JTS Scholar Mansouri researches corporate ‘blah blah’

‘That’s a good question.’ Or: ‘Let me get back to you.’ Evasive or non-answers are everywhere. UG scientist Sasan Mansouri uses language models to detect and analyse evasive language used by listed companies. In this episode of JTS Scholars, he talks about his work.

Text: Jelle Posthuma

Mansouri is an engineer by training, but switched to economics as a young researcher. During his PhD at Goethe University Frankfurt, he researched the use of language models in the financial sector – at the time, a very niche area of research. ‘People said I’d be better off doing ‘real’ research. But I was lucky: during my PhD, the use of language models really took off.’

The focus of his research centres on how companies communicate with the market, for example during so-called earnings calls. During these meetings, listed companies explain their results (such as profit and turnover) to investors and analysts. ‘We used machine learning, to collect and process information coming from companies. My aim is to check the quality of the information. What is blabla, and what is real information.’

Non-answers

Mansouri cites his paper entitled ‘Let me get back to you’. The title is telling, as it deals with the prevalence of non-answers in corporate reporting. ‘When managers avoid bad news and give non-answers, uncertainty surrounding these companies increases. That is what our research shows. Shareholders can take out insurance on their shares, and the premiums for these policies rise when non-answers are given. The market therefore reacted negatively.’

The use of machine learning to detect evasive answers is not limited to businesses. ‘You see it with politicians, for example. When they’re going through a turbulent period in their political adventure, they tend to give more evasive answers to journalists,’ says Mansouri with a touch of understatement. He also mentions the questioning of Facebook CEO Mark Zuckerberg in the US Senate. ‘His answers were full of remarks such as “that’s a good question” or “I don’t know the details”. These are typical statements people use when they want to avoid answering.’

AI text? No donation

Recently, Mansouri has broadened his research into how people perceive artificial intelligence. He describes AI as a double-edged sword: it brings efficiency, but people are also developing ‘algorithm aversion’. In a recent study, he examined GoFundMe, a crowdfunding platform. His research shows that people frequently use language models such as ChatGPT to write the text for their crowdfunding campaigns.

Mansouri: ‘The use of AI for crowdfunding texts is in itself beneficial for society: people have access to better-written texts, making it easier for them to ask for donations. But we also observed another notable outcome: when donors realize that a text has been generated by ChatGPT, they are less likely to donate. Why is that? We want people to make an effort when they ask for money. That’s simply how our brains work.’

Education and Scholarship

Language models, prompts and other AI-related topics also play an important role in his educational activities, he continues. ‘Several colleagues are worried that students might write their theses using ChatGPT. I’m not bothered by that, because in my courses students are actually required to use AI. Many of them are now working in industry, where they make extensive use of language models.’ 

As a JTS Scholar, Mansouri aims to promote collaboration not only between academic disciplines, but also between the university and the corporate world. ‘Take European legislation on AI, for example. We could organize workshops where judges, researchers and business representatives can discuss the implications.’