In this GitHub you can find detailed information about the paper "Guidelines to Develop Trustworthy Conversational Agents for Children" presented at Ethicomp 2022, Turku, Finland on July 27th of 2022.
In the paper we propose some guidelines to develop trustworthy Conversational Agents (CAs) for children. These guidelines were adapted from other ethical guidelines using the Delphi method with a set of experts from different disciplines (Computer Science, Children's rights and AI ethics).
As previous guidelines we considered the Assesment List for Trustworthy Artificial Intelligence (ALTAI) from the HLEG of the European Commission, and the Policy guidance on AI for Children from UNICEF. In the document UNICEF0 we related the sub-requirements from the 9 requirements of the Guidance on AI for Children (rows) to the 7 requirements from ALTAI (columns).
The document UNICEF1 takes the same matrix of UNICEF1 and changes the keywords for stars for visualization. One/Two/Three stars with One/Two/Three linked sub-requirements.
We saw that the Policy guidance on AI for children is highly oriented to politics, and it has a lot in common with the ALTAI document that covers all the non-political points. As the target of our paper was developer's oriented, we followed with the ALTAI document as a reference.
Therefore, we needed to adapt ALTAI self-evaluation questions from AI to CAs and from Humans to Children. In order to do this, we used the Delphi method with a questionnaire to evaluate the risk-level (Likelihood, Impact) of every question adapted to our specific case (CAs, Children). You can find the final answers to the questionnaires from the 4 experts in documents: ALTAI(R1), ALTAI(R2), ALTAI(R3), ALTAI(R4).
By mapping our expert's answers to 1,2,3, and using the measures: Risk = Likelihood x Impact, and Total Risk = CAs Risk x Children's Risks; we computed the Total risk of every question at the attached document: ALTAI Risk.
In order to better understand the obtained risk values, you can have a look to the Partial (for the CAs Risk and the Children Risk) and the Total risk assessment matrices at the document Risk Matrix.
In the document: ALTAI risk order, you will find all the ALTAI questions ordered by Total Risk.
Finally, we use the ALTAI Risk document to group the risk of the ALTAI answers, by the 7 Requirements of ALTAI and its subsections. This allows us to see what sections have higher risk. Results can be found in the document Risk Summary.
A summary of all the experts comments can be seen at the Experts Comments document. Rows indicate what expert did the comment (R1, R2, R3 or R4), and columns indicate the requirement in which the comment was written.
In the document Thematic Analysis, you can find the result of our thematic analysis as columns: Involve children as stakeholders, AI awareness, Tranparency, Risk management and Age appropriate behaviour; mapped to the 7 requirements from ALTAI.