Abstract
The study of data quality in crowdsourcing campaigns is currently a prominent research topic, given the diverse range of participants involved. A potential solution to enhancing data quality processes in crowdsourcing is cognitive personalization, which involves appropriately adapting or assigning tasks based on a crowd worker's cognitive profile. There are two common methods for assessing a crowd worker's cognitive profile: administering online cognitive tests, and inferring behavior from task fingerprinting based on user interaction log events. This article presents the findings of a study that investigated the complementarity of both approaches in a microtask scenario, focusing on personalizing task design. The study involved 134 unique crowd workers recruited from a crowdsourcing marketplace. The main objective was to examine how the administration of cognitive ability tests can be used to allocate crowd workers to microtasks with varying levels of difficulty, including the development of a deep learning model. Another goal was to investigate if task fingerprinting can be used to allocate crowd workers to different microtasks in a personalized manner. The results indicated that both objectives were accomplished, validating the usage of cognitive tests and task fingerprinting as effective mechanisms for microtask personalization, including the development of a deep learning model with 95% accuracy in predicting the accuracy of the microtasks. While we achieved an accuracy of 95%, it is important to note that the small dataset size may have limited the model's performance.

Abstract

Abstract
Large Language Models (LLMs) are widely used today in virtual assistants and content generation. However, there are suspicions that LLMs present confirmation bias, responding in a way that reinforces beliefs or assumptions embedded in users' questions, which can lead to erroneous decision-making, especially in sensitive areas such as healthcare. The objective of this research is to determine how often and under what conditions LLMs present confirmation bias and to identify the causes of this effect. The methodology involves conducting an experiment in which 52 biased healthcare questions are presented to 10 of the most popular models and analyzing whether their responses were biased. This work proves with statistical power the behavior of confirmation bias. We show that confirmation bias in LLMs occurs in all LLMs with a frequency of 20% to 60% of the occasions. The evidence suggests that the bias arises from the training database, the Transformer architecture itself, and the instructions in the fine-tuning phase by the companies behind the LLMs. This research explores pathways for the development of trustworthy LLMs.