We are seeing exciting AI innovations that are transforming our lives and society. Companies are investing and adopting AI to improve their product solutions, business processes and to gain a competitive advantage. As advancements in AI are rapidly growing, societal expectations are growing as well. There are increasing scrutiny on what harms AI systems can cause if there are no guidelines or accountability enforced. As a result, governments are starting to regulate AI in response. On the other hand, data scientists, AI developers and decision-makers face the challenge of finding the right tools to enable them to analyze machine learning models for fairness, safety & reliability, explainability and accountability. In this session, we will explore new practical tools that enable data scientists and companies to incorporate in their Machine Learning lifecycles to build AI systems that are less harmful, more trustworthy and meet compliance requirements.

Recent advances and applications of language technology and artificial intelligence have enabled intelligent automation across various domains such as law, health care, FinTech, etc. Particularly for legal systems, AI based language models have recently been proposed to understand legal language and documents attempting to predict the judgment. While these models demonstrate acceptable performance on judgement prediction problems, they also carry encoded social biases learned from the training data. The concept of bias and fairness within machine learning models have been widely studied across NLP community, but most studies limit themselves to the Western contexts.

In this work, we present an initial investigation of fairness and bias in language models designed to understand legal documents from the Indian perspective. We highlight the presence of learnt algorithmic biases in InLegalBERT, a language model finetuned on legal documents in the Indian context. We show that InLegalBERT shows stereotypical preference in the axes of disparities such as Religion, Caste & Gender and anti-stereotypical nature in the case of the Region axis of disparity. On average, the bias shown by InLegalBERT is around 12.55% higher compared to a standard BERT model. Additionally, we highlight the research requirements in the direction of understanding bias in language models trained on Indian legal documents and its removal, which can potentially assist legal practitioners in future.

This presentation reflects on the research that is undertaken in the Academy of Finland funded, “Emotional experience of privacy and ethics in pervasive systems (BUGGED)” project. In the presentation we will look at how artificial intelligence (AI) is composed from the perspective of privacy. In the presentation I will discuss the nature of AI in light of the types of personal data that are used to inform machine learning system on humans as users. If AI could have an opinion on us based on how we are teaching it through our interests and habits, what would it be?

As we transition from traditional purpose-fit machine learning models to large general-purpose foundation models in consequential application domains, some requirements for safety and trustworthiness remain the same: fairness, robustness, explainability, uncertainty quantification, and transparency. But some possible risky behaviors are new: stereotyping, toxicity, hallucination, plagiarism, and prompt injection attacks. In this talk, we will overview these harms and discuss points of intervention throughout the foundation model development lifecycle to mitigate the harms.

Intelligent technologies are part of our everyday life – not only for adults but children as well. In this talk I will discuss ways for how we can try to understand their effect and role on children’s lives, arguing for a child-led perspective.

We are seeing exciting AI innovations that are transforming our lives and society Companies are investing and adopting AI to improve their product solutions, business processes and to gain a competitive advantage. As advancements in AI are rapidly growing, societal expectations are growing as well. There are increasing scrutiny on what harms AI systems can cause if there are no guidelines or accountability enforced. As a result, governments are starting to regulate AI in response. On the other hand, data scientists, AI developers and decision-makers face the challenge of finding the right tools to enable them to analyze machine learning models for fairness, safety & reliability, explainability and accountability. In this session, we will explore new practical tools that enable data scientists and companies to incorporate in their Machine Learning lifecycles to build AI systems that are less harmful, more trustworthy and meet compliance requirements.