1 Introduction and research question
“This sentence was written by an AI.” Well, the AI did not literally write it by itself and on its own intention, but it comes from the AI translation tool DeepL. DeepL is a German company that “has set itself the goal of eliminating language barriers worldwide by using artificial intelligence” (DeepL n.d.). In this sense, AI has permeated our daily life: not only DeepL but also Amazon’s virtual assistant Alexa, Apple’s Siri, and the currently much-discussed ChatGPT are popular and pervasive examples of AI technology. While companies are not fussy in labeling their newest products as AI, it is still unclear—not to say a huge mess—what exactly AI is. This may be a problem or not—I will say something about this next—but it is not my aim to target this (possible) issue. Instead, from a sociological point of view, there is an interest in how things are classified; this goes back to Émile Durkheim (2011 [1903]: 4), who wrote in collaboration with Marcel Mauss about class grouping:
We may well perceive, more or less vaguely, their resemblances. But the simple fact of these resemblances is not enough to explain how we are led to group things which, thus, resemble each other, to bring them together in a sort of ideal sphere, enclosed by definite limits, which we call a class, a species, etc.
Grouping, thus, includes defining limits. This happens, for example, by claiming a definition, and this is where this paper ties in: by investigating definitions and analyzing the setting of limits. According to this, the question addressed here is not: What is AI? —but rather: How does the class of AI come about? Before I explicate my research question in more detail, I will share some thoughts on why this—especially in the field of AI—is of relevance.
Advertisement
As mentioned above, not having a clear definition is not a problem in general, neither within the field of AI nor in the broader public. But as Pei Wang (2019: 2) outlines, there are different stakeholders and institutions that are interested in a clear definition: actors from within the economic field, since AI is a huge business opportunity and challenge where money is at stake,1 as well as policymakers from politics and law.2 Therefore, it is not surprising that defining AI is high on the agenda; countless episodes of various formats like TED Talks (e.g., Anderson and Thrun 2017), articles (e.g., Legg and Hutter 2007), or podcasts (e.g., Ben Byford 2021) in the past few years have had the sole aim of offering a sound definition of AI. Defining AI has become a trendy social practice.3 In this paper, I would like to shed some light on this.
By investigating a bundle of definitions of AI, I aimed to identify different dimensions of criteria that are crucial for AI. In other words, I wanted to find out which features are decisive for a computer or a machine to be called intelligent. Therefore, my first research question is as follows:
1.
What criteria do definitions of AI use to define AI?
Out of this, two further questions are addressed:
3.
What correlations can be drawn between the found criteria and further parameters?
As it should be clear from the setting of these questions, my aim is not to place a normative argument, like rating definitions or criteria, and it is certainly not the goal to offer my own definition; rather, my analysis follows a descriptive approach. It should also be noted that a lot of terms dealt with herein are—like the term AI itself—conceptually and terminologically challenging. I am aware that the identified dimensions and categories are ambiguous in the way they are interpreted. But since they are all collected from the given material, I take the liberty of adopting the categories in their blurriness and it is not my point to clarify them in this paper.
Advertisement
2 Data and methods
My empirical research consisted of a content analysis of a sample of various text sources. I already intimated that defining AI seems to become more the center of attention the last few years although the term entered the scientific stage back in the 1950s, when John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon proposed the groundbreaking workshop at the Dartmouth College (see Russell and Norvig 2010: 17). To examine if and how definitions of AI have changed over time, the collection of definitions covers, at best, the time since the term was first officially used, dating back to the 1956 proposal paper of the aforementioned Dartmouth Summer Research Project on Artificial Intelligence. Since naturally most scientific journals in the field of AI were established later—for example, Elsevier’s journal Artificial Intelligence started in 1970—a media-based selection was not expedient. A more helpful way here was the search engine Google Scholar, which lists the first hit of the keywords “definition of artificial intelligence” in 1962. Following this search term, Google Scholar offers a total of 1262 text sources (status 7 July 2021), including books, journal articles, dissertations, reports, conference papers, and others.
To have an informative and broad historical sample, I divided the sample into three different time periods and drew a random sample from each of these. The first period lasts from 1962 to 1985, the second from 1986 to 2005, and the third from 2006 to 2021. Since publications about AI increased massively over time and to counteract this imbalance, the earlier periods are set longer. Nevertheless, as expected, the third period contained an exponentially larger number of results compared to the first and the second period (see Table 1). Therefore, these samples constitute a different fraction of the overall number of empirical pieces found in Google Scholar. It is consequently important to note that “inferences from the samples’ estimates to the populations’ parameters have different degrees of confidence” (Abend 2006).
Table 1
Characteristics of the sample
Time Period | Total empirical articlesa | Sample size | Proportion in % | Substitutions | Proportion incl. Substitutions % |
|---|---|---|---|---|---|
1962–1985 | 46 | 15 | 32.61 | 21 | 78.26 |
1986–2005 | 246 | 15 | 6.10 | 29 | 17.89 |
2006–2021 | 970 | 15 | 1.55 | 5 | 2.06 |
For collecting my sample, I used the method of substitution since, especially in the first and the second period, a lot of sources were not useful, either because access to the document was precluded or because they did not contain a definition. Therefore, the total number of sampled sources was 100, from which 45 now form the analyzed sample. The characteristics of my sample are summarized in Table 1.
The large number of substitutions could be an indication that the search term was chosen too broadly. But as substitutions were a lot less in the last period, it is obvious that (1) texts about AI today shed more light on making clear definitions and (2) access to today’s papers is easier.
How representative this data is for research in AI in general cannot be estimated. But as other studies (e.g., Malouin 2020: 65 f.) have shown, the number of published papers related to AI is increasing significantly since 2000. This growth reflects an uptick in publicity and funding but is also seen critically by some academics.5 Besides this trend toward more papers after 2000, the data show a significant peak in 1982–1985: While for a time range of 4 years the sample shows an overall average of 3.27 texts, the period between 1982 and 1985 contains 8 sources. Although especially in 1984 AI seemed to be high on the research agenda, it is probably not George Orwell’s merit, but attributed to the rise of expert systems introduced by Edward Feigenbaum and some offensive funding programs, like the one from the Japanese government (see Anyoha 2017). Regarding the random sampling, it is important to emphasize that the frequency of a definition or criterion should not be taken as a benchmark of its validity, since the accuracy of a definition is hardly statistical by nature. However, for my research interest, which revolves only around the frequency of used criteria within definitions of AI, the statistical approach is suitable.
To analyze the collected text sources, I followed the procedural model of Mayring (2014: 54). After defining the units of analysis, which led to some substitutions because of missing or unclear definitions, I started coding and creating a category system inductively out of the material. Thereby, theoretical thoughts and considerations about my research questions were continuously incorporated. After coding around 50% of my sample, I reevaluated the categories and redesigned them with regard to the analysis purpose. Some sources contained more than only one definition. I then tried to figure out which one complies best with the authors’ view or which one the authors used for their purpose and coded only these. Thus, the definitions are not always created by the authors themselves but could be cited. A list of all coded definitions can be found in the appendix.
In advance, it must be noted that the content analysis carried out required some interpretative judgments, especially the dimensions regarding the criteria of the definitions are methodologically challenging. However, they potentially also offer more significant insights and, as Abend et al. (2013: 610) stated: “it does not follow from the need for human judgment that the instrument cannot be reliable, let alone that any interpretation of meaning or coding judgment is as good as any other.” The interpretative work varies depending on the variable. In particular the metadata, e.g., year of publication or native country of authors,6 are variables where no or only a few interpretations of meaning are involved.
For reasons outlined in the introduction, the goal of the analysis was structuring the material, after which an explorative and a relational design was followed. This means, the aim is (1) to offer new insights into the criteria used to define AI and (2) to show correlations or independencies of different parameters on both content-related and context-related variables. In sum, including the metadata, the coding system included around 20 distinct variables whereof 12 are now considered for this analysis.
To set the variables in dependencies, for the dimensions human likeness, state of “mind”, and successfulness, I put the expressions found in ordinal levels, which allows for statistical tests as well. For this, I had to arguably decide which category is greater or higher. I did this in a general manner with the question “what is harder to reach?”.7 Creating this order is based on theoretical thoughts, required interpretative work on my part as well and was, therefore, challenging. Although I made the categorization and the division of the subcategories inductively from the material, they are not able to represent all characteristics and all properties, and are therefore simplified. Nevertheless, for my analytical purposes, this simplification and categorization were needed and reasonable. If in the following sections, specifications about statistical correlations are given, expressions are coded in the order represented in Table 2. This means, e.g., for the variable human likeness: “Is not a criterion” = 0, “In terms of behavior” = 1, “In terms of success” = 2, “In terms of state” = 3.
Table 2
Criteria used for definitions of AI
1962–1985 (%) | 1986–2005 (%) | 2006–2021 (%) | Total (%) | |
|---|---|---|---|---|
Learning ability | ||||
Is not a criterion | 66.7 | 86.7 | 53.3 | 68.9 |
Is a criterion | 33.3 | 13.3 | 46.7 | 31.1 |
Human likeness | ||||
Is not a criterion | 46.7 | 20.0 | 20.0 | 28.9 |
In terms of behavior | 26.7 | 40.0 | 13.3 | 26.7 |
In terms of success | 13.3 | 26.7 | 20.0 | 20.0 |
In terms of state | 13.3 | 13.3 | 46.7 | 24.4 |
State of “mind” | ||||
Is not a criterion | 46.7 | 60.0 | 46.7 | 51.1 |
In terms of logical/rational operations | 6.7 | 0.0 | 20.0 | 8.9 |
In terms of reasoning | 13.3 | 6.7 | 0.0 | 6.7 |
In terms of knowledge | 20.0 | 6.7 | 0.0 | 8.9 |
In terms of understanding/having meaning | 13.3 | 0.0 | 13.3 | 8.9 |
In terms of mental/cognitive capacities | 0.0 | 26.7 | 20.0 | 15.6 |
Complexity of the problem | ||||
Is not a criterion | 86.7 | 93.3 | 80.0 | 86.7 |
Is a criterion | 13.3 | 6.7 | 20.0 | 13.3 |
Successfulness | ||||
Is not a criterion | 53.3 | 66.7 | 40.0 | 53.3 |
In terms of goal-orientedness | 6.7 | 6.7 | 0.0 | 4.4 |
In terms of satisfying problem solving | 13.3 | 13.3 | 46.7 | 24.4 |
In terms of efficiency | 6.7 | 13.3 | 6.7 | 8.9 |
In terms of the optimal/best solution | 20.0 | 0.0 | 6.7 | 8.9 |
Authors take definition as objective/subjective | ||||
AI definition is something that is objectively | 60.0 | 26.7 | 40.0 | 42.2 |
AI definition is something that is subjectively | 20.0 | 40.0 | 26.7 | 28.9 |
Unclear | 20.0 | 33.3 | 33.3 | 28.9 |
Supports or follows Tesler’s theorem | ||||
No | 86.7 | 66.7 | 86.7 | 80.0 |
Yes | 13.3 | 33.3 | 13.3 | 20.0 |
Total number of articles | 15 | 15 | 15 | 45 |
3 Findings 1: Five criteria used for defining AI
The first findings address research question one and offer a list of criteria that are used for definitions of AI. Out of my material, I identified five main dimensions that—so my argument goes—are central to the dissent over the question “What is AI?”: (1) learning ability, (2) human likeness, (3) state of “mind”, (4) complexity of the problem, and (5) successfulness. As the data show, the dissent is not only about which of these criteria are distinctive for AI, but also about the degree to which a criterion must be fulfilled.
In addition, I coded two further variables regarding the definitions: If the authors claim an objective definition of AI or if they see the definition as something subjective, and if the given definition supports or follows what is known as Tesler’s theorem8: “Intelligence is whatever machines haven't done yet” (Tesler n.d.). The categories and their frequencies found are, grouped by time period, summarized in Table 2. With that, I turn to the individual dimensions and have a look at their development over time.
3.1 Learning ability
Learning ability9 as a criterion for AI is of particular interest since today machine learning is often used as a synonym for AI (see Anderson 2017). As the data show, this criterion became used more frequently in the youngest time period than in the previous two periods. Furthermore, if only publications after 1985 are included, a Pearson’s Chi-square test shows a significant positive correlation between the year of publication and the use of this criterion (r = 0.43, p = 0.017), while over the entire period, the correlation is less striking (r = 0.14, p = 0.369). Even more noticeable is the data in the time 2006–2021: Six of the seven definitions that contained learning ability as criterion were published in 2018 or later. Since the definitions are highly divided in using this parameter or not, I argue that—especially in the newer time—learning ability can be identified as one of the critical dimensions in the dissent over definitions of AI.
3.2 Human likeness
Resemblance to human beings is probably what makes AI so fascinating, not only in science but also in pop culture. As we can see from the data, it is the most used criterion for definitions of AI, although the similarity is applied at different levels. What is especially striking here is that this variable, just like learning ability, shifts significantly over time; while in the first period, almost half of the definitions did not use this criterion, it seems it became more important later. In addition, we can see that the level that has to be reached on this parameter increased as well. A Spearman’s test here shows a significant positive correlation between the year of publication and the use, respectively, the pursued level, of this criterion (ρ = 0.46, p = 0.001). These findings support the argument that what AI is, or better: what counts as AI, is a question of history. However, although we can see some consensus here since most of the definitions use this parameter, there is still dissent over the required level, respectively, over how AI needs to be humanlike. As the most used criterion in my data and as important constituent in the common conception of AI, it could be argued that any definition of AI is partly based on human likeness. If this is true, my analysis must be understood as an inquiry into whether human likeness is used implicitly or explicitly, and in relation to which aspects.
3.3 State of “mind”
The historical development mentioned above can also be found in this dimension, even if it is less obvious. For example, it can be seen that the category “knowledge” particularly occurs around the year 1980, which falls together with the rise of knowledge and expert systems (see Russell and Norvig 2010: 23 f.). Since this parameter shows a large variation—both in referring to and not using it at all—it is arguable that there is no consensus. What is more is that this category, apart from the raised distributions in this study, is probably highly challenging in substantive senses, which makes consensus even less likely. But this is another story.10
3.4 The complexity of the problem
The complexity of the problem is the least controversial dimension, in the sense that most definitions do not consider it crucial.11 This is true for all time periods. With this said, one could claim that there is nothing of interest in this dimension. But I would like to argue why this result is still noteworthy: First, to Marvin Minsky, who is considered one of the founders of the term AI and had—and still has—a large influence on the research of AI,12 this was the only criterion when he defined (artificial) intelligence solely as “the ability to solve hard problems” (Minsky 1985). Second, the hope that AI could solve hitherto unsolved problems often resonates in public discourse, which implies that AI must be able to solve difficult tasks. However, since other dimensions offer more controversial aspects, I turn to the next one.
3.5 Successfulness
The dimension successfulness shows similar distribution characteristics as state of “mind”, which means it is also controversial in use and levels. A general direction cannot be seen, but what is striking is the frequency it has been used since 2006 in terms of “satisfying problem solving”. An interpretation for this could be that there is some disillusion about AI as the best solution for problems, as well as the general expectation that AI can offer a pragmatic solution for our real problems, e.g., climate change, as it is often claimed (see Marr 2021). Anyway, even if the cruciality of this criterion is controversial, it can be noted that at least in the younger period, there is some consensus about the required level.
3.6 Author claims for objectivity
To measure the setting in which the definitions are made, I coded whether the authors claim their definition (or a cited definition) as objectively true or something subjective. This offers some insights into the practice of defining AI since the claim made could be essential for this. Regarding the time, we cannot see significant differences in this variable, but I will come back to this later.
4 Findings 2: The emergence space of definitions
With this more general and historical analysis made, I turn to my third research question: What correlations can be drawn between the found criteria and further parameters? Or to put it another way: Is defining AI and the dissent about it something that can be explained “culturally”? To answer this question, I collected parameters from the material gathered that represent social or “cultural” aspects. These metadata are summarized in Table 3. Subsequently, I compared these parameters with the criteria found to prove if there are some correlations. Before I come to these findings, some comments and limitations must be addressed. First, the term “cultural”—highly disputed in general (see Romano 2002)—is probably not quite appropriate here since my variables are not sufficient to catch culture, nor do all the variables analyzed here represent culture. But I use culture because it is suitable in the sense that it refers to a larger context: the emergence space of the definitions. I could have used external or social factors or something similar, but the word factor is problematic in view of my second caveat: It is not my aim to make causal claims about relationships. Although my findings possibly offer some indications for interdependencies, arguing for causality would need further research as well as a more theoretical explanation.
Table 3
Collected meta data
1962–1985 (%) | 1986–2005 (%) | 2006–2021 (%) | Total (%) | |
|---|---|---|---|---|
Year of publication | ||||
1962–1985 | 100.0 | 0.0 | 0.0 | 33.3 |
1986–2005 | 0.0 | 100.0 | 0.0 | 33.3 |
2006–2021 | 0.0 | 0.0 | 100.0 | 33.3 |
Type of publication | ||||
Book | 0.0 | 20.0 | 0.0 | 6.7 |
Book section | 0.0 | 0.0 | 13.3 | 4.4 |
Conference paper | 13.3 | 0.0 | 6.7 | 6.7 |
Journal article | 13.3 | 20.0 | 46.7 | 26.7 |
Report | 33.3 | 6.7 | 6.7 | 15.6 |
Thesis Ph.D. | 26.7 | 20.0 | 6.7 | 17.8 |
Thesis non-Ph.D. | 13.3 | 33.3 | 20.0 | 22.2 |
Country | ||||
USA | 93.3 | 60.0 | 20.0 | 57.8 |
Not USA | 6.7 | 40.0 | 80.0 | 42.2 |
Gender of the author(s) | ||||
Unknown | 0.0 | 6.7 | 6.7 | 4.4 |
Male | 80.0 | 73.3 | 53.3 | 68.9 |
Female | 20.0 | 20.0 | 13.3 | 17.8 |
Both | 0.0 | 0.0 | 26.7 | 8.9 |
Authors’ field of expertise | ||||
Unknown | 20.0 | 13.3 | 0.0 | 11.1 |
Computer science and informatics | 53.3 | 26.7 | 26.7 | 35.6 |
Philosophy and social science | 13.3 | 33.3 | 6.7 | 17.8 |
Economics and business administration | 6.7 | 0.0 | 60.0 | 22.2 |
Physics and nature science | 0.0 | 6.7 | 6.7 | 4.4 |
Others (e.g., musicology, theology) | 6.7 | 20.0 | 0.0 | 8.9 |
Total number of articles | 15 | 15 | 15 | 45 |
4.1 The medium
The first cultural parameter I consider is the type of publication. In my sample, I found seven different types of mediums: books, book sections, conference papers, journal articles, reports, and theses (Ph.D. and lower than Ph.D.). Cross tabled with the substantive variable human likeness, it can be shown that definitions from reports or conference papers refer much less to human likeness: Only two out of ten text sources from these mediums use this criterion, while over all mediums it is used in up to 71.1% (see Table 4). For an interpretation of this distribution, one must be aware that there is presumably also an interaction with the year of publication, because the types of publication are not distributed equally over time. Nevertheless, compared to other variables this is very striking, which could mean that the question of whether human likeness is crucial for a definition of AI varies with the medium, and thus with the target audience.
Table 4
Cross table: Type of publication
Human likeness is criterion1 | Authors take definition as2 | Supports or follows Tesler’s theorem3 | ||||
|---|---|---|---|---|---|---|
No | Yes | Objective | Subjective | No | Yes | |
(%) | (%) | (%) | (%) | (%) | (%) | |
Book | 33.3% | 66.7 | 100.0 | 0.0 | 100.0 | 0.0 |
Book section | 0.0 | 100.0 | 100.0 | 0.0 | 100.0 | 0.0 |
Conference paper | 100.0 | 0.0 | 100.0 | 0.0 | 100.0 | 0.0 |
Journal article | 16.7 | 83.3 | 22.2 | 77.8 | 66.7 | 33.3 |
Report | 71.4 | 28.6 | 50.0 | 50.0 | 100.0 | 0.0 |
Thesis Ph.D. | 25.0 | 75.0 | 100.0 | 0.0 | 100.0 | 0.0 |
Thesis non-Ph.D. | 0.0 | 100.0 | 50.0 | 50.0 | 50.0 | 50.0 |
Total | 28.9 | 71.1 | 59.4 | 40.6 | 80.0 | 20.0 |
The same is true for whether an author claims a definition objectively true or not: While in journal articles claims for subjectivity of definitions were found more frequently, definitions in Ph.D. theses were most claimed to be objectively true (see also Table 4). A noticeable variance, although not statistically significant, can also be found in whether a definition supports or follows Tesler’s theorem. Only in journal articles and non-Ph.D. theses, this criterion could be found.
The findings presented here, I argue, are a strong indication that definitions of AI are not independent of the medium they are published in—neither regarding the used criteria nor the claims the authors make with the definition. Here, a theoretical link to Ludwik Fleck (2017 [1935]: 146 f.) can be made, who arranges different scientific activities and their results upon its medium: Thus he speaks of “Zeitschriftenwissenschaften” (journal sciences), “Handbuchwissenschaften” (handbook sciences), “Lehrbuchwissenschaften” (textbook sciences), and “populäre Wissenschaften” (popular sciences).
4.2 Field of expertise
While the type of publication shows some significant correlations, the field the author comes from does not seem to be crucial for the criteria used for the definitions of AI (see Table 5). Although we can see some trends over time—definitions today are more often offered from the field of economics (see Table 3)—substantively no significant correlation could be found. This indicates that the dissent over the definition of AI is not a dissent between different research fields. What is most striking here is that comparatively many definitions that support or follow Tesler’s theorem do not come from the main classical fields of AI research.
Table 5
Cross table: Field of expertise
Human likeness is criterion1 | State of “mind” is criterion2 | Supports or follows Tesler’s theorem3 | ||||
|---|---|---|---|---|---|---|
No | Yes | No | Yes | No | Yes | |
(%) | (%) | (%) | (%) | (%) | (%) | |
Computer science and informatics | 37.5 | 62.5 | 43.8 | 56.3 | 81.3 | 18.8 |
Philosophy and social science | 25.0 | 75.0 | 25.0 | 75.0 | 87.5 | 12.5 |
Economics and business administration | 10.0 | 90.0 | 60.0 | 40.0 | 90.0 | 10.0 |
Physics and nature science | 50.0 | 50.0 | 100.0 | 0.0 | 50.0 | 50.0 |
Others | 25.0 | 75.0 | 75.0 | 25.0 | 25.0 | 75.0 |
Total | 27.5 | 72.5 | 50.0 | 50.0 | 77.5 | 22.5 |
4.3 Gender of the author
In the use of the criterion human likeness, there is no difference observable between men and women (Fisher’s test: p = 0.665). However, if we analyze on which level men and women apply this parameter, a different picture is observable: Table 6 shows that male authors use this parameter more often in terms of “behavior or state”, while females focus on “success”. This is evidence for the argument that men and women have different views of what AI is; not in the criteria, but in the levels that need to be reached. Needless to mention: this is a simplified conclusion and to make a general claim about gender differences in the practice of defining AI, deeper research is required.
Table 6
Cross table: Gender of author
Human likeness is criterion1 | Authors take definition as2 | |||||
|---|---|---|---|---|---|---|
No | In terms of … | Objective | Subjective | |||
Behavior | Success | State | ||||
(%) | (%) | (%) | (%) | (%) | (%) | |
Female | 37.5 | 12.5 | 50.0 | 0.0 | 85.7 | 14.3 |
Male or both | 25.7 | 28.6 | 14.3 | 31.4 | 52.0 | 48.0 |
Total | 27.9 | 25.6 | 20.9 | 25.6 | 59.4 | 40.6 |
4.4 Country
What could be suggested regarding the dissent about defining AI is that, in different places in the world, definitions vary. But if we have a look at the corresponding variable (see Table 7), we see that the country the authors come from is statistically independent of all the substantive dimensions. It must be noted here that only text sources in English were taken into account, which is seen as a constraint for this argument.
Table 7
Cross table: Country
Learning ability is criterion1 | Human likeness is criterion2 | State of “mind” is criterion3 | Success is criterion4 | |||||
|---|---|---|---|---|---|---|---|---|
No | Yes | No | Yes | No | Yes | No | Yes | |
(%) | (%) | (%) | (%) | (%) | (%) | (%) | (%) | |
Not USA | 73.7 | 26.3 | 26.3 | 73.7 | 52.6 | 47.4 | 57.9 | 42.1 |
USA | 65.4 | 34.6 | 30.8 | 69.2 | 50.0 | 50.0 | 50.0 | 50.0 |
Total | 68.9 | 31.1 | 28.9 | 71.1 | 51.1 | 48.9 | 53.3 | 46.7 |
5 Conclusion
The main finding of this paper is that the dissent on the question of “What is AI?” can be broken down into five different parameters, namely learning ability, human likeness, state of “mind”, successfulness, and complexity of the problem. Thereby, the latter is not as controversial as the former ones. While since the implementation of the concept of AI in the 1950s, the use of some of these parameters shifted over time, the discourse on definitions of AI, in general, seems to revolve stably around these dimensions. But as the data have shown, there is no overarching consensus; neither on which of these criteria are crucial to mark out AI nor on how these criteria must be fulfilled. However, by focusing closer on some aspects, it can be said that there is some general agreement in the use of two of these parameters: First, the complexity of the problem for most of the definitions is not decisive, and second, in the era since the 1990s, human likeness is accepted widely as a key attribute for AI—although on different levels.
By opposing the found frequencies to the collected metadata, which present some “cultural” aspects from the emergence space of the text sources, it is seen that only a few of them have significant correlations. The data do not show any relation between the content characteristics and the country or the gender of the author or the scientific field in which the text source was written. On the other hand, some substantive variables are not statistically independent of the medium the definition was published in. Since different mediums target different purposes and different audiences, we can assume that writing a definition of AI must be seen in the context of its distribution area and its goal. This finding may correspond with the often-made difference between a working definition and a dictionary definition (see Wang 2019: 3), although I believe that the boundaries of these categories must be considered fluid.
The subsequent question of what implications the findings of this paper have—for the research of AI itself or the social practice of defining AI—is left open. Likewise, no causal relationships could be claimed upon the correlations found here since this requires deeper and wider research as well as theoretical linkages. So far, the results must be viewed with prudence and caution. The data show convincingly that defining Artificial Intelligence, as a form of classifying, is a social practice where different external, cultural, and historical influences come into play and must be taken into account.
I aimed to point out that in the ongoing debate about AI, we should not only focus on the definitions themselves but also pay attention to the context of their social reality. By contrasting various content criteria with various cultural dimensions, this paper attempts to contribute to that end and hopefully offers starting points for further research.
Finally, I do not know what Alexa answers if one asks “her” why “she” is labeled intelligent. One suggestion could be something like: “Because I can listen and talk like a human being.”13 This is—according to the findings presented here—only a partly good answer since it could be seen that the trend regarding the human likeness points more and more toward human-like states (e.g., of consciousness). Speaking, which is a manner of behaving like a human, seems not sufficient anymore to be called intelligent. Sorry, Alexa.
Declarations
Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest.
Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.