by Leah Henrickson
1. Introduction
Natural language generation (NLG) is the process wherein computers produce text-based output in readable human languages. NLG systems are increasingly prevalent in our modern digital climate, prompting the emergence of companies that specialise in generating output for mass readerships and readerships-of-one alike. Narrative Science has worked in partnership with Deloitte to generate client-friendly narrative reports related to such issues as budget optimisation, financial operations, and internal auditing (Krittman, Matthews, and Glascott 2015). Automated Insights has partnered with Bodybuilding.com to produce app-based workout recaps personalised to each individual user to motivate the user to maintain an exercise routine (Figure 1) (Case Studies: Bodybuilding.com). These are only a few examples of NLG’s current applications as implemented by a burgeoning industry based around this technology. There are also, however, more aesthetic endeavours. One early example is The Policeman’s Beard is Half Constructed, marketed as ‘the first book ever written by a computer’, which was published by Warner Books in 1984 (Racter 1984). The Policeman’s Beard comprises poems, prose, and dialogue generated by an NLG system called Racter; each block of text is accompanied by a collage by artist Joan Hall. More recently, hundreds of digital books have been produced as part of the annual National Novel Generation Month. Additionally, Twitter has seen a proliferation of bots that regularly generate tweets according to pre-programmed constraints. One example is the Magic Realism Bot, which proposes a hilariously absurd fantasy story plot every four hours. In this paper, however, I focus primarily on those texts generated for more pragmatic purposes: namely, news articles.
Figure 1: An example of a weekly exercise summary generated by Automated Insights (2017) for Bodybuilding.com
Despite the burgeoning industry centred on NLG, there has not yet been any systematic analysis of computer-generated text reception. Resultantly, we do not know where these texts fit within our current conceptions of authorship and readership. Any act of reading engages interpretive faculties; modern readers tend to assume that a text is an effort to communicate a particular pre-determined message. With this assumption, readers assign authorial intention, and hence develop a perceived contract with the author. I refer to this author-reader contract as ‘the hermeneutic contract’. This is the hermeneutic contract: that reading is accepted as an act wherein the reader receives an interesting and understandable text written by an author motivated by intention-directed agency (Henrickson 2018). The author is regarded as an individual creative genius.
Computer-generated texts bring the hermeneutic contract into question. The hermeneutic contract’s communication principle rests on two assumptions: that readers believe that authors want them to be interested in their texts, and that authors want them to understand their texts. Yet the author of a computer-generated text is often an obscured figure, an uncertain entanglement of human and computer. As yet, the discussion about attributing authorship to computer-generated texts has been limited to theoretical consideration by scholars who tend to argue that computer-generated texts are authorless in the conventional sense of the word. This paper broadens the discussion to include the opinions of ordinary readers – that is, readers who have not extensively considered issues related to NLG. These opinions have been collected through an online reader response questionnaire comprising 500 adult participants. Contrary to the scholarly assertions that computer-generated texts are authorless, these questionnaire results indicate that readers do attribute authorship – and, by extension, perceived agency – to NLG systems.
Note that this work is part of a larger ongoing research project that investigates the social and literary implications of NLG, and the results presented herein are neither comprehensive nor conclusive. Also note that this analysis has been abridged given the time constraints of a conference paper.
2. Questionnaire Demographics and Design
The online reader response questionnaire described in this paper was conducted from late 2017 until early 2018. Of the 500 total participants, 337 self-identified as women, 155 as men, 5 as non-binary, and 3 as other. 139 were aged 18-29, 140 aged 30-39, 76 aged 40-49, 68 aged 50-59, 55 aged 60-69, and 22 aged 70+. 27 countries of residence were represented, with the United States providing a substantial number of responses (213), and the United Kingdom (122) and Canada (90) also providing great numbers. A detailed distribution of country representation is shown in Table 1. 32 noted their highest level of education completed being secondary school, 10 trade/technical/vocational training, 126 an undergraduate programme, 180 a master’s programme, and 152 a doctoral programme. 120 participants were students, 320 employed, 47 retired, and 13 not employed/unpaid workers. More than half of participants identified as working within education, training, and library fields. A detailed distribution of occupational fields represented is shown in Table 2.
Table 1: A detailed distribution of questionnaire participants’ countries of residence
| Country | Number of Participants |
| United States of America | 213 |
| United Kingdom | 122 |
| Canada | 90 |
| Finland | 17 |
| Australia | 9 |
| Germany | 9 |
| France | 7 |
| Sweden | 4 |
| India | 3 |
| Italy | 3 |
| Egypt | 2 |
| Ireland | 2 |
| Israel | 2 |
| Netherlands | 2 |
| Spain | 2 |
| Switzerland | 2 |
| Belgium | 1 |
| Cyprus | 1 |
| Denmark | 1 |
| Hungary | 1 |
| Malta | 1 |
| Mexico | 1 |
| Norway | 1 |
| Poland | 1 |
| Samoa | 1 |
| Singapore | 1 |
| South Africa | 1 |
Table 2: A detailed distribution of questionnaire participants’ occupational fields
| Occupational Field | Number of Participants |
| Education, Training, and Library | 252 |
| Information Technology | 38 |
| Business and Financial Operations | 27 |
| Engineering | 25 |
| Media and Communications | 25 |
| Arts, Design, and Entertainment | 23 |
| Life, Physical, and Social Science | 20 |
| Not Employed | 18 |
| Office and Administrative Support | 13 |
| Legal | 10 |
| Public Sector | 10 |
| Healthcare | 7 |
| Management | 7 |
| Journalism | 6 |
| Retail and Sales | 6 |
| Community/Social Services | 2 |
| Construction and Extraction | 2 |
| Hospitality | 2 |
| Personal Care and Service | 2 |
| Architecture | 1 |
| Building/Grounds Maintenance | 1 |
| Farming, Fishing, and Forestry | 1 |
| Production | 1 |
| Sport | 1 |
Social media platforms such as Reddit, Facebook, and Twitter were used to promote the questionnaire. Additionally, participants were solicited through the SHARP (Society for the History of Authorship, Reading and Publishing), EX-LIBRIS, and DHSI (Digital Humanities Summer Institute) professional listservs. Invitees were encouraged to share the questionnaire in an effort to reach as wide a demographic spread as possible. Nevertheless, these responses are dominated by a highly-educated populace of primarily women and those employed in education, training, and library roles. This populace undoubtedly emerged as a result of the methods of recruitment employed.
In the questionnaire, participants were presented with a 243-word English-language text reviewing Finnish municipal election results called ‘The Finns Party drop most seats across Finland’. This text was generated by an NLG system called Valtteri the Election Bot, developed by the Immersive Automation research team in Finland. The first paragraph of the text is as follows; subsequent paragraphs continue in similar fashion:
‘The Finns Party dropped the most council seats throughout Finland and lost 425 seats. The Finns got 3.5 percentage points fewer votes than in the last municipal election and decreased their voter support by the greatest margin. The party dropped 80501 votes since the last municipal election and has 770 seats in the new council.’1
Questionnaire participants were asked to attribute authorship to this text four times. Each time participants were asked to attribute authorship, they were presented with new information about the text’s process of production. In the first question, participants were only presented with the text itself, which had Valtteri’s name in the byline below the text’s title. In the second, participants were informed that Valtteri was a bot, and that one John Smith (a pseudonym for a member of the Immersive Automation team) had translated the Finnish political party names into English for ease of reading.2 In the third, participants were informed that Valtteri was developed by the Immersive Automation team; in the fourth, that Immersive Automation was funded by numerous public and private bodies. For each of these questions, participants had to select from lists of predefined authors. ‘It is not possible to assign authorship’ and ‘Other’ options were available for all questions. If a participant selected ‘Other’, the participant was required to clarify why. Each question included an optional ‘Why have you selected this option?’ text box for those participants generous enough to elaborate upon their answers. To streamline the analysis herein, only the responses to the fourth and final authorship attribution question – when participants were given all possible options to select – are given serious consideration for quantitative purposes. The other questions, however, did garner valuable qualitative responses that have informed this analysis.
The questionnaire began, though, by asking participants to ‘list three things that come to mind when you think of the word “author”.’ This opening question was included to prompt participants to think about what authorship meant to them, prior to being asked to assign authorship to the text provided. While there have been studies to consider how the increasing prominence of artificial intelligence technologies may lead to altered conceptions of authorship as they pertain to copyright and intellectual property law, none have solicited the opinions of everyday readers (Sorjamaa 2016). The word cloud shown in Figure 2 visualises all 1,500 items listed, with identical listings grouped. The size of each word is relative to the frequency with which it was listed. Variations on similar words have not been aggregated.
Figure 2: A word cloud visualising 1,500 items listed by participants when asked to ‘list three things that come to mind when you think of the word “author”’
This list of word associations was scrutinised with the understanding of the word ‘author’ as both a noun and a verb; an author is a thing of some kind, and/or an activity of some kind. However, as the word association list was coded, four general facets of ‘author’ emerged that nuanced this understanding: ‘author’ refers to (1) an identity that is (2) associated with particular connotations, (3) as well as with particular activities (4) that result in particular kinds of (generally text-based) outputs. It is not the place of this paper to extensively review the results of this word association exercise. For the purposes of this discussion, it is enough to simply say that ‘author’ is hardly as clear a term as one may initially suppose.
3. Attributing Authorship to Computer-Generated Texts
Few scholars have considered how to attribute authorship to computer-generated texts, but those who have tend towards the opinion that computers cannot be considered authors. In his seminal Gödel, Escher, Bach, cognitive scientist Douglas Hofstadter ponders how one may attribute authorship in instances of computer-generated aesthetic works. Although Hofstadter’s discussion centres on the generation of music in particular, his argument may apply to any other form regarded as a product of human creativity. For Hofstadter, computational composition is directed by human intellect, and he regards computers as tools for realising human intention. At Hofstadter’s time of writing, a composing computer could not have been considered a sentient being given the computer’s inflexibility in task execution and lack of unique perspective and self-awareness. Hofstadter (1980, p. 609) argues:
‘If and when, however, people develop programs which have those attributes [flexibility, perspective, and self-awareness], and pieces of music start issuing forth from them, then I suggest that will be the appropriate time to start splitting up one’s admiration: some to the programmer for creating such an amazing program, and some to the program itself for its sense of music. And it seems to me that that will only take place when the internal structure of such a program is based on something similar to the ‘symbols’ in our brains and their triggering patterns, which are responsible for the complex notion of meaning. The fact of having this kind of internal structure would endow the program with properties which would make us feel comfortable in identifying with it, to some extent. But until then, I will not feel comfortable in saying ‘this piece was composed by a computer’.’
More recently, literary scholar Martin Eve (2017, p. 48) has similarly argued that the NLG system’s sense of self ‘is an aggregate of human selves’:
‘Whether or not it has such a self-consciousness, though, is a different matter. However, the absolute history of computer writing rests upon this human writing and labour. Were the human race to die out but the machines to keep on writing, they would continue to produce ever more conservative texts, training themselves upon their own regurgitated outputs with only semi-deterministic random seeds to aid progress and foster change.’
For Eve, the computer-generated text is a manifestation of the many human labour forms that contribute to system output. Without humans to develop and maintain an NLG system, and to regularly provide it with new input to process, the system is rendered impotent. Indeed, Eve goes so far as to assert that ‘we continue to refer to computer poetry and literature as lacking an author’, presumably because readers regard authorship as necessitating the lived experience of a sentient and autonomous – human – entity (Eve 2017, p. 50). Although Eve asserts that authorship remains wholly unattributable in instances of computer-generated texts, other scholars who have recently considered this issue (Bootz and Funkhouser 2014, p. 84) have referred to developers of NLG systems as ‘author-programmers’, assigning some form of authorship related to computer-generated output, however vague, to system developers.
Despite their writing nearly forty years apart, Hofstadter and Eve reach similar conclusions regarding authorship attribution for computer-generated content: authorship cannot be attributed to a computer because the computer is merely a tool for manifesting human vision. This is the nearly-unanimous conclusion reached by scholars. In the rare instances when the word ‘author’ is used in discussions of computer-generated texts, it is qualified, as in ‘author-programmers’.
What, though, do ordinary readers think? Once one appreciates the diverse ways in which the word ‘author’ itself may be considered, the discussion can shift to how readers may attribute authorship to computer-generated texts. Just as the results of the aforementioned authorship association exercise garnered ambiguity, the results related to attribution revealed similarly diverse findings. While I had anticipated that participants would attribute authorship to Valtteri’s developers (if they felt they could attribute authorship at all), many instead attributed authorship to Valtteri itself, distinguishing the system as the author, and the developers as the creators – or even authors – of the system. Surprisingly, some participants identified the funding body of the system’s development team as an author of sorts, alluding to a patronage model of text production. 143 participants answered that it was not possible to assign authorship at all. The remaining 357 participants, however, felt that authorship could be attributed. Table 1 shows the distribution of authorship attribution responses.
Table 4: The distribution of questionnaire participants’ attributions of authorship to a computer-generated news article
| Who is/are the author(s) of this text? | Option | Number of Participants |
| It is not possible to assign authorship | 143 | |
| Valtteri | 179 | |
| John Smith | 7 | |
| Immersive Automation | 90 | |
| Those who fund Immersive Automation | 9 | |
| Other | 72 |
These results are visualised in the form of a bar chart in Figure 3.
Figure 3: A bar chart visualising how participants attributed authorship to a computer-generated news article
4. The NLG System as Author
Perhaps the most notable of these findings is the authorship attribution to Valtteri in particular and, for the sake of brevity, this paper will limit discussion to this finding. When faced with the computer-generated news article in question, 179 participants indicated that authorship should be attributed to the NLG system that generated it. Justifications for this response included:
‘The bot assembled the words and made meanings according to how it was trained to by these people but that doesn’t make them the authors, and just because some people gave some money to the project then they’re not the authors either! (otherwise I have just lost the authorship of my PhD thesis to the AHRC…)’
‘Let me indulge in an analogy here. An author could be sponsored by a foundation, company or government but his/her work is still his/hers. I’m giving the algortihm [sic] personality (only if it is programmed to produce non-predictable outcomes).’
Such justifications distinguish the system from its developers, suggesting that readers feel that the system is capable of creating sufficiently original textual content. In such instances, authorship is regarded as an act of individual expression. The process of assembling words, regardless of developer influence, is in itself enough for the system to warrant the ‘author’ title.
This distinction between system and developer is clearest in participants’ evocation of a parent-child metaphor. A parent, they asserted, cannot be credited as writing a text when the child was the one who penned it. The child (system) produces the text, while the parent produces the child. The parent (developer) is thus one step removed from the final product (the text). Such explanations included:
‘Valterri still did all the work regardless of the creators. If I write something, should my parents get credit?’
‘Even if the company Immersive Automation created the bot, the program Valtteri still composed the writing. If you said it was Immersive Automation wrote the article, it’s like saying that your parents created whatever you wrote, because they created you. Which is obviously not true.’
The parent-child metaphor is, of course, not without flaw. In its infancy, a child cannot be said to have much individual opinion or experience; the child’s actions and verbal output reflect the opinions and experiences of its guardians. More importantly for this discussion, though, is that within the parent-child metaphor rests an underlying comparison of computers to humans. While it was to be expected that participants would use analogies that drew upon their own lived experiences of human relationships, this comparison is particularly significant because it suggests that some readers draw upon their understandings of social networks and behaviours to negotiate where this new medium of text production fits within current cultural contexts. This aligns with a series of experiments conducted throughout the early 1990s that revealed that users automatically respond to digital and computational media as they would to other people or phenomena of the physical world, in accordance with cues exuded by the medium in question (Reeves and Nass 1996). The analysis of the results from this series of experiments notes that participants’ responses were likely not the result of anthropomorphism, as no participant ever argued that a computer should be treated as a person (Nass and Moon 2000, pp. 93-94). Nevertheless, likening computational processes to human processes speaks not only to perceptions of computational capability, but also to more general ways in which readers negotiate attitudes towards new technologies in light of their perceptions of current cultural circumstances.
When readers attribute authorship to Valtteri, an NLG system rather than a human, one may argue that the current conventional understanding of the author as an individual creative genius is at least somewhat called into question. Those connotations of authorship generally associated with humans – as per the questionnaire’s ‘author’ word association exercise, for example: ‘hard working’, ‘responsible’, and ‘education’ – are not so easily applied to NLG systems. However, other connotations – particularly those referring to ‘ownership’ and ‘responsible for intellectual content’ – may find relevance in a discussion of computer-generated texts. If, as in common understandings of authorship, authorship and ownership are considered synonymous (an understanding that is legally reinforced by automatic copyright protection), it would seem that by attributing authorship to Valtteri participants have implied that Valtteri is entitled to some kind of ownership of its output. Yet some questionnaire participants observed:
‘If I gave JK Rowling use of my house and paid for all her meals while she wrote Harry Potter, that would not make me an author or even a co-author of Harry Potter. If I had signed a contract with Rowling then I could possibly have the rights or the IP to Harry Potter, but still not an author. The funders of I.A. therefore are the owners of the text you showed me, but not the authors.’
‘In this case I still think the Valterri bot is the author, but ownership is separated from authorship because the bot cant own anything.’
The questionnaire’s qualitative results indicate that participants did distinguish authorship and ownership. Valtteri may have authored the text, but it is the owner (usually identified as the Immersive Automation team) who is morally and legally accountable for the material under consideration, and it is the owner who is entitled to any financial gain from the system’s output. ‘[T]he bot cant own anything’ because it lacks the financial interests in line with current capitalist cultural values. Accordingly, readers distinguish authorship and ownership in this instance, prompting alternative considerations of authorship that prioritise processes of creation over financial incentive.
5. Conclusion
The questionnaire results that have been presented in this paper indicate that authorship is an immensely vague concept: one that varies from individual to individual, and from context to context. In the ‘author’ word association exercise described above, participants associated ‘author’ with (1) an identity (2) associated with particular connotations (primarily adjectives, but also including some nouns), (3) as well as with particular activities (4) that result in particular kinds of (generally text-based) outputs. These associations are in line with the conventional understanding of authorship wherein the author is regarded as an individual creative genius motivated by intention-driven agency. Yet the authorship attribution exercise in the same questionnaire showed that participants were uncertain about where computer-generated texts fit within their own current conceptions of authorship. In trying to make sense of where the conventional hermeneutic contract fits within instances of computer-generated texts, readers soon realise that these understandings are insufficient. Indeed, computer-generated texts not only challenge traditional understandings of authorship, but also engender new understandings of authorship altogether as readers explore the conceptual gap between human and computer language production. The hermeneutic contract still stands, but needs some nuance to accommodate this new technology of text production, and our digital humanity more generally.
6. References
Automated Insights. (2017). Natural Language Generation in Your Daily Life. Medium. Available at: https://medium.com/@AutomatedInsights/natural-language-generation-in-your-daily-life-53c90c54bef0 [Accessed 28 August 2018].
Bootz P. and Funkhouser, C. (2014). Combinatory and Automatic Text Generation. In: M. Ryan, L. Emerson, and B. J. Robertson, eds., The John Hopkins Guide to Digital Media, 1st ed. Baltimore: John Hopkins University Press, pp. 83-85.
Case Studies: Bodybuilding.com. Automated Insights. Available at: https://automatedinsights.com/case-studies/bodybuilding [Accessed 15 March 2018].
Eve, M. P. (2017). The Great Automatic Grammatizator: writing, labour, computers. Critical Quarterly, 59(3), pp. 39-54.
Henrickson, L. (2018). Computer-Generated Fiction in a Literary Lineage: Breaking the hermeneutic contract. Logos, 29(2-3), pp. 54-63.
Hofstadter, D. R. (1980). Gödel, Escher, Bach: An Eternal Golden Braid. London: Penguin Books.
Krittman, D., Matthews, P. and Glascott, M. G. (2015). Innovation ushers in the modern era of compliance. Deloitte. Available at: https://www2.deloitte.com/content/dam/Deloitte/us/Documents/finance/us-fas-how-natural-language-is-changing-the-game-deloitte-only.pdf [Accessed 15 March 2018].
Magic Realism Bot. Twitter. Available at: https://twitter.com/MagicRealismBot [Accessed 13 April 2018].
NaNoGenMo. Available at: https://nanogenmo.github.io [Accessed 2 August 2018].
Nass, C. and Moon, Y. (2000). Machines and Mindlessness: Social Responses to Computers. Journal of Social Issues, 56(1), pp. 81-103.
Racter [Chamberlain, W. and Etter, T.]. (1984). The Policeman’s Beard is Half Constructed. New York: Warner Software/Warner Books.
Reeves, B. and Nass, C. (1996). The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. Cambridge: Cambridge University Press.
Sorjamaa, T. (2016). I, Author – Authorship and Copyright in the Age of Artificial Intelligence [PhD thesis]. Helsinki: Hanken School of Economics.
Valtteri. Available at: https://www.vaalibotti.fi [Accessed 11 April 2018].
Valtteri the Election Bot. Immersive Automation. Available at: http://immersiveautomation.com/valtteri-election-bot [Accessed 11 April 2018].