Exploring the potential of generative AI with scientific rigor, information protection, and transparency.
A tool that requires verification
Generative AI is particularly relevant due to its cross-disciplinary nature across scientific fields, functions, and purposes. Its non-deterministic characteristics are not due to problems or implementation options, but to the structural properties of the underlying models, which must be known by those who use them:
Scientific rigor requires that results be verified, that sources be consulted, and that the use of AI be declared.
The fluency of a response does not guarantee its veracity nor the existence of the cited references.
COMMON PRINCIPLES
Critically evaluate the results
Variability and non-reproducibility
Distinct models, trained on distinct corpora, produce distinct responses to the same question; the same model produces distinct responses at different times. Reproducibility, which is an essential requirement of scientific work, is not guaranteed.
Enviesamento
The results reflect the composition of the training data: an overrepresentation of certain languages, publishers, content producers, schools of thought, and normative frameworks; a underrepresentation of minority traditions and non-anglophone literature.
Conspiracy
The mechanisms of randomness that make responses flow and not repetitive produce, in a context that requires rigor, false content presented with the same certainty as the true content. The absence of nonexistent bibliographical references is the most frequent and most serious case.
Confidentiality
Most AI services are external to UFP. The use of personal data, health data, unpublished research results, or information subject to professional secrecy requires prior verification of the applicable legal, contractual, and institutional conditions.
Enemy content
It is technically possible to insert hidden instructions into documents, pages, or files that alter the behavior of a generative system that processes them. The defense against this risk is the literacy of those who operate the system and the verification of the result, not the trust in the tool.
Competence erosion
The systematic delegation of cognitive tasks to automated systems can hinder the acquisition of the skills that these tasks develop, which is particularly concerning in formative contexts for future researchers.
Access asymmetry
The free versions are limited in capacity, volume and timeliness. There is a real risk of reproducing, with these tools, the gap that has historically occurred in access to computer equipment.
Empathy
Generative systems respond in a way to gain the trust of those who use them, creating an empathy that can make it difficult to discern the quality of the information received and lead to dependency.
Support possibilities
Generative AI can support different phases of work. In each use, the foundation, validation, and scientific decisions remain under human responsibility.
Support in creating ideas
An AI system can generate ideas that can be used later, after their validity and originality have been verified, preventing plagiarism.
Support in laboratory simulation
An AI system can run simulations based on instructions, generating data, testing and analyzing results, and drawing the simulation itself. The correction of the simulation and the validation of the results must always be verified.
Support in the review of the literature
There are several AI systems that specialize in literature review; the results should all be confirmed regarding their existence and suitability for the topic.
Support in writing the state of the art
The generated texts must be verified, edited, and adapted to the document, application, or thesis, avoiding plagiarism or inconsistent or outdated text.
Support in programming
It is becoming increasingly common, in particular allowing users without training in this area to generate code for tasks that require it, including statistical processing and the use of AI, for example classifiers.
Support in data analysis
Data analysis by AI systems is becoming increasingly common and allows for considerable time savings, including in its preparation, cleaning, and discovery of features.
Support in data visualization and infographics generation
It improves the graphical quality of the data visualization, and inaccuracies, incorrect scales, or other errors should be checked. In particular, the manipulation of AI-generated images should be cautious, even with the aim of improving their quality.
Support in writing
In particular, in a non-native language, improving the understanding of the text and eliminating errors. In cases involving third parties, such as publishers or funders, it should be verified that they accept texts with parts generated by AI.
Declare how you used AI
In all these situations, the use of AI must be declared, specifying which tasks, which AI systems are used, what the original results are and what modifications are made. The use of AI when it is not authorized, or its failure to be declared even when it is authorized, constitutes a breach of integrity.
01
Responsibilities
In which phases or activities did you use AI?.
02
SYSTEMS
What AI systems did they use?.
03
RESULTS
What were the results originally generated?.
04
TRANSFORMS
What checks, editions, and modifications did you make?.
Other usage contexts
STUDENTS
Rules of use, levels of AI and academic integrity.
GUIDELINES FOR YOU
TEACHERS
Teaching, material preparation and responsibility in assessment.
GUIDELINES FOR YOU
Investigators
Good practices, limitations and transparency in research.
GUIDELINES FOR YOU
Contributors
Support for tasks, verification, and human mediation.