Aliso Viejo, California: Are AI models advancing faster than our ability to control them? That seems to be one of the questions thrown up by the case Elon Musk lost on Monday against OpenAI and Sam Altman. Musk had accused Altman of breaching a non-profit contract by shifting ChatGPT's creator company to a for-profit one after the former had 'donated' $38 mn. Musk lost on a technicality. But regulation remains key in the 'AI as beneficial for humanity vs for profit' debate.
Power users of AI elicit about 7x reasoning compute that median users do. Frontier developers submit 17x queries of their peers. The top decile of knowledge workers reclaims 10 or more hours a week, while median workers recover under an hour a day. Capability and oversight are not racing on the same track. There is a deployment gap, a governance coordination issue, and workforce-readiness problem.
A January 2021 paper in American Economic Journal, 'The Productivity J-Curve: How Intangibles Complement General Purpose Technologies', shows that organisations using general-purpose technologies (GPTs) dip in productivity when they rebuild around the new capability, then emerge significantly ahead. Electricity, ICE and enterprise internet each followed this trajectory. We are now in the dip, a phase that will end.
Critics correctly observe that frontier models possess capabilities most enterprises have no immediate use for. This is sometimes presented as evidence of recklessness. But US hyperscalers are projected to spend close to $600 bn on AI infra in 2026, roughly 2x their 2024 outlay. China continues to invest at sovereign scale. Gulf states have repositioned from hydrocarbon rents to compute as a strategic asset.
No single firm or nation can credibly slow the supply side without ceding leadership. A pause in capability development in this market structure is a transfer of standard-setting power to whoever doesn't pause.
The most effective tools available for evaluating and constraining advanced AI systems are AI systems themselves. The toolchain is more developed than policy discourse acknowledges. Reinforcement learning from human feedback shapes model behaviour during training against curated preference data. Constitutional methods train against explicit normative principles.
There are genuine failure modes in production across enterprise deployments every week. Shadow AI bypasses official channels and breaks audit trails. Agentic systems compress latency between flawed reasoning and irreversible action from weeks to milliseconds. Synthetic media has crossed the threshold of human distinguishability for an expanding share of content. The window between a published vulnerability and a functional exploit has shrunk to hours.
Operational governance problems have operational governance answers:
Continuous observability, rather than periodic audit.
Agent autonomy calibrated to the task's risk profile.
Centralised deployment platforms where every system is benchmarked, instrumented and logged before rollout.
Immutable disagreement logs where humans and models reach different conclusions.
The discipline required to deploy responsibly is well understood, but unevenly implemented. Anthropic revised its Responsible Scaling Policy in 2026. Some read the changes as a retreat from binding commitments. What the firm actually did was acknowledge a structural reality.
Catastrophic risk is a function of the entire industry's trajectory rather than any single firm's behaviour. A unilateral pause by a safety-conscious lab transfers leadership to actors with weaker safety cultures, with negative net effects on global outcomes.
So, meaningful safety work increasingly happens at the layer above the firm:
At the pre-deployment evaluation infrastructure stage.
With sector-specific guard rails for healthcare, finance and public services where cost of error is highest.
Liability frameworks for agentic systems.
Concrete benchmarks published by standards bodies, rather than abstract principles negotiated at summits.
Narrow international coordination on catastrophic risk categories.
This work is moving at institutional speed, slower than model release cadence. The lag is a feature of governance institutions, not evidence that the tech is uncontainable. The 'global south' is becoming the governance laboratory. New Delhi AI Impact Summit 2026 committed signatories to the principle that AI benefits should be democratically diffused, rather than concentrated in a small number of foreign labs. India's 2026 AI Governance Guidelines articulate the philosophy explicitly. MeitY calls it 'Innovation over Restraint', supplementing existing law with an AI Safety Institute and AI Governance Group.
Frontier labs are publishing safety frameworks with increasing specificity. Standards bodies are converging on benchmark suites. The 'global south' is demonstrating that pro-innovation governance can coexist with serious risk mitigation. Enterprises are beginning to build agile governance and centralised AI studios that the next phase of deployment requires.
AI will continue to develop. But will organisations and institutions using it keep pace? Irrespective of Monday's Musk v. Altman verdict, that outcome is within our control.
Power users of AI elicit about 7x reasoning compute that median users do. Frontier developers submit 17x queries of their peers. The top decile of knowledge workers reclaims 10 or more hours a week, while median workers recover under an hour a day. Capability and oversight are not racing on the same track. There is a deployment gap, a governance coordination issue, and workforce-readiness problem.
A January 2021 paper in American Economic Journal, 'The Productivity J-Curve: How Intangibles Complement General Purpose Technologies', shows that organisations using general-purpose technologies (GPTs) dip in productivity when they rebuild around the new capability, then emerge significantly ahead. Electricity, ICE and enterprise internet each followed this trajectory. We are now in the dip, a phase that will end.
Critics correctly observe that frontier models possess capabilities most enterprises have no immediate use for. This is sometimes presented as evidence of recklessness. But US hyperscalers are projected to spend close to $600 bn on AI infra in 2026, roughly 2x their 2024 outlay. China continues to invest at sovereign scale. Gulf states have repositioned from hydrocarbon rents to compute as a strategic asset.
No single firm or nation can credibly slow the supply side without ceding leadership. A pause in capability development in this market structure is a transfer of standard-setting power to whoever doesn't pause.
The most effective tools available for evaluating and constraining advanced AI systems are AI systems themselves. The toolchain is more developed than policy discourse acknowledges. Reinforcement learning from human feedback shapes model behaviour during training against curated preference data. Constitutional methods train against explicit normative principles.
There are genuine failure modes in production across enterprise deployments every week. Shadow AI bypasses official channels and breaks audit trails. Agentic systems compress latency between flawed reasoning and irreversible action from weeks to milliseconds. Synthetic media has crossed the threshold of human distinguishability for an expanding share of content. The window between a published vulnerability and a functional exploit has shrunk to hours.
Operational governance problems have operational governance answers:
Continuous observability, rather than periodic audit.
Agent autonomy calibrated to the task's risk profile.
Centralised deployment platforms where every system is benchmarked, instrumented and logged before rollout.
Immutable disagreement logs where humans and models reach different conclusions.
The discipline required to deploy responsibly is well understood, but unevenly implemented. Anthropic revised its Responsible Scaling Policy in 2026. Some read the changes as a retreat from binding commitments. What the firm actually did was acknowledge a structural reality.
Catastrophic risk is a function of the entire industry's trajectory rather than any single firm's behaviour. A unilateral pause by a safety-conscious lab transfers leadership to actors with weaker safety cultures, with negative net effects on global outcomes.
So, meaningful safety work increasingly happens at the layer above the firm:
At the pre-deployment evaluation infrastructure stage.
With sector-specific guard rails for healthcare, finance and public services where cost of error is highest.
Liability frameworks for agentic systems.
Concrete benchmarks published by standards bodies, rather than abstract principles negotiated at summits.
Narrow international coordination on catastrophic risk categories.
This work is moving at institutional speed, slower than model release cadence. The lag is a feature of governance institutions, not evidence that the tech is uncontainable. The 'global south' is becoming the governance laboratory. New Delhi AI Impact Summit 2026 committed signatories to the principle that AI benefits should be democratically diffused, rather than concentrated in a small number of foreign labs. India's 2026 AI Governance Guidelines articulate the philosophy explicitly. MeitY calls it 'Innovation over Restraint', supplementing existing law with an AI Safety Institute and AI Governance Group.
Frontier labs are publishing safety frameworks with increasing specificity. Standards bodies are converging on benchmark suites. The 'global south' is demonstrating that pro-innovation governance can coexist with serious risk mitigation. Enterprises are beginning to build agile governance and centralised AI studios that the next phase of deployment requires.
AI will continue to develop. But will organisations and institutions using it keep pace? Irrespective of Monday's Musk v. Altman verdict, that outcome is within our control.
(Disclaimer: The opinions expressed in this column are that of the writer. The facts and opinions expressed here do not reflect the views of www.economictimes.com.)
Adnan Masood