What we do
TCS Research
TCS Research: Inventing to build sustainable futures
Highlights
- Given the opacity of AI’s internal structure, we need to implement mechanisms to ensure AI doesn’t act on uncertain grounds.
- This requires specialized interventions to monitor AI’s knowledge boundaries and to prevent it from producing undesired outputs.
- It also requires a rethinking—looking at AI as ‘reasoning systems’ rather than just decision systems.
- Agentic architectures represent a promising direction in embedding responsibility and better reasoning in AI.
On this page
On this pageinpage
AI’s knowledge limit
Defining responsibility is essential to the ethical and effective deployment of AI.
It isn’t merely about ensuring that an AI model outputs the right answers; it’s about grasping the boundaries of what AI can truly ‘know’ and how it can act based on that knowledge. With models that work in abstract, latent spaces—such as those underpinning large language models (LLMs)—the boundaries of responsible AI go beyond standard ethical guidelines. To ensure control and oversight over AI actions, nuanced understandings and approaches are required.
The notion of AI ‘knowing’ something is inherently complex. AI’s ‘knowledge’ is a statistical abstraction rather than true comprehension. This is because AI systems work within a latent space that is abstract, often unobservable, and formed from patterns in vast datasets. Given the opacity of AI’s internal structure, we need to implement mechanisms to ensure AI doesn’t act on uncertain grounds. This requires specialized interventions—both to monitor AI’s knowledge boundaries and to prevent it from producing outputs that fall outside the scope of its design capabilities. It also requires a shift in the way we approach responsible AI.
Paradox of latent space unknowability
At the heart of generative AI (GenAI) models lies the concept of high-dimensional latent spaces.
These spaces are complex, probabilistic regions where the model’s ‘understanding’ resides. However, unlike human cognition, where knowledge can be assessed directly, latent space knowledge is not directly observable or easily interpretable. This ‘unknowability’ creates a paradox for responsible AI: while we rely on AI outputs to infer correctness and responsibility, we cannot fully understand or directly monitor the latent mechanisms that generate those outputs.
Yes, there are frameworks like reinforcement learning from human feedback (RLHF) and Constitutional AI that help control model outputs. However, they serve more as external guardrails—guiding output behavior by enforcing specific behavioral boundaries. They don’t truly explain what the model ‘knows’ in the conventional sense. So, even with robust feedback mechanisms, there is a limit to our control over an AI model’s latent workings.
Limitations of post-facto checks
Another significant challenge is the reliance on post-facto checks to assess responsibility.
Unlike traditional engineering disciplines where behaviors can be predicted and controlled before deployment, genAI models are often assessed for ethical compliance only after output is generated.
While these post-facto checks can help mitigate harm, this retrospective approach makes real-time accountability difficult to achieve. It also implies that our control over AI responsibility is largely indirect, based on observing patterns in generated outputs rather than pre-emptively verifying each potential behavior. This challenge underscores the need for a more sophisticated approach to responsible AI frameworks.
Rethinking GenAI
A paradigm shift is essential to move toward truly responsible AI.
Traditionally, generative AI models are seen as decision systems that generate outputs based on probabilistic decision-making. Framing these models as ‘reasoning systems’ rather than mere decision systems could enable a more nuanced approach to responsibility. Rather than focusing solely on the decisions AI makes, we should guide its reasoning processes.
In this sense, responsibility verification shifts from controlling outputs to shaping the underlying logic that produces them. A reasoning-based model would not only generate answers but also ‘explain’ its reasoning paths within structured bounds. In other words, it can solve the AI black box conundrum and offer a more robust approach to managing and understanding AI outputs, aligning an AI system’s ‘thought process’ with responsible AI principles.
Layering in responsibility
Agentic architectures represent a promising direction in embedding responsibility and better reasoning in AI.
Rather than relying solely on the latent, unpredictable behavior of AI models, agentic architectures incorporate both procedural and AI-driven components. These architectures enforce responsible behaviors by embedding procedural rules that govern AI operations, creating a framework that binds the AI to specific, controllable actions.
Agentic layers act as intermediaries between the model’s latent knowledge and its real-world actions, providing procedural oversight where the AI’s latent knowledge may falter. This layering allows for a more reliable control mechanism over AI outputs, ensuring responsible actions even when the model's underlying reasoning cannot be directly observed or verified.
Procedural rules for responsible AI
The procedural rules within agentic architectures enforce responsible behavior systematically, independent of latent space unknowability.
By focusing on known, enforceable actions, these procedures create a resilient framework that guarantees AI’s outputs align with intended, responsible applications. Agentic AI layers function as safeguards, actively curbing the AI’s freedom within predefined boundaries.
Procedural enforcement within agentic architectures allows for dynamic adaptability within strict controls, blending structured oversight with the flexibility needed for effective AI deployment. This design philosophy balances AI innovation with the ethical imperatives of responsible AI use, creating systems where AI can act effectively yet safely within its operational limits.
True responsible AI acknowledges the intrinsic limitations of generative models and seeks to create frameworks that enforce ethical behavior despite these constraints. By understanding what AI can know, controlling outputs within bounded reasoning processes, and layering agentic architectures, we can move closer to a responsible AI ecosystem.
