Artificial Intelligence in Marketing: From Computer Science to Social Science

Large language models (LLMs), like OpenAI's GPT4o, produce results on many tasks that approximate human performance (e.g., Bubeck et al. 2023). Such developments would have been difficult to imagine only a few years ago and have important repercussions for the individual, organizational, and societal impact of marketing activities. To make one example, new findings suggest that LLMs are quickly acquiring “super-human” powers of persuasion (Burtell and Woodside 2023; Matz et al. 2024; Salvi et al. 2024) with far-reaching implications for the practice of marketing. It is difficult to overstate the importance for humankind of the development of thinking machines. Marketing academics and practitioners need help understand the changes taking place and their ramifications.

In this fast-developing landscape, Grewal, Guha, and Becker (2024) offer important guidance. Among other contributions, they point out how artificial intelligence (AI) will impact lives and society to an important extent via activities that are core to marketing. For example, many of the most prominent applications of LLMs are consumer-facing, such as customer service chatbots and productivity-boosting apps like ChatGPT. At the same time, improvements in areas such as face recognition, image generation, and recommender systems are affecting consumers as they shop and browse the web. This commentary adds to Grewal et al.'s analysis by making the case for the importance of a behavioral science perspective in studying and designing marketing applications of AI (for related analyses, see Hermann and Puntoni 2024; Puntoni et al. 2021). I will use the context of AI bias as an illustrative example.

Bias Emerges from an Interaction Between Algorithmic and Human Decisions

The implications of AI deployment for inclusiveness and discrimination are a topic of intense discussion (e.g., Grewal, Guha, and Becker 2024; Hermann, Yalcin Williams, and Puntoni 2023). Many news stories have appeared in recent years where AI technology was shown to be biased and exacerbate inequality. For instance, in 2018 Amazon famously had to withdraw an algorithm used to sort job candidates when the algorithm turned out to favor males over females in hiring decisions (Reuters 2018). The main source of bias discussed in both academic and practitioner circles is the training data. AI technology mimics human decision making by learning patterns in training data. When the training data is biased, algorithmic decisions will replicate and reinforce the same biases. For example, the reason why Amazon's hiring algorithm was sexist is that the algorithm learned to associate strong profiles for software developer jobs with being male because the large majority of software developers in the training data were males.

While the problem of AI bias emerging from biased training data can be hard to solve, it is well-understood, as it belongs to the “garbage in-garbage out” category of problems that has long been familiar to computer scientists. Thus, the antecedents of AI bias as well as its solutions are typically assumed to be a “merely” technical problem. That is a mistaken view. In fact, bias can enter AI predictions via a variety of mechanisms and, for many of these, technical solutions won’t be sufficient. To address the issue of AI bias, behavioral insights are necessary. Below I highlight a few examples.

Lambrecht and Tucker (2019) document sexist outcomes in Facebook advertising. They conducted an ad campaign for STEM jobs and noticed that, while all genders were targeted by the campaign, the ad was more frequently served to male than female Facebook users. The difference was especially large for the 35–44 age group, where the number of female impressions was less than 70% the number of male impressions. The price paid per impression was the same for the two genders and, strikingly, the clickthrough rate was overall higher for females than males, indicating that lower ad effectiveness among females was not the reason for the biased targeting. Instead, biased outcomes emerged as a result of market forces that are independent from the algorithm. Women, especially those in the 35–44 age group, are heavily targeted by advertisers on Facebook. As a result, for the same price per click, more auction bids were lost for female and male eyeballs, resulting in a strongly skewed sample.

Bias can emerge as a result of an interaction between algorithmic and human decision making for a variety of other reasons. A key factor is user adoption patterns. Zhang et al. (2021) examined the effect of the deployment at Airbnb of an algorithm designed to help hosts set prices for their properties. One of the goals of the algorithm was to help reduce a racial bias in nightly rates, as White hosts tend to earn more than Black hosts from comparable properties. The algorithm helps hosts set the right price to maximize revenues and it was thought that this would help reduce the earning gap between Black and White hosts. However, the researchers observed that after introduction of the algorithm the racial gap increased by almost 20%. This unintended effect was the result of differential patterns of algorithm adoption among hosts. Looking only at hosts who adopted the algorithm, the racial gap in earning was reduced by over 70%. The algorithm therefore worked in the intended manner. The problem is that fewer Black hosts adopted the algorithm, with the results of widening the racial earning gap and increasing inequality. In an even more disturbing example, Davenport (2023) documents the effect of the rollout of an algorithm to predict reoffending and court non-appearance in the New Jersey justice system. The algorithm ended up exacerbating racial inequalities because officers were strategic in their decision of whether to check the score of the algorithm for a given defendant. The decisions of whether to rely on the algorithm was a function of whether the algorithm would support harsh penalties for Black defendants, resulting in an increase in racial bias.

In all these cases, bias emerged not as a result of biased training data but as a result of complex and unintended interaction effects between algorithmic and human decision making. Social scientists in disciplines as varied as marketing, economics, and psychology will continue to document instances of this kind, where extensive testing, performance benchmarks, and other technical solutions will not suffice to prevent bias in AI outcomes.

Conclusions

The recent progress in the field of AI is the result of multiple trends. First, the ability to collect vast datasets about human behavior thanks to innovations like the Internet, where human behavior is easily trackable, and more recently the Internet-of-Things—even your vacuum cleaner today has sensors and an app. Second, the ability to store vast amounts of data due to improvements in memory storage and especially the building of a large cloud computing infrastructure. Third, the continued, non-linear trajectory in the power of computer chips over time (“Moore's Law”) and, specifically relevant to the domain of AI, developments in Graphics Processing Units like Nvidia's. While originally developed for gaming, these chips turned out to be ideally suited for the type of parallel processing that is required for the training of AI models. The benefits of greater computing power are compounded by the massive investments that companies are making into increasing the computational power available for training AI models. Last but not least, vast improvements in machine learning and related fields. Among many recent landmarks, the most important is the development of Transformer Models. Proposed just a few years ago by a research team at Google (Vaswani et al. 2017), this architecture is now powering all LLMs. To give a sense of the amount of academic attention, as of August 2024 the original paper has been cited over 30.000 times since the start of the year alone.

The confluence of these trends has given us AI systems that can appear to reason like a human expert and that can complete tasks needing sophisticated knowledge and decision-making skills. Listing these trends helps us foresee how the current turbulent, and arguably hyped, conditions are likely to develop. Even if the momentum in AI investments slows down, several of these trends are likely to persist for the foreseeable future. For example, fast progress is still being made in improving existing models and in developing new ways to solve problems. Moreover, just like the investments made during the “Dot.com bubble” of the early 2000s created of the infrastructure behind the Internet we know today, the investments being made in AI will likely pay dividends even if, in case of a crash, these dividends won’t necessarily accrue to those making the investments. While it is impossible to make firm predictions about the advancements in AI models that we are likely to see in the coming years, it seems safe to predict that models will keep getting better and that companies will keep learning how to best deploy them. In turn, it is safe to predict that the impact of AI on the field of marketing will continue to grow, either because these models are used to provide more value and new benefits to customers or because they are used by companies to make decisions or automate processes that impact customers (Puntoni and Wertenbroch 2024).

As AI systems become ubiquitous, it is crucial for marketers to realize that AI has ceased to be solely a technical topic for computer scientists and engineers. It is now also a social science topic, and in many important areas progress can only be made with the contribution of social scientists and via the application of insights about human behavior and psychology. There is much to be optimistic about the potential for AI technology to improve outcomes in crucial domains like healthcare and education, let alone the potential for economic growth and better standards of living. Ultimately, more intelligence must be a good thing. But like all other powerful technologies, responsible and accountable decision making will be necessary to minimize the risk of harm. To conclude with the question from Grewal, Guha, and Becker (2024) title, “is AI changing the world for better or for worse?” The answer is in our hands.