Who Made This? Algorithms and Authorship Credit

Responses to Assistance

A producer claiming authorship credit for a work product given assistance may create a troublesome disjuncture between norms regarding how work is produced and audience expectations about that work. Even though assistance may be both common and socially conventional in some domains (Phillips, 1997), outsiders—such as consumers, observers, or other stakeholders—will likely see some forms of assistance in a negative light. For example, a patron considering buying artwork from a well-known local artist may be less likely to do so upon learning that the artist did not actually paint the piece but only generated a high-level concept. Similarly, readers may respond negatively upon learning about a fiction author who only develops plotlines and leaves the actual writing to unnamed teams of ghostwriters, and concertgoers may dislike a musician who lip-syncs to their own recording during purportedly live performances.

Theory and empirical research relating to authenticity provides one lens to analyze when and why a specific producer might be attributed credit by laypeople, stakeholders, or consumers. Specifically, theories of nominal authenticity propose that individuals see work as genuinely authored or “real” is when it is certified by spatiotemporal verification: Did the producer actually create a particular work, or is it a “fake” or “fraud” (Dutton, 2003; Lehman et al., 2019; G. E. Newman & Bloom, 2012; G. E. Newman & Smith, 2016)? These theories fall short, however, in resolving the authorship attribution problem because they typically categorize these judgments as binary: a painting is either a “genuine” Picasso or not, depending on whether he actually painted it. Realistically, the opacity of collaborative work suggests that these judgments may be more nuanced—and nonbinary—when people are given information about a producer’s involvement alongside assistance. Jauregui (1996) notes that domains of painting, sculpting, music performance, and even athletics are “. . .seldom individual endeavors”, but are instead contexts where industry experts “. . .pick thresholds for determining authorship” in line with established norms. These arguments echo broader philosophical debates about the meaning of authorship (Barthes, 2001; Foucault, 2017; Wilson, 2004), and they have inspired empirical research concerning how people attribute credit in similarly opaque human teams where work is potentially unequally distributed (e.g., Biernat & Sesko, 2013; Heilman & Haynes, 2005; Maglio et al., 2020). Despite these studies, the authorship attribution problem is complicated by emerging realities about the changing nature of work and production: Humans can be assisted—in part or wholly—by sophisticated algorithms driven by large datasets and/or artificial intelligence.

Algorithmic Assistance With Work

Human assistance with work products can both undermine perceptions of authorship and possibly precipitate negative economic consequences for producers (Carroll & Kovács, 2021). Do audiences respond similarly to algorithmic assistance and human assistance? Algorithmic processes assist and augment human work in many contexts and ways (Brynjolfsson & McAfee, 2014). For example, algorithms can generate unique artwork, transform images to match different artistic styles or color palettes, write poetry using natural language processing, or predict a business’ likelihood of success (e.g., Deep Dream Generator, 2022; Midjourney, 2022; OpenAI, 2022). To many analysts, the sophistication of contemporary algorithmic processes implies that humans can be augmented by intelligent technologies in most important work domains (e.g., Raisch & Krakowski, 2021).

Accordingly, a stream of research investigates how people respond to and interact with algorithmic systems. Much research focuses on when and why people listen to algorithmic advice or perceive algorithms’ decisions as (un)fair (e.g., Castelo et al., 2019; Dietvorst et al., 2015; Jago & Laurin, 2022; Longoni et al., 2019; D. T. Newman et al., 2020). These studies underpin emerging theories concerning the “psychology of technology” and the development of comprehensive models explaining how people respond to, and think about, algorithmic processes, particularly given human frames of reference (Logg, 2022).

Extensive algorithmic assistance with work generates a theoretical puzzle: If an algorithm significantly assists a producer, do people still credit that producer with authorship? We posit that both human and algorithmic assistance undermine attributions of authorship credit compared with producers’ unassisted work. Moreover, we hypothesize that algorithmic (vs. human) assistance results in producers receiving greater authorship credit, even if the work these agents performed is otherwise identical. We additionally propose a theoretically novel mechanism underlying this process, namely, that this effect occurs because people assume algorithmic assistants require more producer oversight than do human assistants.

This prediction rests on three assumptions people typically make regarding algorithmic work. First, employing, maintaining, and interpreting algorithmic processes to help with work typically appears to require a level of technological sophistication that few people might possess, compared with the skills necessary to manage other humans (Edison & Geissler, 2003). Second, research on algorithm aversion suggests that people see algorithms that make mistakes as flawed, requiring more producer contribution or intervention (i.e., oversight processes) to improve decisions (Dietvorst et al., 2015). This argument dovetails with research suggesting that people see algorithms as lacking various elements of mind, including the potential for noticing and correcting mistakes (Bigman & Gray, 2018; Gray et al., 2007; Reich et al., 2022). Finally, if people assume algorithms are less capable of iterative improvement, then they may also assume that human assistance reflects a more concerted effort for a producer to do relatively little work while simultaneously maintaining the economic and social benefits stemming from the realization of a specific work product (e.g., a celebrity commissioning an uncredited ghostwriter to produce an autobiography, see Smith & Newman, 2014).

Importantly, these attributions are rooted in perceptions of the amount of creative or physical work a producer exerts, not the intrinsic deservingness of an assisting agent. Because people do not believe algorithms have the experiential capacity for suffering (e.g., Gray et al., 2007), people may be relatively willing to assign producers credit-given algorithmic assistance simply because they perceive the assisting agent as incapable of financial, political, or emotional harm. Although perceptions of both producer and assistant undoubtedly matter in terms of parsing authorship credit in work arrangements, this argument focuses primarily on perceptions of the producer’s effort, as opposed to the assistant’s deservingness. We formalize this logic in the following hypotheses:

Critically, we propose that these variations in how people interpret algorithmic and human assistance occur even when these agents’ contributions toward a particular work outcome are otherwise identical. In many—if not most—production domains, identical work is increasingly possible given algorithms’ sophistication alongside organizations’ increasing access to powerful computational resources.

Although much research explores the consequences of automation itself (e.g., an algorithm making a decision in place of a human), the studies reported below examine how people respond to augmentative algorithmic processes (e.g., an algorithm helping a human make a decision). Algorithms may be changing perceptions of authorship and credit, which we argue ultimately coalesces credit in ways that benefit producers who choose to outsource elements of work to such technologies. One potential consequence of technological change is inequality: Powerful people and organizations, such as owners or developers, often benefit to a greater extent than others when technology advances (Brynjolfsson & McAfee, 2014). In parallel, producers with both the necessary resources and skills to augment work with algorithmic processes may largely be seen as more responsible for that work, potentially exaggerating existing advantages and reputations.

This research also expands theory broadly related to authenticity, particularly spatiotemporal conceptualizations of the construct, for example, “normative” or “historical” authenticity (Newman & Smith, 2016). Specifically, the hypothesized arguments suggest that algorithmic assistance clarifies attributions of authorship credit in ways that human assistance does not, suggesting that different theories relating to authenticity and ownership should similarly reflect differences in how people perceive human and nonhuman assisting work contributions. Finally, these explorations into the “producer-side” mechanism of oversight in explaining authorship credit attributions offer a framework for how individuals, groups, and even organizations can implement and justify work assistance in ways that preserve beneficial attributions of credit.

Overview of Studies

We tested these two hypotheses across four experiments. The experiments used a variety of samples as well as work domains (e.g., art, construction, sports analytics, small-business proposals) to enhance the generalizability of findings. We first investigated whether individuals attribute more authorship credit to producers using algorithmic (vs. human) assistance (H1) in samples of community college students (Study 1) and American adults (Study 2) while also examining the mediating role of oversight (H2). We next tested the mechanistic role of oversight using a moderation design, experimentally manipulating whether a producer extensively supervised an assisting agent or not (Study 3). Finally, we investigated the consequences of these effects for entrepreneurship, asking passersby on an urban university campus about a business proposal where an algorithm (vs. human) meaningfully assisted with a small-business idea (Study 4). We preregistered Studies 2, 3, and 4 using AsPredicted.org, report all measures and manipulations for all studies, and did not exclude any participants’ data. We aimed ex-ante for 100 participants per experimental cell in all studies, save for Study 1, which was open-enrollment, and Study 4, where we aimed for 50 participants per experimental cell because of the estimated length of data collection. We collected age and gender demographics in all studies but did not collect participants’ race or ethnicity information. All experimental materials and data are posted via the Open Science Framework: https://osf.io/wp3z6/?view_only=c96aee7557844924969a30343672c3a7.

Study 1

In Study 1, we pursued three goals. First, we wanted to verify empirically our intuition that assistance—whether from a human or algorithm—reduces perceptions that a producer deserves credit (i.e., authorship and recognition) for a specific piece of work. In the study, we measured assigned authorship credit both before and after participants learned about work assistance. Second, we sought to investigate whether individuals attributed a producer with more authorship credit given an algorithm’s assistance, compared with analogous human assistance (H1). Finally, we tested whether perceptions that algorithmic assistance requires relatively more producer oversight served as a mechanism for this credit effect (H2).

Results and Discussion

Authorship

At Time 1, we found no significant difference between the human (M = 5.73, SD = 1.12) and algorithm (M = 5.72, SD = 1.13) conditions—which had not yet been presented to participants—in terms of attributed authorship credit, t(348) = 0.045, p = .962, d = .004, CI₉₅ = [−.20, .21]. We primarily included this condition to verify that, absent any information about an assistant, participants were willing to ascribe credit to the producer for this particular artwork. However, at Time 2—following the experimental manipulation—participants in the algorithm condition (M = 5.13, SD = 1.30) attributed the producer with significantly more authorship credit compared with participants in the human condition (M = 4.48, SD = 1.14; t(348) = 4.93, p < .001, d = .53, CI₉₅ = [.28, .70]). A significant interaction stemming from a 2 (Condition, Between-Subjects) × 2 (Authorship Timing, Within-Subjects) repeated-measures ANOVA verified that the authorship credit decline was significantly greater in the human condition, compared with the algorithm condition, F(1, 348) = 24.96, p < .001, η²p = .07, CI₉₀ = [.03, .11]. As we expected, this model also indicated a main effect of timing (F(1, 348) = 200.32, p < .001, η²p = .37, CI₉₀ = [.30, .42], suggesting that participants attributed less credit to the creator after learning about any form of assistance. This model also indicated a main effect of agent such that participants broadly ascribed more credit-given algorithmic assistance (F(1, 348) = 8.77, p = .003, η²p = .03, CI₉₀ = [.005, .06]). Separate general linear models using assistant type, each potential moderator (anthropomorphism, technological affinity, art interest, hours spent creating paintings/drawings each week, or hours spent creating art each week), and their interaction as predictors indicated that none of these variables significantly moderated the relationship between agent type and authorship at Time 2 (i.e., following the assistant type manipulation; all F < 1.24 and all p > .267). However, the main effect of assisting agent remained significant in each model (all F > 18.82, all p < .001). These results suggest that people not only see producers as more deserving of credit when using algorithmic (vs. human) assistance, but also that these attributions are not influenced by personal anthropomorphic tendencies, attitudes toward technology, or interest in or experience with art.

Oversight and Distributive Justice

Participants, on average, indicated that the producing artist oversaw the algorithm’s work to a significantly greater extent (M = 5.06, SD = 1.11) compared with the human assistant’s work, M = 4.70, SD = 0.99; t(346) = 3.21, p = .001, d = .35, CI₉₅ = [.13, .56]. Put differently, participants assumed the algorithm required more supervision than the human in executing its work tasks, operating less independently. Participants also indicated they were significantly more concerned about potential distributive justice violations in the human condition (M = 4.57, SD = 1.05) compared with the algorithm condition, M = 4.18, SD = 1.37; t(346) = 3.02, p = .003, d = .32, CI₉₅ = [.11, .53]. A 5000-iteration bootstrapped mediation model using assistant type (0 = person, 1 = algorithm) as the independent variable, authorship (measured at Time 2) as the dependent variable, and both oversight attributions and distributive justice concerns as simultaneous mediators revealed that both oversight (CI₉₅ = [.06, .27]) and distributive justice concerns (CI₉₅ = [.004, .12]) mediated the lower authorship credit-given human (vs. algorithmic) assistance with the artwork (see Figure 1). An exploratory 5000-iteration agent → oversight → justice concerns → authorship credit serial mediation model indicated that this serial pathway was not significant (CI₉₅ = [−.001, .02]), whereas the two pathways remained significant on their own (oversight: CI₉₅ = [.05, .27]; justice: CI₉₅ = [.003, .11]), suggesting they operated somewhat independently (see Supplemental Materials for more detailed summaries of mediation results across all studies).

OPEN IN VIEWER

Figure 1.

Mediation to Authorship Credit Through Both Oversight and Distributive Justice Concerns (Study 1). All Coefficients Refer to Unstandardized Betas.

Study 1 provides initial evidence for both H1 and H2 in the work domain of artistic production. Consistent with research on authenticity, participants saw assistance of any kind as threatening to artists’ authorship credit over a painting. However, they also attributed more credit to this artist when algorithmic assistance was used, compared with human assistance, even when the exact work and assisting functions the two agents performed were identical. Finally, we observed evidence that perceptions of oversight—specifically, that the algorithmic assistance requires more supervision on the part of the claiming creator—explained some of this effect.

Given that authorship credit in situations can be rooted in perceptions of both a claiming producer as well as perceptions of their assistant(s), people’s attributions about both agents can meaningfully influence this process. These results provide initial evidence that “producer-side” assumptions about a creator’s effort or contributions to a work project—in the form of oversight— explains (some of) the hypothesized credit discontinuity, in addition to “assistant-side” perceptions that algorithms are not deserving of credit, echoing mind perception research that algorithms cannot meaningfully suffer and thus may not be emotionally, politically, or even financially deserving of authorship (e.g., Gray et al., 2007). Study 1 also suggests that these results are robust when exposed to a series of potential moderators, including participants’ natural anthropomorphism of algorithms, technological affinity, and personal interest in the work domain.

Study 2

In Study 2, we used a similar paradigm as Study 1—a credit-claiming producer receiving algorithmic vs. human assistance—across a variety of different work domains. Our goal in doing so was to investigate if participants’ assessments of greater authorship (H1) as explained by more required oversight (H2) given algorithmic (vs. human) assistance generalized beyond painted art, to other work domains. In Study 2, we also measured oversight assumptions to test as a mediator. We preregistered this study using AsPredicted.org: https://aspredicted.org/56jj2.pdf.

Results and Discussion

Oversight and Authorship Credit

Consistent with our preregistration plans, we first collapsed across the five different work domains to test whether people’s oversight and authorship attributions differed as a function of type of assistant. Results indicated that participants assumed the algorithm required more oversight (M = 5.51, SD = 0.99) compared with the human assistant, M = 5.18, SD = 1.04, t(502) = −3.63, p < .001, d = .32, CI₉₅ = [.15, .50]. Participants also attributed the producer with more authorship credit when using algorithmic assistance (M = 5.46, SD = 1.06) compared with human assistance, M = 4.99, SD = 1.30, t(502) = −4.49, p < .001, d = .40, CI₉₅ = [.22, .58], despite the two agents’ contributions being identical.

Differences Across Work Domains

We next conducted two 2 (Agent: Person vs. Algorithm) × 5 (Work Domain) ANOVAs predicting oversight and authorship, respectively. These models investigated whether the effects meaningfully changed across the different work domains. Predicting oversight, we observed a main effect of agent type consistent with the above results, F(1, 494) = 12.88, p < .001, η²p = .03, CI₉₀ = [.007, .05]. We also found no effects of work domain, F(4, 494) = 0.36, p = .839, η²p = .003, CI₉₀ = [.00, .006], and no interaction between agent type and work domain, F(4, 494) = 0.02, p = .999, η²p < .001, CI₉₀ = [.00, .00]. Predicting authorship credit, we again observed a significant effect of assisting agent by type, F(1, 494) = 19.75, p < .001, η²p = .04, CI₉₀ = [.02, .07], but no main effect of work domain, F(4, 494) = 0.26, p = .906, η²p = .002, CI₉₀ = [.00, .003], and no interaction between these two variables, F(4, 494) = 0.82, p = .515, η²p = .007, CI₉₀ = [.00, .02]. Both of these models suggest that participants’ assumptions about oversight and authorship given human versus algorithmic assistance were relatively similar across the different work domains.

Mediation Through Oversight

As with Study 1, we estimated a 5,000-iteration bootstrapped mediation model investigating whether participants’ beliefs that the algorithmic work required greater oversight on the part of the creator mediated their assignment of greater authorship credit attributions for algorithmic (vs. human) work. This model used authorship credit as the dependent variable, assisting agent (0 = person, 1 = algorithm) as the independent variable, and oversight as the mediator. As with Study 1, we observed significant mediation, CI₉₅ = [.10, .35], indicating that assumptions that algorithms require more oversight mediated the increased authorship credit assigned to the producer (see Figure 2).

OPEN IN VIEWER

Figure 2.

Mediation to Authorship Credit Through Oversight (Study 2). All Coefficients Refer to Unstandardized Betas.

Along with Study 1, Study 2 indicates that algorithmic assistance does not undermine the authorship credit attributed to a producer to the degree that human assistance does. Specifically, we observed evidence for both H1 and H2 across a variety of different work domains. When producers were assisted by algorithms, participants believed they exercised more oversight—and thus deserved more authorship credit—compared with when producers were assisted by humans. Importantly, these effects occurred despite exactly identical assistant work functions in each domain, for example, generating and coloring a specific image in a logo or calculating heat and air pressure when glassblowing.

Study 3

In Study 3, we tested H2—that people assume algorithms require more producer oversight, which explains increased credit attributions—using a moderation design. Specifically, in addition to a control condition resembling Studies 1 and 2, we incorporated a second condition where participants received unambiguous information that a producer exercised significant supervisory oversight over either human or algorithmic assistance in creating software that predicts the outcomes of baseball games. This “moderation-of-process” (Spencer et al., 2005) approach allowed us to examine more thoroughly this oversight mechanism above and beyond the mediation strategies employed in Studies 1 and 2, particularly given that other assistant-side (vs. producer-side) mechanisms can reasonably influence credit attributions. In addition, Study 3 also aimed to explicitly address two potential confounding factors. First, Studies 1 and 2 examined uncredited assistance (i.e., instances where it was later discovered that a producer was assisted by a human or algorithm), which could possibly change people’s attributions, compared with upfront credited assistance. In Study 3, participants read about a claiming producer’s work alongside either explicitly credited human assistance or explicitly credited algorithmic assistance with that work. Second, it is possible that people assume algorithmic and human work differs in terms of quality, which could manifest in perceptions that creators have differential influence over and/or deserve different levels of credit for final work products. As such, we measured perceptions of work quality to use as a covariate. We preregistered this study using AsPredicted.org: https://aspredicted.org/e98jr.pdf.

Results and Discussion

A 2 (Assistant: Person vs. Algorithm) × 2 (Oversight Condition: Control vs. Oversight) ANOVA indicated a main effect of assistant condition, F(1, 399) = 7.80, p = .005, η²p = .02, CI₉₀ = [.003, .05], no significant effect of oversight condition, F(1, 399) = 1.48, p = .224, η²p = .004, CI₉₀ = [.00, .02], but crucially, a significant interaction between the two, F(1, 399) = 6.07, p = .014, η²p = .02, CI₉₀ = [.002 .04] (see Figure 3). Simple effects’ analyses revealed that, in the control condition, participants attributed greater authorship credit to the creator given algorithmic assistance (M = 5.77, SD = 0.74), compared with human assistance, M = 5.29, SD = 1.11; F(1, 399) = 13.82, p < .001, η²p = .03, CI₉₀ = [01, .07]. However, this difference vanished in the high oversight condition, where participants attributed similar credit to the producer given algorithmic (M = 5.66, SD = 1.01) and human (M = 5.63, SD = 0.81) assistance, F(1, 399) = 0.54, p = .816, η²p = .000, CI₉₀ = [.00, .01]. Analyzed differently, the oversight information significantly increased participants’ credit attributions given human assistance, F(1, 399) = 6.78, p = .010, η²p = .02, CI₉₀ = [.002, .04], but not given an algorithm’s assistance, F(1, 399) = 0.78, p = .379, η²p = .002, CI₉₀ = [.00, .02].

OPEN IN VIEWER

Figure 3.

Authorship Credit as a Function of Assisting Agent and Oversight Condition (Study 3). Error Bars Represent 95% Confidence Intervals of Means.

An identical ANOVA predicting assumptions about the quality of the assisting agent’s contribution indicated no significant main effects or interactions, Fs(1, 398) < 1.60, ps > .209, η²ps < .005. Including this measure as a covariate (ANCOVA) in the above model did not change the significance thresholds for any effects: Specifically, this model still indicated a main effect of assistant condition, F(1, 397) = 11.15, p = .001, η²p = .03, CI₉₀ = [.007, .06], no main effect of oversight condition, F(1, 397) = 0.50, p = .479, η²p = .001, CI₉₀ = [.00, .01], and a significant interaction between the two, F(1, 397) = 8.04, p = .005, η²p = .02, CI₉₀ = [.003, .05]. Similarly, mediation models using condition (0 = person, 1 = algorithm) as the independent variable, authorship credit as the dependent variable, and quality as the mediator indicated that quality did not mediate the effect of assisting agent on authorship, CI₉₅ = [−.09, .09]. Accounting for oversight as a moderator of either the assistant to quality link (CI₉₅ = [−.21, .17]) or the quality to authorship link (CI₉₅ = [−.02, .02]) revealed no moderated mediation, and quality did not significantly mediate the effect of assisting agent in either the control or high oversight conditions (CIs₉₅ = [−.13, .15], [−.13, .11], [−.11, .09], and [−.10, .08], respectively).

In sum, Study 3 provides evidence via moderation about the role of oversight in explaining the credit differential between human and algorithmic assistance observed in Studies 1 and 2. Although there are other potential explanations for this effect (e.g., that algorithms are less deserving of credit because of their not needing resources or lacking emotions and the capacity for suffering; Gray et al., 2007), participants no longer distinguished between the amount of credit a producer deserved once that producer exercised significant oversight over an assistant’s (relatively large) contribution to work. This moderation result reinforces the mechanistic importance of people’s assumptions about a producer’s role in work when it comes to crediting authorship, in addition to any assumptions or beliefs people may have about the intrinsic moral or economic deservingness of an assisting agent. One limitation worth noting is that—although Study 3 adhered to our heuristic of recruiting 100 participants per experimental cell and employed relatively heavy-handed manipulations—this sample size may have been somewhat underpowered to detect an attenuated interaction (Simonsohn, 2014). In this study, we additionally observed no evidence that people assume human and algorithmic work differs in quality, nor that quality perceptions or influence credit attributions.

Study 4

Our primary goal in Study 4 was to investigate downstream consequences stemming from the increased authorship credit producers receive given algorithmic assistance, compared to analogous human assistance. If people see producers’ work as more “their own” when those producers are assisted by algorithms, does this manifest relatively greater support for those producers’ work? Given how algorithms can inform—and even autonomously make—a variety of different business decisions including conceptualization, initial planning, and staffing and budget forecasting (e.g., Growth Science, 2022), we investigated this question in the context of entrepreneurship. We predicted that people would assign an entrepreneur more credit when assisted by an algorithm (vs. human) in developing a small-business proposal, which would in turn increase support for that business and confidence in its success. We preregistered this study using AsPredicted.org: https://aspredicted.org/nf9xw.pdf.

Results and Discussion

Consistent with previous studies, participants attributed significantly more authorship credit to the entrepreneur receiving algorithmic assistance (M = 5.64, SD = 0.93) compared with human assistance, M = 4.90, SD = 1.26; t(107) = 3.48, p = .001, d = .67, CI₉₅ = [.28, 1.05]. Participants were also significantly more likely to endorse university investment in the startup given algorithmic assistance with the business plan (46/53) compared with human assistance, 36/52; χ² (N = 105) = 4.73, p = .030; four participants—three in the human condition and one in the algorithm condition—elected not to answer this investment question. Finally, participants provided a lower estimated chance of the organization failing within two years when an algorithm assisted in creating the business plan (M = 46.64, SD = 19.25) compared to another person, M = 57.15, SD = 23.57; t(107) = −2.54, p = .012, d = .49, CI₉₅ = [.10, .87].

We next estimated two 5,000-iteration bootstrapped mediation models to investigate whether higher authorship credit attributions from algorithmic assistance (0 = person, 1 = algorithm) mediated participants’ greater willingness to recommend the university invest in the business as well as lower estimates that the business would fail. We observed significant mediation through authorship credit predicting both outcome measures, CI₉₅ = [.19, 1.44] and CI₉₅ = [−9.24, −.62]: Greater authorship credit from algorithmic assistance mediated a greater willingness to vote for university investment in the company, as well as estimates that the business was less likely to fail. The condition manipulation fell to nonsignificance in both models accounting for authorship credit (p = .375 and p = .132, respectively; see Figure 4).

OPEN IN VIEWER

Figure 4.

Mediations Through Authorship Credit to Investment Votes and Failure Estimates (Study 4). All Coefficients Refer to Unstandardized Betas.

Study 4 demonstrates how and why authorship credit stemming from algorithmic and human assistance can matter in the form of stakeholder support for assisted producers. In the context of business planning, participants believed that an entrepreneur’s small-business proposal was both more deserving of support and more likely to succeed when that entrepreneur was assisted by an algorithm, compared with another person. Crucially, participants’ attributions that the algorithm-assisted entrepreneur deserved more credit for this proposal and its contents mediated these effects. As such, Study 4 illustrates that the nature of assistance producers receive can shape audience and consumer support for different work initiatives, even when—as with Studies 1–3—the exact same support is provided by humans or algorithms. As people outsource more and more decision processes to algorithmic agents, this study suggests that they will better maintain the often-beneficial credit attributions for that work, compared with if they outsourced that same work to other humans.

General Discussion

Contemporary artist Jeff Koons prompted a great deal of public discourse regarding artistic credit when he announced that he had approximately 150 people on his payroll and never personally picked up a paintbrush (Sesser, 2011). Instead, Koons revealed that he had developed a system whereby his assistants used stencils to paint according to his intentions (Yu, 2020). Other contemporary artists have ignited similar discussions. Sculptor Henry Moore, for example, routinely poses for photographs showing him sculpting large blocks, even though his work is enlarged by assistants from palm-sized originals and finished in bronze foundries (Saunders, 1993). Glassblowers also often collaborate with assistants and technicians to accomplish specific forms and styles, ultimately claiming credit for pieces they may or may not have played an important role in physically creating (O’Connor, 2005). In a variety of work domains, such as artistic production, idea generation, analytics, or business planning, these kinds of exemplar situations raise important questions about credit and production: Does a producer who was assisted by someone deserve credit for that work? Here, we found that the nature of this question is changing given the increasing proliferation of algorithms able to augment—or even autonomously emulate—human work.

In the studies presented here, we found that the type of agent that provides production assistance impacts how laypeople assign credit for work (Studies 1–3) as well as the corresponding support they offer producers (Study 4). Specifically, across a variety of work domains, levels of collaborative contribution, and economic contexts, we found evidence that people assign more credit to credit-claiming producers when algorithms assist those producers, compared with when humans provide the same assistance (H1). The difference emerged, in part, because people assumed algorithms require more producer oversight (H2).

Although automation processes may be changing definitions of what audiences consider “real,” one key consideration is whether people judge an object as actually produced by the person to whom it is formally credited. As algorithms assist with work in an increasing variety of domains, the attribution of credit is changing in ways that favor producers’ claims of authorship and any subsequent benefits associated with such claims. Accordingly, theories of authenticity will likely benefit from empirically investigating the philosophical puzzles collaborative work inspires. Do people really regard an artist as having painted the painting if they later learn that they did not do the work on their own (e.g., Study 1)? Here, we find that the type of assisting agent involved in collaboration can meaningfully shift people’s attitudes and inferences in such situations.

Future research could explore how other facets of situations change credit inferences. For example, a producer’s hierarchical relationship with an assistant (e.g., an artist being assisted by an employee, student, or algorithm they themselves programmed) may be important to observers, as are the assistant’s inferred motivations (or lack thereof given algorithms’ seeming mindlessness; Gray et al., 2007) for engaging in the work. Relatedly, the studies reported here showcase how assumptions about producers’ roles and effort can meaningfully shape credit attributions given both human and algorithmic assistance. Research on mind perception and on human–computer interaction broadly indicates that assumptions about assisting and/or (potentially) morally exploited agents could meaningfully deprive nonexperiential algorithms of authorship credit, as such agents do not necessarily socially, politically, or emotionally benefit from such attributions (Gray et al., 2007; Waytz & Norton, 2014). Independent from this research pointing to discontinuities in perceived assistant deservingness, the studies presented here indicate that people’s assumptions about producers’ effort and/or involvement is also shaped by algorithmic augmentation. Future research could more thoroughly explore production- and assistant-side interpretations of effects such as these, including the elasticity of social attributions, such as credit assignments given augmented work.

Automation tends to concentrate wealth among the people and organizations using new technologies (Brynjolfsson & McAfee, 2014). These results suggest that credit—and the benefits it manifests—may become similarly concentrated as human labor is increasingly outsourced to algorithms that people assume require greater oversight. Importantly, these attributions may occur regardless of how much supervision these technologies actually require from producers. A musician, for example, could outsource most of a song’s composition and recording to an algorithm; these studies suggest that they will still maintain relatively desirable attributions of production credit. Although producers may perceive strong incentives to conceal assistance altogether, those wishing to capitalize on the benefits associated with authorship may come to highlight technologies’ roles in creation processes. So, somewhat paradoxically, intentionally dehumanizing assistants may benefit producers when it comes to audience perceptions.

There are many potential boundary conditions for these effects that the present studies did not fully explore. People’s (or producers’) experience with work domains, their own signaled technological affinity, or the specific functions of technologies could all meaningfully influence their assumptions about credit derived from algorithmic or augmentative work. Although we did not observe evidence for moderation along these lines in Study 1, it stands to reason that these attitudes and traits may become more relevant as algorithms become more commonplace and accepted. In addition, algorithms are becoming increasingly collaborative, fostering environments where such technologies might appear to require less producer oversight, thereby attenuating the relative credit effects observed here. For example, voice assistants such as Google Assistant, Siri, and Alexa can engage in back-and-forth conversations with users trying to accomplish a task, and digital artists can increasingly use artificial intelligence to generate art via keywords, refine these keywords, and work collaboratively “with” art-focused algorithms in tandem (e.g., DALL-E, which generates art using natural language prompts), as opposed to treating such technologies as simply a tool. Future research could better explore people’s attributions in such situations, although the exact nature of producer-algorithm production processes could remain relatively opaque to interested stakeholders or consumers, even as such technologies evolve. Some such technologies might actually require less oversight than analogous human assistance. Although we observed consistent evidence—via both mediation and moderation—that people assume algorithms require more oversight in a variety of different work and social contexts, it stands to reason that as technologies evolve and are used for new or different purposes, people’s assumptions about producers’ roles relative to algorithms will also shift.

These experiments also broadly speak to how people’s assumptions about machines manifest in different domains, which relates to emerging theorizing regarding the psychology of technological change. Specifically, we found evidence in Studies 1 to 3 that people assume algorithmic work requires more producer oversight, compared with analogous human work. This effect could be due to perceptions that algorithmic systems are “mindless” in some-but not necessarily in all-important domains (Bigman & Gray, 2018). People may also perceive algorithms as uncreative and unable to fix mistakes (Dietvorst et al., 2015). However, this is often not the case: Many algorithmic systems threaten disruption by their capability to engage in precise augmentative work. Put differently, algorithms may not necessarily require the level of oversight people assume they require, creating a potential discontinuity between producers and audiences. Further complicating these discussions are arguments regarding the myriad agents who could possibly receive credit for algorithmic work, including those utilizing the algorithm, developers who designed it, and/or companies that funded its creation or implementation. Each of these agents can presumably make authorship claims in a variety of social and business contexts, raising questions about the extent to which algorithms themselves are deserving of work credit, compared with more emotionally capable humans involved in their design or deployment (see Waytz & Norton, 2014).

The studies presented here also raise the intriguing question of whether algorithmically augmented work is perceived to be “authentic” or not. On one hand, increased attributions of authorship credit could boost the perceived authenticity of work producers do not even physically craft or engage with, for example, a famous artist generating an idea for artwork and using keywords entered into an AI-powered algorithm to generate an image (e.g., OpenAI, 2022). On the other hand, people do not believe algorithmic work is “authentic” on its own, despite the fact that connections to human creators can foster authenticity attributions (Jago et al., 2022). Given simultaneous advances in computing and the increasing sophistication of new technologies able to autonomously engage in and generate high-quality work, future research could more precisely disentangle these competing perceptions. It could also explore how algorithmic augmentation shapes other relevant perceptions of both creative and noncreative work in different economic (e.g., an artist selling a painting; Studies 1 and 2) and social (e.g., a nonprofit generating a fundraising plan; Study 4) contexts.

Algorithms are increasingly easy to use, commercially available, and can engage in a variety of work and creative tasks, including generating artwork using keywords, writing, making decisions, generating ideas, and processing natural language to facilitate meaningful conversations and interactions. As algorithms become more commonplace, an intriguing question arises about people’s assumptions regarding these systems—specifically in comparison to analogous human work—and whether these assumptions will change over time. Future research concerning the psychology of technology will likely benefit from cataloging these potentially shifting attitudes and assumptions across time, given the changing landscape of algorithms and the many tasks they can successfully assist with or otherwise accomplish.

Conclusion

The question of whether credit-claiming producers actually “made” a specific object, work product, or decision can be deceptively complex. Indeed, research on authenticity and collaborative work credit often struggles to answer such questions. Systematically examining observers’ reactions to, and inferences about, such situations sheds empirical light on these puzzles, including how these situations can manifest in important social and commercial domains. In this research, we found that the credit given to producers using algorithmic assistance ultimately benefits those producers: Collaboration with algorithms concentrates producers’ relative legitimacy in claiming credit for work. Research on authenticity suggests that producers’ answers to questions of “Did you make this?” can have important economic and social consequences. As algorithmic technologies increasingly assist with work and creation processes, they may similarly be lessening audiences’ skepticism toward producer responses of “Yes, I did.”

Supplemental Material