The advance of artificial intelligence in recent years, driven by the launch of ChatGPT, has generated a mixture of admiration and fear. Activities previously considered exclusively human can now be performed by machines. This innovation, combined with market competitiveness and the search for convenience, has meant that artificial intelligences based on LLMs (Large Language Models) have been incorporated into our daily lives very quickly, becoming indispensable.
However, we need to be careful in using these new artificial intelligence tools because they have limitations that are not always so clear and can go unnoticed, so I am writing this article to point out characteristics that we need to know to use this technology safely, mainly motivated by conversations with several government and institutional relations professionals over these past few months.
The challenge of LLM-based artificial intelligences
A critical point is that LLMs are not capable of reasoning in a human-like manner [1]. A study indicated that, when subjected to tests requiring logical reasoning, the most advanced model got only 47% of the answers right, while a human being would easily reach around 96% [2].
In other words, this happens because the architecture of these new technologies are basically super complex computer software that can identify patterns invisible to us in gigantic masses of data, and with that, they gain an ability to predict the next step by extrapolation.
Although this response by extrapolation is useful in several applications, such as translating a text, in more delicate situations where there is not enough data, they bring incorrect answers, or at best, very generic ones.
This can represent a challenge for Government and Institutional Relations professionals, who, in many cases, need to work with multiple poorly correlated themes that, for example, range from tax matters to the deep technical issues of the industries they operate in, and in the daily rush, they may end up trusting the AI's response, which are usually written in a highly convincing but completely wrong way, without realizing that the AI does not have all the data to mold a quality response.
Therefore, it is essential to be careful when asking AI for recommendations or analyses involving the crossing of information, especially on themes that you do not fully master, or where there is a scarcity of data.
So, to illustrate my point, I did an experiment. I asked the ChatGPT 4o model to generate an image of a wine glass filled to the brim; however, you will quickly notice that AI does not know what the "brim of a glass" is, consequently, ChatGPT is incapable of generating an image of a wine glass full to the brim, because in the end, how many times do we see a picture of an overflowing wine glass compared to glasses half full? Imagine for regulatory objects that are much more complex, abstract, and with much rarer descriptions in the literature we find on the internet.
Does this mean that AIs are useless for Governmental and Institutional Relations?
FAR FROM IT! LLMs can be extremely useful in the day-to-day work of Government and Institutional Relations professionals, provided they are used for the right tasks involving pattern identification, such as:
Summarizing Information
AIs are highly effective in extracting relevant information from a text and creating concise and accurate summaries.
Formatting Content
If you need to organize an unstructured text, such as the processing of a legislative proposition or the minutes of a meeting, AI can help structure it in a clear and understandable way.
Correcting Grammar and Spelling
AI can improve the fluidity of the text, correcting grammatical and spelling errors with high precision.
Translating Content
AIs were originally developed to improve translation accuracy, becoming excellent tools for this purpose.
Searching for Information
Searching for information is an essential activity in the daily lives of Government and Institutional Relations professionals. AIs can optimize this process, searching for data based on keywords or semantic analysis. At Sigalei, for example, we have the Automatic Curation feature, which uses AI to make searches more accurate and efficient.
Comparing and Grouping Data
AI also excels in classifying and organizing information, identifying patterns, and facilitating the extraction of themes and tags.
Final Considerations
I hope this article has helped clarify how you can safely use AI in your day-to-day Government and Institutional Relations. LLMs are powerful tools but require judgment and knowledge to be used in the best possible way.
See you in the next article!
References:
[1] YAX, Nicolas; ANLLÓ, Hernan; PALMINTERI, Stefano. Studying and improving reasoning in humans and machines. 2023. Available at: https://doi.org/10.48550/arXiv.2309.12485. Accessed on: Mar 21, 2025.
[2] DUA, Dheeru et al. DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs. [S. l.]: arXiv, 2019. Available at: https://arxiv.org/abs/1903.00161. Accessed on: Mar 21, 2025.
