries of their field. Each scientist was concerned with discovering new bi-
ological mechanisms that give fundamental insights into biology. Having
identified the six laboratories, I then contacted the scientists and asked
them to participate in my research. All six agreed to participate in the
study. I then interviewed the scientists to determine what their current re-
search projects were, what the scientists in their labs were doing, and what
their plans for the coming year were. Following this consultation I then
selected four laboratories as being most suitable for investigation.
The goal of this research was to identify the points in time at which
innovative scientific thinking occurs, capture this thinking on audio- and
videotape, and then analyze the processes involved in the scientists'
thinking and reasoning. To this end I spent a year in the four selected
molecular biology laboratories. I spent the first four months becoming
familiar with the scientists in the laboratory, staying in the labs during
the day, attending lab meetings, interviewing the scientists in the lab,
and reading grant proposals and drafts of papers. I discovered that the
laboratory meeting is one of the central places in which new ideas and
concepts are generated. Each laboratory had a weekly meeting that all of
the members of the lab attended. The senior scientist, who manages the
lab, is present as well as the postdoctoral fellows, graduate students, and
technicians. In the lab meetings, a scientist presents his or her latest re-
search, which is conducted with the senior scientist. Members of the lab
ask questions about the research and propose new experiments, hy-
potheses, and interpretations, often forcing the presenting scientist to
reconceptualize his or her ideas. Totally new concepts are generated and
modified by members of the laboratory at some of the meetings. Often
the senior scientist plays a crucial role in the development of new ideas
and concepts. The scientists' reasoning at lab meetings is often sponta-
neous, and the on-line interactions concern some of the most creative
moments in science. The finding that lab meetings are a central source
of creative thinking and reasoning is also important because the rea-
soning that occurs at these meetings occurs through presentations and
spontaneous interactions in which the scientists develop their ideas. Be-
cause the scientists talk out loud during the meetings there is an exter-

464
 HOW SCIENTISTS THINK

nal record of thinking and reasoning. Using this method it is possible to

directly monitor thinking and reasoning rather than uncover reasoning
through post hoc interviews, questionnaires, or think-aloud protocols.
The scientists externalize much of their thinking through interactions
with other scientists in the lab. Thus, by recording laboratory meetings
it is possible to gain access to on-line thinking and reasoning without
influencing the way the scientists think.
Following my initial data collection phase, I evaluated the best method
of collecting data on scientific thinking. I found that the laboratory meet-
ings provide a much more accurate picture of the conceptual life of a lab-
oratory than interviews, lab books, or papers. In fact, I found that the sci-
entists were often unable to remember the steps in the development of a
particular concept. The laboratory meetings provided a far more veridi-
cal and complete record of the evolution of ideas than other sources of
information. Thus, I selected the laboratory meetings as the core source
of data and the interviews and papers as supplemental sources of infor-
mation. Thus, the particular method that I used to collect data revolved
around the discovery that the laboratory meetings are central to the con-
ceptual life of a laboratory.
I constructed a before-during-after design for uncovering the effects
of laboratory meetings on the scientists' theories and methods: Before a
lab meeting I interviewed the scientists to find out what their hypotheses
were, what they thought the data meant, and what they were going to do
next. I then audio or videotaped the scientists during the lab meeting. Af-
ter the meeting I interviewed the scientists to determine whether the lab
meeting affected their knowledge. I also interviewed the senior scientists
about their conceptualization of the research project. This was a cyclical
process in which I observed the scientists present work a number of times.
By the end of the year, I had collected data on 19 scientific research pro-
jects. In addition to recording laboratory meetings I conducted interviews
with members of the laboratory, was given copies of grant proposals and
drafts of papers, and attended lectures by the senior scientists and many
impromptu meetings. Thus, I collected data on all aspects of scientific re-
search with the laboratory meeting as the central focus.

465
 KEVIN DUNBAR

The Laboratories
Data on 21 scientists in the four laboratories were collected, as well as data
from the 4 senior scientists. My current analyses focus on the 4 senior sci-
entists and 19 scientists in the laboratories. Twelve of the scientists were
postdoctoral fellows, 5 were graduate students, and 2 were research tech-
nicians. The four laboratories that were studied were either developmen-
tal biology labs or just worked with pathogens (disease-causing viruses
and bacteria). Furthermore, the senior scientists varied in terms of expe-
rience. Two were full professors, 1 was an associate professor, 1 one was
an assistant professor. By varying the types of subdomains that the scien-
tists work in and their level of experience, it is possible to determine
whether these factors influence their research.
All the scientists allowed me free access to their laboratories, to inter-
view anyone in the laboratory, to attend any meeting, to read and keep
copies of their grant proposals (including the pink sheets), to attend their
talks and lectures, and to read drafts of their papers. Thus, I was given full
access to the day-to-day activities of the laboratories. In addition, the sci-
entists frequently asked me to attend impromptu meetings and discus-
sions, and they often called when they felt that interesting events were oc-
curring in the lab.
I selected research projects for study on the basis of whether the pro-
ject had just started or was about to begin. In addition, I consulted ex-
tensively with the senior scientists in choosing the research projects to in-
vestigate. Once I had selected the projects, I then met with the senior
scientists, postdocs, graduate students, and technicians that were involved
in the research. All members of the four laboratories agreed to cooperate.
Laboratory A. Laboratory A was run by a senior scientist who has over
300 publications and numerous awards. This laboratory has had many dis-
coveries that have appeared on the front page of the New York Times, Sci-
ence, Nature, Cell, and so forth. His laboratory consisted of 22 postdoc-
toral fellows, 5 graduate students, and 4 technicians. I selected four research
projects to follow. Two of the four research projects were successful and
led to scientific discoveries. Importantly, neither I nor the scientists in-
volved realized that a discovery was about td be made when I started fol-

466
 HOW SCIENTISTS THINK

lowing their research. It was only after a few months of following the re-
search projects that the discoveries were made. Thus, I had collected data
before, during, and after a discovery had been made. One of the researchers
discovered a new gene that controls cell differentiation, and another had
discovered how certain cells proliferate into certain regions of the body.
Importantly, the latter discovery actually occurred during a laboratory
meeting at which I was present and was tape-recording; that is, I have the
moment of discovery on tape. This project forms the basis of the research
discussed in the section titled "Anatomy of a Conceptual Change." Of the
two remaining projects, one was unsuccessful and the other had not pro-
gressed significantly within the data collection period.

Laboratory B. Laboratory Β was run by a scientist who has made many

important discoveries in molecular biology. He has numerous publications
and has trained many now eminent scientists. His current research pro-
gram involved determining a general model of how certain genes control
traits in a novel type of bacterium. His laboratory had 3 postdocs, 5 grad-
uate students, and 1 technician. I have analyzed two of the research pro-
jects that were conducted in his laboratory. One of the research projects
has resulted in two publications; however, the scientists were unable to
reach their goal of discovering the function of a component of a gene. The
other project made minimal progress.

Laboratory C. Laboratory C was run by an associate professor who

has made a number of important discoveries on how DNA and RNA are
coded in two different types of parasites. The lab consisted of 4 postdocs,
2 graduate students, and 1 lab technician. I followed research projects con-
ducted by the 4 postdocs. All the research projects resulted in significant
breakthroughs that have been published in the major scientific journals
such as Science.

Laboratory D. Laboratory D was run by an assistant professor who is

already famous for his work on viral mechanisms and his creative ap-
proach to uncovering gene function. The laboratory consisted of 4 post-
docs, 6 graduate students, and 2 lab technicians. His current research pro-
gram is centered on discovering the mechanism by which certain genes in

467
 KEVIN DUNBAR

the HIV virus allow the virus to infiltrate into the host organism. He has
evolved a research program that has employed a number of novel and in-
genious techniques to discover how this works. These research projects are
now leading to a new model of an important component of HIV activity
that has wide-ranging theoretical and practical implications for molecu-
lar biology. The director of Laboratory D also invented a new genetic tech-
nique. I was present for the implementation and development of this tech-
nique. This technique has been widely referenced and reviewed in many
major scientific journals.
Data Analysis
Transcription. Transcriptions and coding were done by two indepen-
dent transcribers with backgrounds in molecular biology.
Coding. All coding was conducted by coding the transcriptions into a
computerized database. Multiple coders were used, and reliability checks were
conducted by independent coders. In this section, I provide a very general
overview of the coding techniques used. I provide a more detailed account
of coding in the method discussions of other sections of this chapter. The
basic unit of analysis is the statement or utterance. A statement is essentially
equivalent to a clause or sentence. Statements were chosen as the basic unit
of analysis as they contain a verb phrase, which in turn contains the core
mental operation (proposition or idea) that the presenter is employing at
the time. Thus, we treat statements at meetings in the same way that state-
ments are treated in standard protocol analyses (cf. Ericsson & Simon, 1993).
I used the corpora of statements made to build a representation of scientists'
mental operations. Using techniques borrowed from protocol analyses, state-
ments can be aggregated by episodes, solution steps, and processes. One can
switch between different levels of analyses, depending on the questions that
one is asking of the data. The MacSHAPA coding and database software sys-
tem was used to code the data (Sanderson et al., 1994).

Summary of Results
The research reported in this chapter provides a snapshot of my current
analyses and interpretation of the cognitive processes involved in creativ-

468
 HOW SCIENTISTS THINK

ity in science. I now address three main sources of creative cognition. First,
I present an analysis of the role of analogy. Second, I outline my analyses
of scientists' treatment of unexpected findings. Third, I discuss some of
the findings on distributed reasoning. Finally, I present a case study of a
conceptual change that involved all three of the aforementioned strategies.

ANALOGY
Analogy has been regarded as a very important psychological process in-
volved in creative cognition and has been the focus of intense investiga-
tion over the past 15 years, culminating in a number of detailed models
of the cognitive processes involved in analogical reasoning (e.g., Forbus,
Centner, & Law, 1995; Holyoak & Thagard, 1989,1994).l Accounts of anal-
ogy distinguish between two components of an analogy: the target and
the base. The target is the concept or problem that the scientist is at-
tempting to solve or explain. The base is another piece of knowledge that
the scientist uses to understand the target, or explain the target to others.
When the scientist makes an analogy he or she maps features of the base
onto features of the target. By mapping the features of the base onto the
target new features of the target may be discovered, or the features of the
target can be rearranged so that a new concept is invented, or the scien-
tist can highlight a specific feature of the target for other people. To illus-
trate this discussion of analogy I borrow an analogy that Rutherford
(Rhodes, 1986) ostensibly used in his research. When Rutherford was at-
tempting to understand the structure of the atom he made an analogy to
the solar system. In this case, the target was the atom and the base was the
solar system. Rutherford ostensibly mapped the idea that the planets re-
volve around the sun onto the atom, and he argued that the electrons re-
volve around the nucleus. Thus, a number of historians have argued that
by drawing an analogy to the solar system, Rutherford was able to pro-
pose a new account of the structure of the atom. By mapping the feature

'Many cognitive accounts of analogy start with a reference to analogy in science and have noted that the
types of distant analogies alluded to in the literature on the history of science are rarely used by partic-
ipants in psychology experiments.

469
 KEVIN DUNBAR

of the planets revolving around the sun, Rutherford was able to align his
data with those predicted by a solar analogy. According to this view, the
analogy resulted in a major restructuring of his knowledge, and a scien-
tific discovery was made.2
The Rutherford example highlights two key assumptions that re-
searchers in the creativity literature have made about the role of analogy
in science. The view of analogy in the creativity literature has been that
when a scientist makes an analogy (a) the source is usually from a very
different domain,3 and (b) the role of analogy is to restructure the scien-
tist's knowledge in a gestaltlike manner (e.g., Boden, 1993; Koestler, 1964).
One of the questions I want to ask here is whether this is a valid picture
of the role of analogy in science. The question can be divided into a num-
ber of more detailed questions: Do scientists use analogies at all? If they
do, are they the distant analogies that have been talked about in the his-
torical creativity literature? Do less distant analogies play any role in sci-
ence, as the emprical psychological work suggests (see Forbus et al., 1995;
Holyoak & Thagard, 1994)? Does analogy work alone, or does it work in
conjunction with other mental operations? Is analogy involved in scien-
tific discoveries and conceptual change in science?

Method
I investigated the use of analogy at 16 meetings (4 meetings for each of
the 4 labs). All analogies were coded by two independent coders. Any time
a scientist referred to another base of knowledge to either (a) explain a
concept or (b) use that other base of knowledge to modify the concept, it
was coded as an analogy. Three representative analogies follow:

la. Within organism: An HIV to HIV analogy. "Urn. In the case of HIV
it's 5 bases away, umm. So, um to study RT (reaction time) using a
-There is some controversy about whether the solar system analogy played a causal role in Rutherford's
discovery of the structure of the atom. Whatever the real case may be, my point is that researchers have
used such examples to emphasize the critical revolutionary role that this particular type of distant anal-
ogy plays in scientific discovery and conceptual change.
3
Most cognitive accounts of analogy have made no assumptions about how distant the source and the tar-
get are in science. In fact, Holyoak and Thagard (1994) have made a list of the most important analogies
in science over the past 2,000 years and have found very few distant analogies. However, when researchers
do allude to analogy in science they tend to give examples wherein the source and the target are distant.

470
 HOW SCIENTISTS THINK

substrate that more closely mimics the in vivo situation is difficult.

Because um, number one, you will need to anneal six surface
strands together. Number two, it is really doubtful that since there is
only a five base pair here, where they hold this complicated struc-
ture together."
Ib. Other organism: An Ebola virus to Herpes virus analogy. "The prob-
lem with Ebola is that it is AT rich. So you can't really do some
analysis, analysis of homology with the, uh, genome because of this
very AT rich, uh, richness. That would not be the case for herpes
and could give a better answer for some of the putative homology."
Ic. Nonbiological or distant: Monkeys to PCR (polymerase chain reaction)
analogy. "You know, just because you can see 10 molecules that still
isn't working in my book. A monkey will eventually type Shake-
speare, given the opportunity. PCR is not unlike that. You do it a
billion times and you probably will find one thing that happened to
be right."

Note that instances where a scientist stated that X was like Υ were not
coded as analogies. That is, statements of similarity that neither gave ex-
planations nor resulted in the mapping of features from the base to the
target were not coded as analogies. Once the analogies were found, they
were coded along a number of dimensions. The coding dimensions are
specified in the section dealing with that dimension.

Results
Frequency of Analogy Use
There were 99 analogies used during the 16 meetings (M =6.1 analogies
per meeting). The range of analogy use was 2 to 14 analogies per meet-
ing. All four labs used analogies. There were a total of 25,30,31,13 analo-
gies, respectively, for Labs A, B, C, and D. Thus, analogies were frequently
used at laboratory meetings.
Range of Analogy Use
The range over which the analogies were used was coded. Range is an in-
dex of how far apart the base and target were for each analogy. Analogies

471
 KEVIN DUNBAR

were coded as being within organism, other organism, or nonbiological.

Within-organism analogies are those wherein the base and the target are
from within the same organism. In the previous within-organism exam-
ple (la), the scientist has drawn an analogy between the way the HIV virus
works in an in vivo context and how an in vitro HIV could be made by
mapping from the in vivo HIV onto the in vitro HIV construct. Other-
organism analogies are those in which the base and the target are from two
different organisms, as in Example Ib (analogy between the Ebola virus
and the Herpes virus). In this analogy, the scientist points out the differ-
ences between Ebola and Herpes to show why Ebola is a better organism
to research a particular question. Nonbiological or distant analogies are
those in which the base is taken from a nonbiological domain. In Exam-
ple Ic (Nonbiological or distant: Monkeys to PCR analogy), the scientist
highlights the fact that a finding could be due to chance by drawing an
analogy between a monkey typing Shakespeare and the polymerase chain
reaction, generating a chance result. Note that this type of distant analogy
has received the most attention in the literature.
Almost all of the 99 analogies were either within organism (40) or
other organism (57). There were only 2 nonbiological analogies. Thus, the
bulk of analogical reasoning happened when the base and targets were
from the domain of biology. This result is very important. Most accounts
of analogy in science focus on distant analogies, yet only 2 of the 99 analo-
gies used by the scientists were of this type.

Goals and Analogy Use

Categories of goals were formulated by searching for goals in the database
rather than imposing them on the data a priori. From this emerged four
dominant goals: formulate a hypothesis, design an experiment, provide an
explanation, and fix an experiment (when an experiment went awry the
scientists often drew analogies to procedures used in other experiments
and proposed replacing one step in the faulty experiment with a step from
an analogically similar experiment). Almost half (45) of the analogies
occurred when the goal was to provide an explanation. Usually, the
explanations were of methodological issues. There were 21 analogies for

472
 HOW S C I E N T I S T S THINK

design an experiment, 10 for fix an experiment, and 23 for formulate a

hypothesis.
Next, I discuss the relation between goals and range. Table 1 reports
the number of analogies for each combination of goal and range. The table
shows a number of interesting relations between a scientist's goals and the
range over which the analogy is drawn. First, it can be seen that the two
nonbiological analogies were used to make explanations; they were not
used to formulate hypotheses. Although there were only two nonbiologi-
cal analogies in the 16 meetings coded, there were two other nonbiologi-
cal analogies in the database. All four of these nonbiological or distant
analogies were used to explain a concept to members of the laboratory.
Thus, nonbiological or distant analogies are rare and generally used for
explanations rather than to generate new hypotheses and concepts.
I now turn to a discussion of the within-organism and other-
organism analogies and goals. There was little difference in range between
designing and fixing experiments. Scientists were equally likely to draw an
analogy from the same organism or a different organism when designing
or fixing an experiment. The major interaction of goals with range was in
hypothesis generation. The scientists tended to use analogies to other or-
ganisms when formulating a new hypothesis. For example, a scientist
might in attempting to determine the function of a gene in one organism
(e.g., a gene in malaria) draw an analogy to a gene in another organism

Table 1

Scientists' Goals for Within-Organism, Other-Organism,

and Nonbiological Analogies

Within Other
Type of goal organism organism Nonbiological

Hypothesis 3 20 0
Design experiment 9 12 0
Fix experiment 5 5 0
Explain 23 20 2

473
 KEVIN DUNBAR

(e.g., a similar gene in clams). If the scientist knows what the gene does
in one organism (e.g., in clams), she or he can then map the functions of
that gene over to the organism that they are working on (e.g., the similar
gene in malaria). Thus, rather than the source of hypotheses being analo-
gies made to nonbiological or distant domains, when formulating hy-
potheses the scientists make analogies to other organisms.

How do Scientists Generate Their Analogies?

How do scientists retrieve the sources for between- and within-organism
analogies? One possibility is that the scientists recall specific experiments
conducted in their labs or journal articles they have read. If we break
down the range of the analogies by whether the scientists were recalling
specific cases when they were making analogies (such as specific experi-
ments that were conducted in the past, references to particular research
articles, or experiments conducted by researchers in the field), we can see
that 31 of the 40 within-organism analogies recalled a specific case. In con-
trast, only 6 of the 57 other-organism analogies recalled a specific case.
Thus, when scientists make analogies to the same organism, they tend to
recall a specific case. However, when scientists make analogies to a differ-
ent organism, they do not recall specific cases. In addition, 22 of the 31
within-organism analogies recalled cases of previous experiments con-
ducted in the lab. Thus, when the researchers made analogies to the same
organism, the bulk of the analogies were to previous experiments con-
ducted in that lab.
How did the scientists use analogies to other organisms without re-
calling a specific case? Psychological research has shown that individuals
have great difficulty going outside their current problem to make an anal-
ogy (e.g., Gick & Holyoak, 1983), yet the scientists were able to transcend
this problem. How? An analysis of analogies to other organisms revealed
that the scientists had two main ways of circumventing this analog re-
trieval problem. First, molecular biologists have a tool available to them
that gives them another way of retrieving base analogs: homology. Sec-
ond, the scientists had an abstract knowledge of the biological mecha-
nisms that exist in other organisms. Scientists can use their knowledge of

474
 HOW S C I E N T I S T S T H I N K

biological mechanisms to search memory for organisms that use a par-

ticular biological mechanism.
I discuss each of the ways of retrieving base analogs in turn. Scientists
use homology to determine the molecular structure of a gene by se-
quencing each base pair in the gene. The scientists then type the sequence
of their gene into a computer and search a database of genes for a gene
that has a similar coding. If the scientist finds a gene or genes with a sim-
ilar sequence (i.e., a homologous gene), and the function of that gene is
known, the scientist can infer that the gene may have the same function
in their organism. That is, the scientist maps the function of the homol-
ogous gene onto the gene being investigated. Thus, homology allows the
scientist to both retrieve analogs and propose new hypotheses about gene
function. Not only does the homology allow the scientist to infer new hy-
potheses concerning the biological function of the gene, but the scientist
can also use the methodologies that the previous researchers used when
conducting their research. Importantly, the same homology can provide
new hypotheses and new methods that the scientist can use in his or her
research. The scientists in my study generated 31 of the 57 other-
organism analogies by using homology. Thus, homology allowed them to
generate other potential base analogs. As can be seen from Table 2, the sci-
entists used homology to infer biological mechanisms and the methods
that they should use in their experiments.
I now turn to analogies to other organisms that were not based on
homology. The scientists were more likely to use biological mechanisms

Table 2

Types of Knowledge Retrieved by Analogies on the Basis of

Homology and Nonhomology for Other-Organism Analogies

Type of knowledge Homology Nonhomology

Biological mechanism 10 16
Experimental method 17 7
Problems with methods 4 3

475
 KEVIN DUNBAR

as a retrieval cue. For instance, the scientists might think that "E.coli per-
forms a particular function by splicing the protein at the AT site; perhaps
our organism splices the protein in the same way." The scientists' knowl-
edge of biological mechanisms is often tied to particular organisms, and
these organisms become part of the analogy.

Summary and Discussion of Analogy Results

Analogy was frequently used in all of the laboratories. Most of the analo-
gies that were observed in the current study were biological. Only 2 of the
99 analogies were nonbiological or distant. These findings shed new light
on the role of analogy in science. Most historical accounts of analogy in
science have tended to focus on very distant analogies; yet, the results of
these investigations suggest that distant analogies are not an important
component of contemporary science. There are a number of reasons for
the differences between these findings and those discussed in the litera-
ture on the history of science or the creativity literature. First, many of the
distant analogies that scientists have mentioned in the history of science
may not have had a role in the making of a discovery. In fact, a number
of historical analysts have argued that the Rutherford solar system anal-
ogy, and the snake analogy mentioned by Kekule in his discovery of the
structure of the benzene ring had no role in the respective discoveries
(Rhodes, 1986; Wotiz & Rudofsky, 1984).4 The data presented here sug-
gest that it may be the case that scientists use distant analogies to explain
a new concept to an audience rather than that distant analogies have a
causal role in making a discovery. I am currently monitoring the scien-
tists' publications to see if more distant analogies seep into their accounts
of their findings.5 Second, the types of analogies that the scientists use in
on-line reasoning are easy to forget. In fact, in postlab meeting interviews

4
I thank Bill Brewer for bringing this article on KekulS's discovery of the benzene ring to my attention.
5
Some researchers have suggested that perhaps the evidence of distant analogies is an index of the matu-
ration, or lack thereof, of the development of a field (with presumably a higher incidence of distant analo-
gies occurring at the beginning of a field). However, there is nothing in my data that supports this view.
Note that the scientists in my study were pioneering totally new concepts, in an uncharted conceptual
space. In this view one would expect to see many distant analogies relative to the other types, which was
not the case.

476
 HOW SCIENTISTS THINK

the scientists rarely remembered the analogies that were generated during
the meeting. Thus, analogies are often used as a scaffolding that the sci-
entists use in the construction of new theories and methodologies. Once
the new concepts and methods have been advanced the analogy can be
discarded. Many of these analogies will not make their way into the note-
books of the scientists, and thus the historical record will not show that
the within-organism or other-organism analogies had a role in the dis-
covery of a new concept or invention of a new method.6
More than one analogy may be involved in a particular discovery, and
one particular analogy may not be responsible for a particular conceptual
change, but a group of quite different analogies may be causally involved
in making a breakthrough. Again, because no one analogy made a major
restructure of knowledge, the scientists may not have recalled a particu-
lar within-organism or other-organism analogy as being a factor in the
discovery. However, when all of the analogies that are involved in making
a discovery are examined, only analogies of very specific types will be seen
to have played a major role in scientific reasoning and discovery. More-
over, as is shown later, analogy is not the only mechanism that comprises
conceptual change. In the next three sections I show that other key cog-
nitive mechanisms produce conceptual change. Thus, analogy, while im-
portant, is but one of a complex of mechanisms that produce conceptual
change. At the close of this chapter, I explain what this complex of mech-
anisms is and show how together they contribute to scientific discovery
and conceptual change.

UNEXPECTED FINDINGS AND

CONFIRMATION BIAS
There is a large literature in psychology and philosophy of science on what
happens when scientists get unexpected results from their experiments. In
the psychological literature researchers have investigated this in terms of

*M>· in vitro investigations of analogical reasoning also reveal that research participants have little aware-
ness of, or memory for, the mental steps involved in making a discovery, even directly after having made
a major conceptual shift (Dunbar & Schunn, 1990; Schunn & Dunbar, 1996).

477
 KEVIN DUNBAR

confirmation biases; individuals tend to seek evidence that is consistent

with their hypothesis and ignore evidence that is not. Researchers have re-
peatedly found evidence that research participants engage in this type of
r> behavior and have argued that scientists have similar reasoning biases (cf.
- Klayman & Ha, 1987; Tweney, Doherty, & Mynatt, 1982). However, before
one accepts the generality of the results of these types of experiments it is
important to note some of the large differences between the tasks that re-
search participants perform and what scientists do. Most important, there
fc is no actual scientific knowledge involved in the psychological tasks; the
to-be-discovered concepts are arbitrary, and the links between hypothe-
sis, experiment, and data are straightforward. For example, the 2-4-6 task
i. is one that has been widely used (Wason, 1960). In this task, the experi-
- menter asks an individual to determine the rule underlying a sequence of
•5 numbers. The individual is given a triad of numbers, such as the numbers
g 2, 4, 6, and is told that this number triad is an example of the rule. The
ο individual is then told that she or he can generate other triads and that
-' the experimenter will determine whether the triad is an example of the
> rule. Finally, the individual is told that when certain she should state the
- rule. Many research participants tend to generate triads that are consis-
- tent with their hypotheses; they attempt to confirm their hypotheses.7 On
Ir the basis of experiments such as these, researchers have argued that this
- confirmation bias is a general phenomenon that both lay people and sci-
''. enlists must avoid if they are to reason correctly.
Although the confirmation bias view of science has received much em-
pirical support, another related phenomenon is the issue of unexpected find-
ings. A number of researchers have argued that a useful strategy in science
is to focus on unexpected findings. According to this view, scientists work
with a heuristic such as "if the finding is unexpected, then set a goal of dis-
}:
covering the causes of the unexpected finding" (cf. Dunbar, 1993, 1996;
f
Kulkarni & Simon, 1988). This view of reasoning is quite different from that
1
implied by the confirmation bias viewpoint. According to this viewpoint,

7
The common rule that research participants must discover is "numbers of increasing magnitude." Re- ·
search participants generally propose the rule "even numbers increasing by 2" and only generate triads
consistent with this rule.

478
 HOW S C I E N T I S T S T H I N K

when unexpected findings are inconsistent, scientists should focus on the

finding rather than ignore it. Thus, there are two conceptions of what sci-
entists may do. Of course, it may be the case that under certain circum-
stances the scientists may focus on unexpected findings, and under other
circumstances they may ignore the findings and behave like the participants
in psychology experiments (cf. Tweney, 1989). The goal of the following
analyses was to investigate these questions in a real scientific environment.

Method
My graduate student Lisa Baker and I decided to investigate the role of
unexpected findings by analyzing the scientists' reactions to unexpected
findings at four laboratory meetings in Lab A (see also Baker & Dunbar,
1996). We chose Lab A because scientists there had obtained many ex-
pected and unexpected findings and thus provided much data to investi-
gate these issues. We had two independent coders code every unexpected
finding in each of the four lab meetings. All findings in which the scien-
tist had previously predicted a different result, or expressed surprise at the
obtained result, were coded as unexpected. All findings that were consis-
tent with the predictions were coded as expected. A third category of find-
ings consisted of those that occurred in exploratory experiments. Here the
scientist did not have any predictions one way or the other and conducted
the experiment to see what would happen. The results of these types of
experiments were coded as exploratory.

Results
Our first step was to determine how common expected, unexpected, and
exploratory findings were? In four meetings there were six experiments
reported with 70 conditions. There were 22 expected, 18 unexpected,
and 30 exploratory findings. Clearly, unexpected findings are common.
We coded all expected and unexpected findings on the basis of whether
the scientists tried to explain away their results or whether they built
theories with the findings. To do this, we coded the number of reason-
ing blocks the scientists engaged in following both expected and unex-
pected findings. A reasoning block was a group of statements that in-

479
 KEVIN DUNBAR
1 VI ! Κ Τ

volved reasoning about a particular finding. One finding can generate

many different reasoning blocks. The number of reasoning blocks gen-
erated by expected and unexpected findings can be used as an index of
how much attention scientists give to these types of findings. There was
more reasoning for unexpected (179 reasoning blocks) than expected
(42) findings. Furthermore, when confronted with unexpected findings,
scientists were much more likely to engage in theory building than to
attempt to explain the results away. Thus, 161 reasoning blocks were con-
cerned with theory building, and 18 reasoning blocks were concerned
with attributing the result to some sort of error. Thus, scientists do pay
attention to unexpected findings.
The previous analyses applied purely to whether the scientist who con-
ducted and presented the research was likely to attend to an unexpected
finding. We next investigated whether the group also attended to unex-
pected findings. We calculated the number of reasoning blocks that any-
one other than the presenter devoted to unexpected and expected find-
ings. Again, we found much more reasoning by the group when faced with
unexpected compared with expected findings. As a measure of group at-
tention to unexpected findings we also counted the number of interac-
tions for expected and unexpected findings. We found 23 interactions for
expected findings and 176 interactions for unexpected findings. These re-
sults indicate that the group also pays attention to unexpected findings
and uses the findings to propose new hypotheses and experiments.
Another question that can be asked about the scientists' use of unex-
pected findings is whether there was any difference between the scientists'
treatment of unexpected findings that were consistent with their hypoth-
esis and those that were inconsistent with their hypothesis. An unexpected
finding that is consistent with a scientist's hypothesis can occur, for ex-
ample, when the scientist expects a certain type of result to occur, but the
size of the effect is much greater than expected. In this type of situation,
the result is consistent with the hypothesis but the size of the effect is un-
expected. An unexpected inconsistent finding is one in which a qualita-
tively different type of outcome occurs. We coded the 18 unexpected find-
ings along these dimensions and found that 8 unexpected findings were

480
 HOW SCIENTISTS THINK

consistent with scientists' expectations and that 10 unexpected findings

were inconsistent with their expectations. We then coded the findings that
resulted in the proposal of new hypotheses. We found that 4 of the 8 con-
sistent findings resulted in new hypotheses and 8 of 10 inconsistent find-
ings resulted in new hypotheses. These results indicate that the scientists
attended to the unexpected findings even when the findings were incon-
sistent with their hypothesis.
We recently have been conducting new analyses of scientists' reactions
to unexpected findings to determine whether the time at which an unex-
pected finding occurs affects whether an unexpected finding is attended
to. We found that there are two dimensions of an unexpected finding that
determine whether the unexpected finding is attended to. The first is
whether the unexpected finding is unexpected relative to a core hypothe-
sis in the field or to an auxiliary hypothesis that the scientist has proposed
to get the experiment to work. Another dimension is how early or late in
the research project the unexpected finding occurs. We found that when
the unexpected finding occurs early and is not a core hypothesis, the sci-
entists will not devote much attention to it. However, if the unexpected
finding occurs early and is unexpected relative to the central assumptions
of the field, the scientists will focus on the finding. When the unexpected
finding occurs late in the research project the scientists will attend to it
regardless of whether it is a core or an auxiliary hypothesis. Note that the
situation in which the scientists ignore unexpected findings is very simi-
lar to that of individuals in psychology experiments: The individuals are
early in the experiment, and the hypotheses are not core assumptions.
Thus, it is only under very restricted circumstances that one finds a sim-
ilarity between the results of psychology experiments on confirmation
bias.
Our analyses of unexpected findings indicate that scientists do attend
to unexpected and inconsistent findings. Why do the scientists attend to
unexpected and often inconsistent findings? One reason is that in real sci-
ence unexpected findings are frequent. The fact that unexpected findings
are frequent may have a major effect on the scientists' ability to deal with
these types of findings. It may be the case that the longer a scientist is in

481
 KEVIN DUNBAR

the field, the more unexpected findings the scientist has encountered and
the more likely it is that the scientist has developed strategies or heuris-
tics for dealing with them. Thus, the way a scientist deals with unexpected
findings depends on the specific strategies he or she has developed to at-
tempt to reconcile them.8 Participants in psychology experiments are un-
likely to have developed strategies for dealing with unexpected findings
and may prefer to focus on their current goal, ignoring unexpected results
(as in Dunbar, 1993). As they encounter more and more evidence that is
inconsistent, they are eventually forced to attend to unexpected findings.

DISTRIBUTED REASONING
Most cognitive research on scientific reasoning focuses on individuals rea-
soning about a problem. However, much of modern science is conducted
by groups of scientists rather than individuals. Furthermore, much of the
cognitive work has demonstrated that individuals make many different
types of reasoning errors. In this section, I investigate whether reasoning
in groups can circumvent certain individual reasoning errors. In particu-
lar I explore the issue of distributed reasoning in science. Distributed rea-
soning happens when different members of a group reason about topics
such as a hypothesis, experiment, methodology, or interpretation of a re-
sult while adding new elements to the topic under discussion. The ques-
tion that I will ask is whether distributed reasoning of this sort helps
circumvent problems that individual participants display in standard ex-
periments.
One of the major tasks for both individuals in psychology experiments
and scientists confronted with new data is to determine what types of in-
ductions to make from new data. There are infinitely many inductions that
can be made from a set of data, and this is a potential place where differ-

"Lovett and Anderson (in press) have shown that history of success plays a role in determining what strat-
egy research participants use to solve a problem. They have shown that research participants use both
their history of success and the current problem-solving context to determine the type of problem-solv-
ing strategy to use. I argue that scientists use a similar sei of heuristics. Whether they will use unexpected
findings or not will depend on both the history of success and current context.

482
 HOW SCIENTISTS THINK

ent members of the group can make different inductions from the same
data. To examine this, I explored the role that the group played in the types
of inductions that a scientist in an HIV lab made during his talk. At this
talk the scientist presented five sets of findings and made 11 inductions
about the mechanisms that the HIV virus uses. The members of the lab
often disagreed with the inductions that the scientist made and modified
his inductions. The other members of the lab limited (3), expanded (1),
replaced (2), or discarded (1) a total of 7 of the 11 inductions.
This pattern of challenging inductions was ubiquitous across all labs
and provides important information about the role of distributed rea-
soning. Individuals have great difficulty generating alternative inductions
from data and also have great difficulty limiting and expanding induc-
tions. Distributed reasoning helps circumvent these difficulties. When dis-
tributed reasoning occurs, the group quickly focuses on the reasoning that
has occurred, and the other members of the laboratory generate different
representations. These new representations make it possible for them to
propose alternative inductions, deductions, and causal explanations. Thus,
distributed reasoning provides new premises and models that an individ-
ual may not be able to generate when reasoning alone.
Another issue relevant to distributed reasoning is the number of peo-
ple involved in the reasoning. In the previously mentioned HIV lab I in-
vestigated the number of inductions and deductions that were shared. That
is, how many inductions and deductions occur in which one premise is
provided by one person, and another premise is provided by another per-
son. We found that 30% of inductions and deductions were shared by
more than one individual. We also found that 12% of all inductions and
deductions had more than two participants. Furthermore, inductions of
one individual sometimes formed the basis of a deduction for other indi-
viduals.
Distributed reasoning consists of scientists performing cognitive op-
erations on information (e.g., induction) and then passing the results of
the operation on to other scientists in the group. The other scientists then
use the results of the first operation as the input to further cognitive op-
erations. Together, the results of these cognitive operations are then used

483
 KEVIN DUNBAR

to build new cognitive representations: scientific theories and new exper-

iments. How and when the information is passed between individuals de-
pends on the goals of the individuals and the group, as well as the knowl-
edge bases that the scientists have at their disposal. The generation of
different representations during distributed reasoning helps scientists cir-
cumvent one of the major reasoning difficulties that individuals have: that
of generating alternative hypotheses, explanations, theories, and experi-
ments.
The results of these analyses of distributed reasoning are different from
the results of brainstorming experiments and creativity in group experi-
ments. Many studies have shown that when a group of people is asked to
generate novel concepts, the group performs no better than individuals.
However, in the research reported in this chapter it can be seen that groups
of scientists do generate new concepts and that distributed reasoning is
an important factor. The difference in findings is twofold. First, in psy-
chology experiments the participants are not part of a group sharing com-
mon knowledge and values. Usually, individuals in psychology experi-
ments are randomly thrown together for the purpose of the experiment.
Second, the types of problems given to the participants are often arbitrary
and require little background knowledge. In the science labs investigated
in this chapter, the scientists had overlapping backgrounds and shared
goals about the research. Furthermore, the members of the lab had slightly
different types of knowledge that they could bring to bear on the prob-
lem. Taken together, these results suggest that entirely new experiments
on group reasoning need to be conducted using real groups' reasoning
about real problems, with significant background knowledge and diver-
sity of knowledge. The prediction is that in this type of situation groups
of individuals perform more creatively than individuals.

ANATOMY OF A CONCEPTUAL CHANGE

The account of the cognitive processes underlying scientific creativity of-
fered so far is static. I have demonstrated that analogy is an important part
of current day science, that scientists reason about unexpected findings,

484
 HOW SCIENTISTS THINK

and that distributed reasoning is a potentially important concept in sci-

ence. I now turn to the issue of how all three aspects of scientific reason-
ing form a complex of mechanisms that work together to produce a con-
ceptual change in a group of scientists at a meeting.
Many recent analyses of theory change in the history of science have
focused on the notion of conceptual change (e.g., Carey, 1992; Nersessian,
1992; Thagard, 1992). Conceptual change has been defined as changes in
scientific theories that occur when new concepts are proposed and old
concepts must be radically changed or replaced to accommodate the new
concepts. One example of this type of conceptual change noted in the lit-
erature is the 16th-century shift from a unitary concept of heat and tem-
perature to two new concepts: one involving heat and one involving tem-
perature (Wiser & Carey, 1983). It is conceptual change of this type that
I now discuss.
Here I provide a dynamic account of a conceptual change that oc-
curred in Lab A and use this example to show how in this situation dif-
ferent forms of reasoning worked together to produce entirely new con-
cepts. To preserve the anonymity of this lab, I have been obliged to change
the names of the diseases and the specific mechanisms involved in the dis-
eases. Alas, I have also had to render intentionally vague specific aspects
of the scientists' discussion that factored critically in the conceptual
change. Nonetheless, I have tried hard to leave intact the essence of the
complex of mechanisms that contributed to these scientists' conceptual
change.
Let me begin with some background on the discovery. A postdoctoral
fellow had recently come to a world-famous immunology lab. He had de-
cided to investigate the way that B-cells cause a particular autoimmune
disease. He had been conducting experiments in collaboration with an-
other postdoc in another lab. Their work began with an analogy. Twenty
years before, a researcher had noticed that an autoimmune disease in rab-
bits called CVX was very similar to a human autoimmune disease. Since
then, the CVX diseases in rabbits has been used as a model for the human
disease LOA. The postdocs investigated the disease in yet another organ-
ism (hamsters) because the postdoc's lab used hamsters and had facilities

485
 KEVIN DUNBAR

that could be used to investigate the mechanisms underlying the CVX dis-
ease that few other laboratories had. Overall, the motivation for his re-
search was based on analogies between the human LOA disease, the rab-
bit CVX disease, and the hamster CVX disease.
One May afternoon the postdoc gave a talk about his latest experi-
ments. He began with analogies between the human LOA disease and the
CVX disease in rabbits, noting where the similarities and differences be-
tween the two diseases arose. He then moved to analogies between the
CVX disease in rabbits and in hamsters. The first set of experiments re-
sulted in a small amount of discussion and suggestions for future exper-
iments. Then the postdoc started to discuss some experiments in which
the results were very unusual. The postdoc had conducted a straightfor-
ward experiment. He had two conditions: one that caused colmenia dis-
ease in the joints and the other that caused the CVX disease in the heart.
Both the heart and the joints are immune-privileged sites that do not nor-
mally allow B-cells in. In fact, the only types of B-cells that have been
found in the heart are CVX B-cells, and the only B-cells that have been
found in the joints are the colmenia B-cells. The postdocs expected that
the B-cells that cause the disease in the heart would go to the heart and
the B-cells that cause the disease in the joints would go to the joints. In-
stead, they found both types of B-cells in the heart and in the joints. This
was an unexpected finding. The postdoc reached the part of his presenta-
tion wherein he discussed these results. He was surprised and excited by
what he found. The result was unusual. It was at this point that the con-
ceptual change began to unfold.
The director of the lab was intrigued. He asked the postdoc how it
happened. The postdoc said he did not know. The director then made the
question more specific. He asked the postdoc what properties were com-
mon to the colmenia and CVX B-cells that allowed them entry into the
heart. The postdoc made an analogy to some other experiments that an-
other postdoc in the lab had conducted and induced that the CVX and
colmenia cells were both methylated. The director and other postdocs in
the lab then made a series of inductions and deductions that led to a causal
explanation for the unexpected finding. The reasoning was distributed

486
 HOW SCIENTISTS THINK

over the members of the lab. However, the explanations they offered did
not account for some other aspects of the findings, and another round of
distributed reasoning occurred. This distributed reasoning resulted in a
conceptual change: They proposed two new biological mechanisms to re-
place the unitary concept they had all assumed up to that point. Previ-
ously, it had been assumed that CVX cells only go to the heart and that
colmenia cells only go to the joint; that is, B-cells have organ-specific at-
tractions. The assumption was that once these cells got into the organ,
they started the disease in that organ. Thus, there was one mechanism that
caused both the entry into the organ and the initiation of the disease. The
members' distributed reasoning led them to the conclusion that entry into
the organ and the initiation of the disease were caused by two different
mechanisms. They then had to propose what these mechanisms could be.
They proposed two mechanisms that could together account for the CVX
B-cells' causing the disease. One postdoc drew an analogy back to the hu-
man disease and mapped the mechanisms that had been proposed for the
CVX disease onto the human disease. They modified their new model to
fit the analogy to the human disease and thus ended up proposing a new
model that not only explained the mechanisms underlying the three dis-
eases but also had major ramifications for whole classes of autoimmune
diseases.
By proposing two new mechanisms the scientists had to also change a
number of other concepts in their knowledge of autoimmune diseases. It
was at this point that everyone in the lab realized that a conceptual change
had occurred, and they all shouted in excitement. This was followed by some
further analogies in which other postdocs suggested other experiments. Fi-
nally, a postdoc made an analogy to the methods that other researchers have
used and the methods that the postdoc had used, explaining why their ri-
val's lab had not made the discovery they had just made.
This account of a conceptual change reveals some important charac-
teristics of the mechanisms underlying conceptual change. First, there was
no one reasoning mechanism underlying the conceptual change. Analogy,
induction, deduction, causal reasoning, and distributed reasoning were all
involved. Second, analogy was a significant component of the conceptual

487
 KEVIN DUNBAR

change, but all of the analogies that were used were either to the same or-
ganism or to other organisms: Conceptual change can and does occur
without distant analogies. Third, the scientists had little memory for any
of the on-line analogies used at the meeting. I asked the postdoc who con-
ducted the research how the "discovery" was made. I asked this question
one week later, a month later, three months later, and nine months later.
On none of these occasions did he recall the spontaneous analogies used,
or that distributed reasoning was involved. Thus, much of the on-line cog-
nitive processes that went into the conceptual change would have disap-
peared without a record if I had not taped the original meeting.

CONCLUSION: CREATIVE C O G N I T I O N
IS A TINKERER
The investigation of the cognitive mechanisms involved in on-line scien-
tific thinking and reasoning reveal a number of important mechanisms
underlying creative cognition. The main idea is that no single cognitive
process is responsible for creative thought. I have found that scientists use
a variety of cognitive mechanisms to produce any single new concept or
theory. Creative ideas and novel concepts arise through a series of small
changes produced by a variety of different cognitive mechanisms. It may
be the case that reasoning and conceptual change are related in much the
same way that a series of minor mutations produce major changes in or-
ganisms during evolution. In conceptual change, small mutations in con-
cepts occur due to analogy and other reasoning mechanisms. Overall, a
series of small changes will produce major changes in a concept. Concep-
tual change, like evolutionary change, is the result of tinkering. From a
psychological point of view this account of conceptual change explains
why it is so hard to discover the underpinnings of creativity. The many
incremental steps that are involved in creative cognition are often lost and
forgotten, and the act of creation becomes a mythical entity in which the
final step in the creative process is often seen as the cause of the new con-
cept. This leads to the proposal of entities such as distant analogies and
insight as more important in creativity than they really are.

488
 HOW SCIENTISTS THINK

A further question is whether the cognitive processes underlying cre-

ative conceptual change are different from the processes underlying sim-
ple changes in concepts. I would argue that they are not. Exactly the same
types of cognitive processes that are involved in the more mundane as-
pects of conducting science were involved in the moments of true con-
ceptual change outlined previously. The question then arises as to what
has made these scientists so productive and what has launched them to
the forefront of their fields? The answer lies in their choice of research
topics. Each of the scientists has developed research programs around dif-
ficult topics for which there were few simple answers or an abundance of
ready-made techniques available. To conduct their research the scientists
had to invent new techniques and engage in research that was risky. Thus,
the factor that unifies the creative scientists in this sample is their ability
to take risks. Each of the scientists conducted both high- and low-risk ex-
periments in their laboratories. Although taking risks does not in itself
lead to success, risk taking in combination with the use of the various rea-
soning strategies discussed in this chapter provide the context in which
discoveries can be made.
The view of creativity offered here is quite different from that offered
in the creativity literature. Authors such as Boden (1993) have proposed
that the main way that analogy is involved in creative discoveries in sci-
ence is by having major restructuring of concepts. Here, I have argued that
analogy is involved in a very different way. Many very specific analogies
are made that in conjunction with other reasoning mechanisms produce
both modifications in existing concepts and entirely new concepts. The
reason for the difference between my conclusions and that of others in the
creativity literature is the differences in methodologies used. By looking
at on-line reasoning rather than scientists' patchy reconstructions of a sci-
entific discovery or breakthrough, it is possible to discover the specific cog-
nitive mechanisms underlying creative thought. As I have argued elsewhere
in this chapter, much of the cognition involved in creative thought works
as a form of scaffolding. Once a new concept is generated the cognitive
scaffolding is thrown away and scientists cannot reconstruct the cognitive
steps that went into the discovery. Because of this, scientists, like histori-

489
 KEVIN DUNBAR

ans, reconstruct their creative moments, often from their lab books. Un-
fortunately, many of the key cognitive steps made in a discovery do not
end up in the lab books. Thus, many of these reconstructions are based
on partial information, and, as a result, myths surrounding the creative
process develop.
An important question about the research presented in this chapter is
whether the findings are generalizable to other domains. There are nu-
merous reasons to expect that these findings are indeed generalizable. First,
I have observed similar types of reasoning strategies in biology laborato-
! ries at other universities (Dunbar, Patel, Baker, & Dama, 1995). Second,
; we have observed similar reasoning in clinical situations wherein medical
Γ doctors reason about patients (Dunbar et al., 1995; Patel, Dunbar, & Kauf-
j man, 1995). I am now starting to investigate whether the same types of
reasoning strategies occur in a business context.
Molecular biologists have some special tools that other scientists and
| nonscientists do not have, such as the use of the structure of DNA to search
I for homologies in a database. These scientists have the advantage of a way
| of representing their data that makes it possible to quickly and efficiently
! search for analogs. By representing their knowledge in a standardized fash-
i ion and searching for structural patterns that are similar to the one they
ί are interested in, the scientists solve the problem of how to retrieve rele-
I vant analogies. Thus, homology makes finding base analogs easier. Ulti-
ί mately, using homology gives the scientists another route to access base
; analogs. Once the scientists retrieve these base analogs, they use the same
i cognitive processes for constructing analogies as they do when they search
their own memories for base analogs. Can scientists in other domains re-
trieve analogs in a similar fashion? The answer depends on the way the
knowledge in a field is codified. If knowledge is coded in a structural man-
ner, then it should be possible for the scientists to search for analogs with
a similar structure and generate new analogies. It will be interesting to see
whether the new databases that have arisen in virtually all fields will al-
low scientists to encode structural information, thereby allowing the sci-
entists in a field to retrieve source analogs. This would ease one step in
drawing analogies and could serve as a useful aid to scientists in all fields.

490
 HOW SCIENTISTS THINK

Overall, the research reported in this chapter demonstrates that it is

possible to investigate complex creative cognition in real-world contexts.
This in vivo research makes it possible to discover fundamental mecha-
nisms of creative cognition and how multiple cognitive processes work to-
gether to produce conceptual change. Furthermore, this in vivo approach
both makes it possible to discover what aspects of in vitro research are
generalizable and suggests new types of experiments that can be conducted
in the cognitive laboratory.

REFERENCES
Baker, L, & Dunbar, K. (1996). How scientists use unexpected findings. Manuscript
in preparation.
Boden, M. (1993). The creative mind: Myths and mechanisms. New York: Basic Books.
Carey, S. (1992). The origin and evolution of everyday concepts. In R. N. Giere (Ed.),
Minnesota studies in the philosophy of science. Vol XV: Cognitive models of sci-
ence (pp. 89-128). Minneapolis: University of Minnesota Press.
Dunbar, K. (1993). Concept discovery in a scientific domain. Cognitive Science, 17,
397-434.
Dunbar, K. (1995). How scientists really reason: Scientific reasoning in real-world
laboratories. In R. J. Sternberg & J. Davidson (Eds.), Mechanisms of insight (pp.
365-395). Cambridge, MA: MIT Press.
Dunbar, K. (1996). Online analogical reasoning in scientific laboratories. Manuscript
submitted for publication.
Dunbar, K. (in press). Problem solving. In W. Bechtel & G. Graham (Eds.), A com-
panion to cognitive science. Cambridge, MA: Blackwell.
Dunbar, K., Patel, V., Baker, L., & Dama, M. (1995, November). Group reasoning
strategies in knowledge rich domains. Paper presented at the 36th annual meet-
ing of the Psychonomic Society, Los Angeles, CA.
Dunbar, K., & Schunn, C. D. (1990). The temporal nature of scientific discovery:
The roles of priming and analogy. Proceedings of the 12th Annual Meeting of the
Cognitive Science Society (pp. 90-100). Hillsdale, NJ: Erlbaum.
Ericsson, Κ. Α., & Simon H. A. (1993). Protocol analysis: Verbal reports as data (rev.
ed.). Cambridge, MA: MIT Press.
Forbus, K., Centner, D., & Law, K. (1995). MAC/FAC: A model of similarity-based
retrieval. Cognitive Science, 14, 144-206.

491
 KEVIN DUNBAR

Gick, M. L., & Holyoak, K. J. (1983). Schema induction and analogical transfer. Cog-
nitive Psychology, 14,1-38.
Holyoak, K. J. (1995). Problem solving. In E. Smith & D. Osherson (Eds.), An invi-
tation to cognitive science (2nd ed., pp. 267-296). Cambridge, MA: MIT Press.
Holyoak, K. J., & Thagard, P. (1989). Analogical mapping by constraint satisfaction.
Cognitive Science, 13, 295-355.
Holyoak, K. J., & Thagard, P. (1994). Mental leaps. Cambridge, MA: MIT Press.
Klahr, D. (1994). Searching for cognition in cognitive models of science. Psycoloquy,
5,68.
Klahr, D., & Dunbar, K. (1988). Dual space search during scientific reasoning. Cog-
nitive Science, 12, 1-48.
Klayman, J., & Ha, Y. (1987). Confirmation, disconfirmation, and information in hy-
pothesis testing. Psychological Review, 94, 211-228.
Koestler, A. (1964). The act of creation. London: Macmillian.
Kulkarni, D., & Simon, H. A. (1988). The processes of scientific discovery: The strat-
egy of experimentation. Cognitive Science, 12, 139-176.
Lovett, M. C., & Anderson, J. R. (in press). History of success. Cognitive Psychology.
Nersessian, N. (1992). How do scientists think? Capturing the dynamics of concep-
tual change in science. In R. N. Giere (Ed.), Minnesota studies in the philosophy
of science. Vol. XV: Cognitive models of science (pp. 3-44). Minneapolis: Univer-
sity of Minnesota Press.
Patel, V., Dunbar, K., & Kaufman, D. (1995, November). Distributed reasoning in
medical and scientific contexts. Annual Meeting of the Psychonomic Society, Los
Angeles, CA.
Rhodes, R. (1986). The making of the atomic bomb. New York: Simon & Schuster.
Sanderson, P. M., Scott, J. J. P., Johnston, T., Mainzer, J., Watanabe, L. M., & James,
J. M. (1994). MacSHAPA and the enterprise of exploratory sequential data
analysis (ESDA). International Journal of Human Computer Studies, 41,633-681.
Schunn, C. D., 8c Dunbar, K. (1996). Priming and awareness in complex reasoning.
Memory & Cognition, 24, 271-284.
Thagard, P. (1992). Conceptual revolutions. Cambridge, MA: MIT Press.
Tweney, R. D. (1989). Five questions for computationalists. In J. Shrager 8c P. Lang-
ley (Eds.), Computational models of scientific discovery and theory formation (pp.
471-484). San Mateo, CA: Morgan Kaufmann.
Tweney, R. D., Doherty, M. E., & Mynatt, C. R. (Eds.). (1982). On scientific thinking.
New York: Columbia University Press.

492
 HOW SCIENTISTS THINK

Wason, P. C. (1960). On the failure to eliminate hypotheses in a conceptual task.

Quarterly Journal of Experimental Psychology, 12, 129-140.
Wiser, M, & Carey, S. (1983). When heat and temperature were on. In D. Centner
& A. L. Stevens (Eds.), Mental models (pp. 267-297). Hillsdale, NJ: LEA.
Wotiz, J. H., & Rudofsky, S. (1984, August). Kekule's dreams: Fact or fiction? Chem-
istry in Britain, pp. 720-723.

493