top of page

Responsible Enterprise AI: Why LLMs flatter to deceive (Part Two)

Writer's picture: Ram BalaRam Bala

Elements of Responsible AI
Elements of Responsible AI


Fundamental conflicts in AI

The Open AI leadership drama has brought the fissures in AI into full view. On the one hand, AI doomers want a complete moratorium on AI research and development until we have government regulated safeguards in place. On the other hand, techno-optimists clearly believe in the power of AI to create abundance and prosperity. Our research (and research conducted by others over the years) has shown that while there are great many productivity benefits of AI to organizations and individuals, if the technology malfunctions due to either malicious intent or just bad data or bad management, the potential risks could go beyond those organizations and individual users to society at large. Risks to individuals and organizations range from safety and security to reputation and financial ruin as well as legal problems.


Our position at Samvid is one of cautious optimism. We strongly believe in the power of AI to do good, and view responsible deployment as key to ensuring reliable enterprise-wide adoption and the benefits that this will eventually bring.


But what do we mean by Responsible AI? In this post, we articulate our thoughts on the different dimensions of responsible AI and what that means for building such systems.


What is Responsible AI?

Responsible AI constitutes an approach that encompasses the ethical and legal aspects in the creation and utilization of artificial intelligence (AI) systems. At Samvid, we are aligned with this important value and are focused on crafting, evaluating, and rolling out AI systems with a focus on ensuring three major concepts - Transparency, Traceability and Trustworthiness


Transparency

Stories abound of Large Language Models (LLMs) displaying hallucinations that fabricate information and distort facts. While in many consumer use cases, this may be harmless and in fact even necessary when creating content such as stories, this can have deleterious effects in a professional context. A widely known example is the case of a lawyer who used ChatGPT to create a legal brief, which had fictitious case references. Although this issue was caught by the presiding judge, it resulted in punitive measures for the lawyer who submitted the brief, not to mention the damage of trust to the concept of AI use in the legal arena. However, this example is very clear in highlighting the role of transparency in enterprise AI use. Businesses deal with a lot of data that needs to be verifiably true in order to be usable. Therefore, any responsible enterprise AI solution has to enable transparency by drawing a clear line from its user output to the proprietary and public documents it sources its information from.


Traceability

Despite their wide use, LLMs remain enigmatic black box AI systems. Transformer models like ChatGPT, DALL-E, and Bard comprise of deep neural networks, which are built on the fundamental idea that the brain is a machine that can, among other things, focus on identifying trends and matching patterns. Once exposed to sufficient examples of a coffee mug, individuals quickly develop the ability to recognize a coffee mug in a store or at a cafe. Researchers still can’t fully explain how precisely the brain makes that connection. Similarly, even with a complete list of input variables, deep learning systems are convoluted entities of complex functions where the interrelationship between variables leading to a final prediction is often not easily explainable. While this might be fine if the goal of the AI system is simply a classification task such as separating pictures of cats from dogs, managerial decision makers have to contend with defending their decisions in front of multiple stakeholders. Thus, any AI system they use for decision making has to not only give them the right information but also explain how this information was arrived at starting from the original data provided. This goes beyond transparency into uncovering the connections between different forms of data and how they affect decisions.


At Samvid, the autonomous intelligent systems we are designing will employ Neeyums that utilize explainable AI techniques rooted in causal inference from input parameters, data visualizations and domain expertise. Our goal is to back recommendations with a complete trace - which part of the data was used, how features were engineered, what model was used, and why a certain prediction was made. For example, a user query such as “estimate shipping cost” should not only point to the data sources used for this forecast but also justify why a particular forecasting model was used from a family of models that might potentially fit the data.


Trustworthiness

Ultimately, the true test of Enterprise AI is adoption and we know that customer trust is key to such adoption. Customer trust is the faith a consumer has in a product and / or company. This faith is usually a consequence of robust, reliable and repeatable experiences. LLMs by themselves do not provide such reliability and repeatability guarantees. Therefore, using them out-of-the-box is not a viable option for enterprises.


At Samvid, we are building reasoning systems that employ domain specific “Tree of Thought” approaches that verify every step and ensure robustness of output. In addition, we employ an iterative self-improving process that evaluates every outcome against the achievement of user objectives, typically captured by a reward function. This reward function is further adjusted through a feedback mechanism that transmits user trust metrics back to our AI system.


Much has been written about existential AI risk. However, the immediate reality for most enterprises is the responsible deployment of AI within the walls of their own organizations with a view to increasing productivity and better decision making. At Samvid, we have arrived at a concrete definition of responsibility in AI based on the three pillars of Transparency, Traceability and Trustworthiness, something we strive to infuse into our product offerings to customers, while ensuring security, privacy and regulatory compliance (a topic we will address in our next post).

101 views0 comments

Recent Posts

See All

Comments


© 2023 by Samvid

Powered and secured by Wix

bottom of page