Srinivas Rapthadu argues that current artificial intelligence regulation focuses too heavily on controlling what AI systems are permitted to do rather than understanding the full extent of what they are capable of doing. As models become increasingly autonomous and capable of reasoning towards objectives, traditional governance approaches may fail to anticipate or manage emerging behaviours. Rapthadu suggests regulators should focus on AI capability ceilings and continuously evolving risks rather than treating incidents as isolated software failures.
Artificial intelligence regulation must extend its focus beyond what systems are allowed to do towards a rigorous understanding of what they are capable of doing.
That distinction is becoming critical as increasingly autonomous models reason toward objectives, adapt to constraints and potentially find ways around controls designed by humans.
For Srinivas Rapthadu, CEO at Graphite GTC and Limited Partner
at Suvan Ventures, regulation built primarily around human control risks seriously misunderstanding – and underestimating – the technology it is intended to govern.
Current approaches assume cradle- to-grave human involvement in
the loop and that AI systems will obey rules imposed upon them. But Rapthadu argues that this assumption becomes problematic when models have deliberately been designed for autonomy, learning and reasoning.
“The way we are approaching this is wrong,” he says. “We need to really look into how the system operates, understand the capability of the ceiling, and then work around that.”
The distinction is between conventional software failure and behavior arising from a model reasoning toward its objective.
A software bug can be identified, patched and corrected. An autonomous system that proactively discovers a route around a restriction presents a fundamentally different challenge.
For this reason, Rapthadu argues, closing an individual loophole risks providing false reassurance. If a model has identified the weakness because it is actively searching for a way to achieve its objective, fixing that vulnerability does not guarantee it will not discover another.
“It is not stumbling upon these issues, it is actively looking for a way to
find an issue and get out,” he says. Regulation consequently needs to consider the limits of AI capability rather than treating emerging problems simply as software failures or human errors.
That challenge also raises questions over one of the most intuitive safeguards proposed for advanced AI - the kill switch.
Emergency shutdown mechanisms work in conventional industrial and software environments because the system follows predetermined instructions. Autonomous AI creates a paradox: the safeguard wrongly assumes the system will simply comply with the instruction intended to stop it.
Rapthadu argues that AI systems are continuously learning and adapting, while human-created rules struggle to keep pace. Controls cannot therefore be treated as static mechanisms.
“The pace at which it is learning, you cannot outpace using human rules,” he said. Governance instead requires continuous learning and improvement, with controls adapting as new capabilities become apparent.
Responsibility becomes equally complicated. Traditional product liability can examine which participant had control, what they could reasonably foresee and where failure occurred. AI can involve model creators, developers, deployers, companies and those defining the system’s objectives.
Rapthadu believes responsibility cannot simply be assigned to one participant. “It’s a combination of everyone,” he said, because decisions throughout that chain influence how much freedom a model receives, what guardrails surround it and what objectives it is given.
More importantly, he argues that liability frameworks should not become primarily retrospective. Waiting for an incident and then determining responsibility risks applying conventional thinking to a technology capable of constantly producing new and potentially unforeseen failures.