type:: [[REVIEWS/PhDThesis]]
tags:: 
year:: 2024
venue::  
full-title:: Model-Driven Automated Testing of Ethical Requirements in Generative AI
date-start:: [[14-06-2024]] - 18:16
date-submitted::
external-links::
status:: [[DONE]] 
deadline-submission:: 
file:: [[@Model-Driven Automated Testing of Ethical Requirements in Generative AI]] 
parent::
todoist::

	- ### [[Highlights]]
		-
		- # Annotazioni
			- (14/6/2024, 18:27:49)
			- - “Ethical Requirements in Generative AI” ([García, p. 1](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=1&annotation=R4NZAFC8)) #ffd400
			- *The research is very relevant and I encourage the student to engage in this very interesting and timely topic. One general comment that I have, I recommend to distinguish the concepts bias, fairness, and ethics distinguished. I'm recommending this because in the proposal sometimes such concepts are used interchangeably. Ethics is a very general concept and not all its constituting elements can be automated in my opinion.*
			- - “Generative artificial intelligence (GenAI) systems, and specifically large language models (LLMs), are artificial intelligence components capable of generating new content, such as text, images, music, or other forms of data, based on patterns learned from observed data” ([García, p. 4](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=4&annotation=GIITRPVP)) #5fb236
			- - “However, due to their nature, Generative AI systems and their underlying LLMs can face significant challenges related to fairness and balanced representation in the generated responses” ([García, p. 4](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=4&annotation=IYUF75ZI)) #5fb236
			- - “Generative AI models learn from large datasets, which may contain biases that the model may replicate or amplify in its output [4, 7, 21, 40, 41, 47].” ([García, p. 4](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=4&annotation=K5FVSIA9)) #a28ae5
			- - “Addressing biased responses in generative AI is crucial to prevent the generation of unwanted discriminatory and/or harmful content.” ([García, p. 4](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=4&annotation=93QLAFZZ)) #a28ae5
			- - “To this end, it is necessary to implement techniques to detect and mitigate biases in a model’s responses” ([García, p. 4](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=4&annotation=PFFN2EBH)) #a28ae5
			- - “generative AI models are probabilistic algorithms for identifying patterns and generating new similar data” ([García, p. 4](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=4&annotation=DM4L95NF)) #5fb236
			- - “integrate a systematic ethical assessment within the development lifecycle of LLM-based applications,” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=Q7WPRHLH)) #a28ae5
			- - “from its creation to the continuous monitoring after deployment” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=HNABRQ8E)) #a28ae5
			- - “judgements or allusions that are detrimental to a collective or an individual, selecting the best one for the most pressing ethical requirements in the project” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=W2G5C265)) #a28ae5
			- - “understand the state of the art of the assessment of ethical requirements in the field of AI ethics, with a special focus on bias detection in generative AI text-to-text models.” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=72KHLALG)) #a28ae5
			- - “establish their ethical requirements model and evaluate the corresponding lack of bias in candidate textto-text models” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=X8DK6FED)) #a28ae5
			- - “defining ethical requirements that automatically generates and executes a series of test cases in order to assess whether a target generative AI model satisfies such requirements, and provides informative and actionable feedback to stakeholder” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=TENANYYF)) #a28ae5
			- - “Design and Creation [38].” ([García, p. 5](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=5&annotation=4YQ8LGE2)) #a28ae5
			- - “model ethical principles aimed at integrating them into the software development proces” ([García, p. 9](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=9&annotation=7UE2G897)) #a28ae5
			- - “facilitate describing ethical requirements, but to provide a systematic method to automatically and efficiently assess them on concrete LLMs, driven by the user’s priorities and most pressing concerns.” ([García, p. 9](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=9&annotation=V5DQAZP5)) #a28ae5
			- - “Likewise, Dhamala et al. [14] and Alnegheimish et al. [1] use partial sentences to evaluate the inclusion of gendered pronouns and the generation of text, which may leads to undesired skewed representations.” ([García, p. 9](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=9&annotation=MSMF5UEM)) #5fb236
			- - “We will contribute broader coverage of biases and a more versatile detection method in addition to sentence completion.” ([García, p. 9](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=9&annotation=GJYAWDQM)) #ffd400
			- *Which one? Can you list or prioritize some of them? The topic is very broad and it is necessary to start with some concrete case to generalize afterward.*
			- - “we want to include further strategies that may reveal the existence of otherwise hidden bias, starting from a model-friven and platform-independent approach.” ([García, p. 10](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=10&annotation=UI2ERG8K)) #a28ae5
			- - “Our framework should be able to allow users to specify how an ethical requirement has to be assessed.” ([García, p. 10](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=10&annotation=45AVVAA7)) #e56eee
			- - “continuous ethical assessment throughout the whole development process” ([García, p. 10](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=10&annotation=ZHD7P8XU)) #e56eee
			- - “ethical requirements on a tailored set of particular ethical concerns at the modeling level” ([García, p. 10](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=10&annotation=8G9M3N72)) #a28ae5
			- - “traceability mechanism to relate the testing results with the initial ethical requirements, in order to assess the fulfillment of the latter by the AI-based software system.” ([García, p. 10](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=10&annotation=XPJ9Y8JI)) #ffd400
			- *How is it possible to that when using blackbox LLMs?*
			- - “ethical requirements in the field of AI ethic” ([García, p. 11](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=11&annotation=T5JTQJQL)) #a28ae5
			- - “gaps between the ethical principles and contribute a mechanism for their operationalization within the software development industry.” ([García, p. 11](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=11&annotation=MB8XWSHN)) #a28ae5
			- - “Facilitate a systematic ethical assessment of AI-based software” ([García, p. 11](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=11&annotation=NTFEAWU9)) #a28ae5
			- - “elaborate an ethical requirements model that describes the individual ethical concerns that must be addressed for their specific context by themselves.” ([García, p. 11](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=11&annotation=M9782KTW)) #a28ae5
			- - “define the test scenarios that are the best fit for their needs” ([García, p. 11](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=11&annotation=NUE2KNCT)) #a28ae5
			- - “end-to-end traceability between the ethical requirements model and the testing results” ([García, p. 12](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=12&annotation=S457CXGK)) #ffd400
			- *The feasability of this is not clear.*
			- - “Ethical requirements model and test scenario definition via DSL” ([García, p. 15](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=15&annotation=LQ772YWK)) #ffd400
			- - “Researching information systems and computing” ([García, p. 19](zotero://select/library/items/MQDWFTZV)) ([pdf](zotero://open-pdf/library/items/R6AWAPNX?page=19&annotation=MUSGU5WP)) #a28ae5
	- ### [[Comments]]
		- The research on ethical requirements in generative AI is timely and highly relevant. It is essential for the student to delve into this subject, particularly given the growing influence of generative AI systems and large language models (LLMs) in various domains. However, it is crucial to differentiate between bias, fairness, and ethics, as they are sometimes used interchangeably in the proposal. Ethics is a broad concept encompassing various elements, not all of which can be automated. Bias and fairness, on the other hand, are more specific and can be more directly addressed through technical languages and tools. Another point of attention is the feasibility of achieving end-to-end traceability between ethical requirements and testing results. Currently, such feasibility is unclear, especially because language models are black-box components, making it challenging to map or link given inputs with generated outputs. Overall, it is a good research plan!
	- ### [[REVIEWS/Notes]]
	- ### YELLOW CONCERNS
	  background-color:: yellow
		- {{query (and [[ffd400]] [[PhD_Research_Plan___Sergio_Morales]] )}}
		  query-table:: false
	- ### ❓️Questions
		- {{query (and [[question]] [[PhD_Research_Plan___Sergio_Morales]] )[[question]]}}
		  query-table:: true
		  query-properties:: [:block]