LLM Ethics

Download the document HERE

 

 

Consciously or unconsciously, virtually all internet users employ some form of artificial intelligence (AI) in their daily lives, particularly systems based on language models. For example, language models are used when writing a text message or email, and the "autocomplete" tool suggests how to finish the sentence. Language models are also used to support academic activities, such as automated translation systems or programs that detect grammatical errors, among others. Furthermore, all types of web platforms and mobile applications now provide access to AI tools for generating text and audiovisual content.

 

However, some uses of AI are risky; therefore, their use must be informed, transparent, ethical, and responsible. AI tools are not suitable for every type of activity, and certain uses can be counterproductive to the research process. Some uses of AI can pose risks to both users and third parties. Furthermore, it is essential to be aware of the ethical, environmental, and human rights implications associated with the use of these tools.

 

In this context, the objective of this section is to promote the responsible and ethical use of these technologies in learning processes and professional life, as well as to prevent violations of academic integrity and situations in which authors, consciously or unconsciously, engage in academic dishonesty.

 

The use of AI as a support tool for carrying out different learning activities is permitted. The parameters for the use of these tools, described below, distinguish between “low-risk” and “high-risk” tools, according to the risks that the respective AI poses to the pedagogical process, users, and third parties.

 

• Rules for the use of low-risk AI. Low-risk AI tools may be used freely; that is, those that allow for the correction or revision of content generated by authors or those that allow them to collect and process data. For example, grammar checking, translation, audio-to-text transcription, and internet search tools, among others. It is recommended that when such tools allow for significant tasks (e.g., translation), their use be indicated in the corresponding section (e.g., methodology).
• Rules for the use of high-risk AI. AI tools used to generate synthetic text are considered high-risk. This includes LLMs accessible through chatbots, such as ChatGPT, Bard, Bing, and Claude, etc. AI-generated content may only be included in individual and group submissions when it meets four requirements:
  
  1. Informed use. Prior to use, authors must investigate who or which company developed the tool, how it was developed, how it works, what functions it can perform, its terms of use, what the company does with the information it collects from the user, and what limitations and/or risks it presents to the user and/or third parties.
  2. Transparent use. Authors must clearly and explicitly state which tool they used and how they used it in the methodology section or any other section they deem relevant. This is a requirement currently requested by scientific journals and publishers, such as Nature and Cambridge University Press.
  3. Ethical Use: It must be possible to distinguish what was written or produced directly by the authors and what was generated by an AI tool. Regarding this point, general citation rules apply, for example, using quotation marks if including verbatim excerpts, in addition to the corresponding reference. To cite synthetic text produced by LLMs, the rules suggested by the APA style team may be consulted. Violations of this rule, particularly regarding the second requirement, will be addressed using an approach similar to that applied in cases of plagiarism. Furthermore, ChatGPT is not a person and therefore cannot be considered a co-author. Language models like ChatGPT do not meet the criteria for attributing “authorship.”
  4. Responsible Use. It is recommended that the use of these AI tools be limited to the early stages of research, to inspire or suggest directions, not to produce content that will later be included in the original manuscripts. In any case, if authors choose to use information obtained through an LLM, they must demonstrate that they rigorously verified this information with other reliable sources, given that these systems tend to offer inaccurate, erroneous, and false information.

• UsoPlausible uses of LLMs:
  
  1. Exploring New topics and inspiration; that is, in the early stages of research.
  2. Automating basic tasks. Supporting exercises that involve recognizing and/or reproducing patterns in texts, for example: translating, summarizing, modifying the tone and style of the text, correcting spelling and grammar, audio-to-text conversion, programming, etc., although the result is not entirely reliable and the LLM output should always be carefully reviewed.
  3. Trivial writing. Supporting minor writing tasks.
  4. Sparring. Testing arguments or ideas in interaction with the chatbot.