Tech Trend 05: Responsible AI: Building a sustainable framework

GenAI risks

Risks associated with GenAI, especially in LLMs, include model-induced hallucinations, ownership, and technological vulnerabilities such as data breaches, as well as compliance challenges arising from biased and toxic responses. There have been recent examples of authors being credited with non-existent articles and fake legal cases have been cited by GenAI tools. Inadequate control over LLMs trained on confidential data can lead to data breaches, which, according to a recent EY survey, is the single biggest hurdle to GenAI adoption in India.

Toxic information and data poisoning, intensified by insufficient data quality controls and inadequate cyber and privacy safeguards, adds another layer of complexity, diminishing the reliability of GenAI outputs and jeopardizing informed decision-making. Additionally, the broader spectrum of technology risks of deepfakes to facilitate crime, fabricate evidence and erode trust necessitates proactive measures for secure GenAI adoption.

 Potential intellectual property rights (IPR) violations during content and product creation also raise legal and ethical questions about the origin and ownership of generated work.

Other risks include:

• Bias and discrimination
• Misuse of personal data
• Explainability
• Misuse of personal data
• Predictability
• Employee experimentation  
• Unreliable outputs
• Limitations of knowledge
• Evolving regulation
• Legal risks

Building guardrails against risks  

To capitalize on the competitive advantage and drive business, GenAI models and solutions are implementing safety guardrails to build more trust. The tech giants have created the frontier model forum. Its objectives include advancing AI safety research, identifying best practices, and collaborating with policymakers, academics, civil society, and companies. The forum aims to ensure that AI developments are handled responsibly and deployed responsibly. A model’s performance is evaluated and measured against designated test sets and quality considerations. Model monitoring and performance insights are leveraged to maintain high quality standards. 

With various models evolving, implementing robust data governance policies that comply with privacy regulations will help companies mitigate risks. There are seven key domains to establish a robust framework and governance processes that align with industry-leading standards of responsible AI. These are business resiliency, security operations, model design development, governance, identity and access management, data management, and model security.

Regulations so far

The growing need for AI regulations has resulted in a complex and diverse array of global regulations to navigate AI risks. China has been a forerunner in designing a new law that focuses on algorithm recommendations, including generative and synthetic algorithms. EU’s AI Act is the first major legislation to stress on a risk-based approach. It categorizes AI applications into different risk levels, ranging from unacceptable to low and with high-risk applications subject to more stringent requirements. The law prohibits AI systems that pose an ‘unacceptable risk,’ such as those utilizing biometric data to deduce sensitive traits like individuals' sexual orientation. Developers of high-risk applications, such as the ones that use AI in recruitment and law enforcement, must show that their models are safe and transparent.

While India is developing its own law, the Ministry of Electronics and IT (Meity) has recently issued an advisory to AI platforms to take permission before launching AI products in the country. The government has asked intermediaries to tag any potentially deceptive content with distinctive metadata or identifiers to trace its source and thus aid in tracking misinformation or deepfakes and its creators.  Meanwhile, in the US, the AI Executive Order directs agencies to move toward adoption with safeguards in place. 

The G20 nations have also committed to promoting responsible AI in achieving the Sustainable Development Goals (SDGs). Additionally, 28 countries, including India, China, the US, and the UK, signed the Bletchley Declaration AI summit, pledging to address AI risks and collaborate on safety research. Also, HITRUST has released the latest version of the Common Security Framework comprising areas specifically addressing AI risk management. Along with the global agreements, Responsible AI needs local regulations as well.