Can generative AI increase inequalities in the workplace?

Less than two years after ChatGPT was launched into the public sphere by the US research organization OpenAI, its use has become ubiquitous in offices and classrooms.  

Generative AI (GenAI) systems are increasingly used in favor of basic web browsers thanks to their ability to understand and process natural language tasks. Large language model (LLM) systems provide elaborate answers to specific questions, saving users the chore of searching through pages of web results skewed by SEO, paid-for placements and the constantly changing preferences of search engines. 

The study suggests that GenAI's impact on verbal interactions is more beneficial for high performers

Where previous waves of automation primarily affected blue-collar work, such as the use of robotics in manufacturing, the bespoke knowledge offered by GenAI impacts knowledge-intensive skills. Easy access to customer service-based information or answers to questions regarding legislation and regulations creates efficiencies in roles typically carried out by people with specialist knowledge or training.  

But while early research on the impact of GenAI shows that this access to written answers provides a boost to low-performing professionals and evens out productivity levels, a new study from Antonio Roldán, Director of the Center for Economic Policy (EsadeEcPol), suggests the technology’s impact on interactions implying the performance of complex tasks is more beneficial to high performers. Rather than smoothing out performance, he suggests, the use of GenAI could increase inequalities in the workplace. 

The full working paper can be found in EsadeEcPol’s website. 

Crucial skills for high earners

High earners, managers and leaders in all sectors tend to possess strong cognitive skills in areas such as social intuition, critical thinking, argumentative agility and the capacity to persuade others. The research from Roldán is one of the few analyzing the impact of GenAI on these cognitive skills, particularly on in-person social interactions.  

In a three-day randomized control trial carried out at Esade’s DecisionLab and CUNEF University, Roldán assessed the debating performance of 142 law, economics or business undergraduate students. Information about each participant, including socio-demographic characteristics, academic achievements, debating experience, native language and previous experience using ChatGPT, was collected at baseline.  

Randomization took place in three phases: the treatment or control group participants were assigned to; the partners they were given; and the debating position assigned in each round. Each debate in the series had four rounds related to a different policy topic, with the randomization meaning that students could be debating either a treatment or control student, and potentially debating the same person twice.  

Debating the impact of GenAI

In the first round of debates, ChatGPT was prohibited for everyone. Treatment students were then given a 20-minute training session on ChatGPT, while control students did not receive such training. After receiving the topics for the following rounds, all students were given 45 minutes to prepare. AI tools were strictly prohibited for the control group. 

In the following debates, the treatment group was allowed to use ChatGPT throughout the contest, while the control group was only allowed to use conventional internet resources. Vouchers were offered to the winning debaters to incentivize performance. The debates were recorded and sent to independent expert judges, who were not informed of the experiment. 

Stronger performers are better at interpreting the information provided by GenAI, rather than merely repeating it

Overall, the independently scored results revealed that ChatGPT had a positive but not significant effect on overall debating performance. However, the crucial finding came when the academic ability of students was taken into account: it revealed that the use of ChatGPT had a significantly more positive benefit on the students who were recorded as having a higher academic ability. The students who had lower levels of academic ability did not benefit at all from using ChatGPT to inform their debate. 

A strong argument

The evaluation also revealed that the high-skilled students who used ChatGPT had higher scores in four out of five areas: credibility, superiority, refutation and rhetoric. But the students with lower ability who also had access to GenAI displayed no positive benefits in the same categories.  

While the use of ChatGPT by the students with lower academic ability did help to improve the clarity of their presentations, it didn’t improve their ability to decipher the content the technology presented as being either good or bad. This finding suggests that stronger performers have a better ability to interpret the information they’re provided, rather than simply repeating it. 

In a real-life setting, simply replicating (aka “copying and pasting”) answers provided by GenAI may help to improve performance of people with lower ex ante abilities, particularly on written tasks. But in social environments where in-person interactions and conversations rely on the ability to persuade and influence, the knowledge extracted from ChatGPT proves much more beneficial to those who already possess these skills to a higher degree. 

If the results of the research can be replicated in other contexts, such as negotiation or sales, the implications of GenAI on work inequalities can be better understood.