Everyone does it, but no one admits it. This refers to the use of generative artificial intelligence (GenAI). However, the introductory sentence is not entirely accurate, as many media outlets and other companies already disclose when their published content has been created with the help of AI. But the actual number is likely to be higher, says Lino Ballof, research assistant at Prof. Michael Stich's Chair of Accounting at TUM Campus Heilbronn.
Ballof knows what he is talking about, because together with Luca Fober, a graduate of the Management & Technology program at TUM, he has researched how many companies use Large Language Models like ChatGPT and similar tools to create text. More specifically, the two looked into whether German-listed companies use AI tools to create their annual reports. To this end, they selected a sample of 1,225 reports from all available annual reports of the 40 DAX, 50 MDAX, and 70 SDAX companies from 2017 to 2024 and subjected them to a double analysis.
Exponential increase
The first step was a word analysis: Ballof and Fober examined whether certain terms were used significantly more frequently after 2022 – the year ChatGPT was released – than before. They focused on words that had been identified in earlier studies as indicators of possible AI use. The result: “We observed an exponential increase in words such as ‘crucial,’ ‘enhancing,’ and ‘meticulously’.” This was an initial indication that GenAI was indeed at play. However, word analysis has its weaknesses: as valuable as it is for large amounts of text, it is not very meaningful for individual documents or sentences.
Ballof and Fober therefore analyzed the reports in a second step using the AI detection software GPTZero. “The tool can already assess the probability of individual sentences being generated by AI. But it also works on an aggregated level, i.e., with large amounts of text,” he explains. The use of GPTZero confirmed the results of the word analysis: The proportion of sentences that were written with at least a 90 percent probability of AI assistance rose from 0.44 percent to 2.77 percent between 2017 and 2024 – more than sixfold.
More efficiency, less transparency
“The absolute proportion seems small, but the growth rate is enormous. In addition, annual reports are relatively inflexible in their structure: many sentences are carried over from year to year. That's why I assume that the AI-generated proportion of newly added sentences is even higher,“ says Ballof. A third aspect also catches attention:” AI is often used as a support tool: you get initial ideas, adapt them, and refine them. Tools such as GPTZero often don't recognize this because they are designed to avoid false hits.” Here, Ballof draws a connection to the previously mentioned unreported figures.
But why do companies use AI for text generation? They expect more efficient work processes and faster results. They hope for higher quality because there are fewer grammatical and typing errors. Last but not least, GenAI makes it easier for small companies with limited resources to appear more professional: they don't have to hire additional staff to write their annual reports. But Ballof also mentions risks: “Everyone is familiar with hallucinations that AI tools regularly produce. You have to be extremely careful here.” Another disadvantage, in his view, is the current lack of transparency: it often remains uncertain whether a person or a company has used AI or not.
“Human in the loop” as the ideal solution
What are the possible solutions? “In our paper, we advocated the human-in-the-loop approach. Put simply, this means that AI provides support, but humans are responsible.
AI should only be used as a supporting tool and not as a co-author or even main author. Above all, a responsible person must take responsibility for any errors when adopting AI content,” says the scientist. He also advocates stronger regulation. For him, the EU AI Act is an important step in the right direction. At the same time, he warns against overregulation, which would slow down the momentum and efficiency of the new technology and thus potential productivity gains.
If all this is guaranteed, Ballof believes that annual reports – and other corporate texts such as press releases, social media posts, website texts, and product descriptions – will be created entirely by AI in the future. This could happen in three to seven years—if the technology continues to develop as it has done so far and the regulatory framework can keep pace. From Ballof's point of view, this would be a considerable step forward: “If this happens in the right structure, then I do believe that repetitive work will generally be automated and people will have time to focus on more enjoyable tasks.”
Ballof knows what he is talking about, because together with Luca Fober, a graduate of the Management & Technology program at TUM, he has researched how many companies use Large Language Models like ChatGPT and similar tools to create text. More specifically, the two looked into whether German-listed companies use AI tools to create their annual reports. To this end, they selected a sample of 1,225 reports from all available annual reports of the 40 DAX, 50 MDAX, and 70 SDAX companies from 2017 to 2024 and subjected them to a double analysis.
Exponential increase
The first step was a word analysis: Ballof and Fober examined whether certain terms were used significantly more frequently after 2022 – the year ChatGPT was released – than before. They focused on words that had been identified in earlier studies as indicators of possible AI use. The result: “We observed an exponential increase in words such as ‘crucial,’ ‘enhancing,’ and ‘meticulously’.” This was an initial indication that GenAI was indeed at play. However, word analysis has its weaknesses: as valuable as it is for large amounts of text, it is not very meaningful for individual documents or sentences.
Ballof and Fober therefore analyzed the reports in a second step using the AI detection software GPTZero. “The tool can already assess the probability of individual sentences being generated by AI. But it also works on an aggregated level, i.e., with large amounts of text,” he explains. The use of GPTZero confirmed the results of the word analysis: The proportion of sentences that were written with at least a 90 percent probability of AI assistance rose from 0.44 percent to 2.77 percent between 2017 and 2024 – more than sixfold.
More efficiency, less transparency
“The absolute proportion seems small, but the growth rate is enormous. In addition, annual reports are relatively inflexible in their structure: many sentences are carried over from year to year. That's why I assume that the AI-generated proportion of newly added sentences is even higher,“ says Ballof. A third aspect also catches attention:” AI is often used as a support tool: you get initial ideas, adapt them, and refine them. Tools such as GPTZero often don't recognize this because they are designed to avoid false hits.” Here, Ballof draws a connection to the previously mentioned unreported figures.
But why do companies use AI for text generation? They expect more efficient work processes and faster results. They hope for higher quality because there are fewer grammatical and typing errors. Last but not least, GenAI makes it easier for small companies with limited resources to appear more professional: they don't have to hire additional staff to write their annual reports. But Ballof also mentions risks: “Everyone is familiar with hallucinations that AI tools regularly produce. You have to be extremely careful here.” Another disadvantage, in his view, is the current lack of transparency: it often remains uncertain whether a person or a company has used AI or not.
“Human in the loop” as the ideal solution
What are the possible solutions? “In our paper, we advocated the human-in-the-loop approach. Put simply, this means that AI provides support, but humans are responsible.
AI should only be used as a supporting tool and not as a co-author or even main author. Above all, a responsible person must take responsibility for any errors when adopting AI content,” says the scientist. He also advocates stronger regulation. For him, the EU AI Act is an important step in the right direction. At the same time, he warns against overregulation, which would slow down the momentum and efficiency of the new technology and thus potential productivity gains.
If all this is guaranteed, Ballof believes that annual reports – and other corporate texts such as press releases, social media posts, website texts, and product descriptions – will be created entirely by AI in the future. This could happen in three to seven years—if the technology continues to develop as it has done so far and the regulatory framework can keep pace. From Ballof's point of view, this would be a considerable step forward: “If this happens in the right structure, then I do believe that repetitive work will generally be automated and people will have time to focus on more enjoyable tasks.”
