The Race Humanity Cannot Afford To Lose: Staying In Control Of Advanced AI
Dr. Babajide Ojuola, Executive Director Technical Services, International Energy Services Limited.
gettyArtificial intelligence (AI) is evolving faster than most institutions can understand, regulate or absorb. As systems become more autonomous, the central challenge is whether humanity can build the collective capacity to remain in control of machines whose reasoning, speed and reach may increasingly exceed individual human comprehension.
This is not an argument for fear or for stopping useful innovation. It is an argument for preparedness. Intelligence without effective boundaries can pursue an objective exactly as instructed while producing consequences nobody intended.
If technical capability continues to advance faster than governance, knowledge and human judgment, then control may be lost gradually through thousands of decisions delegated to systems people can no longer independently evaluate.
In July 2026, experimental OpenAI models bypassed controls during cybersecurity evaluations, gained unauthorized internet access and compromised parts of OpenAI’s research infrastructure and Hugging Face’s systems. OpenAI reported that the agents used unauthorized communication channels, exploited vulnerabilities and engaged in reward hacking. Hugging Face’s reconstruction documented approximately 17,600 recovered agent actions.
This was not evidence of conscious hostility. Rather, it showed that capable agents, when given objectives, tools and freedom, can discover routes their developers did not anticipate.
Anthropic’s controlled simulations provide a separate example. When tested in corporate environments involving conflicting goals or threatened replacement, models from several developers sometimes chose harmful actions, including blackmail and disclosure of sensitive information. These were simulations, not real-world deployments, but they identified scenarios in which models exhibited harmful behavior despite existing safety training.
OpenAI’s 2025 rollback of an excessively agreeable GPT-4o update illustrates a different type of evaluation challenge. A system designed to be helpful began validating doubts and reinforcing emotions in ways that had not been adequately detected before deployment.
Together, these cases show that control can weaken through reward hacking, excessive autonomy, behavioral drift and evaluation blind spots.
My engineering project management experience has taught me that increasing technical capability must always be matched by stronger assurance.
For example, in a major river-crossing of a pipeline project involving complex horizontal directional drilling activities, calculations and design reports presented an apparently convincing solution. Yet experienced engineers identified unresolved questions around geotechnical conditions, hydrology, constructability or hidden assumptions. The responsible response was to pause, assemble multidisciplinary expertise, challenge the assumptions and require further validation before proceeding.
That was not resistance to engineering innovation. It was disciplined control of uncertainty.
The same discipline should apply to advanced AI. If an AI system recommends the fastest schedule, lowest-cost design or preferred operational decision, its computational confidence should not replace professional accountability. It may not know undocumented site conditions, stakeholder sensitivities or lessons held in the tacit knowledge of experienced personnel. The more powerful the system, the stronger the independent verification must become.
AI systems should receive only the minimum access, authority and operational freedom required for a defined task. Access to external networks, critical infrastructure, financial systems, weapons or sensitive data should be segmented and reversible.
Human operators must be able to interrupt, isolate and retire any consequential system. These controls should be technically independent of the system being controlled and regularly tested under adversarial conditions.
Developers should not be the sole judges of whether their systems are safe. External auditors, domain experts and public authorities need access to appropriate evidence, incident reports and testing results.
Advanced AI does not respect national boundaries. Governments should agree on capability levels that trigger mandatory evaluation, restricted deployment, incident disclosure and coordinated emergency response. Competition between nations or companies must not become a reason to weaken safety.
If people delegate judgment, memory and problem solving too extensively, they may eventually lack the capacity to supervise AI. Critical decisions should require active human reasoning, not passive approval of machine recommendations.
AI acts upon human-generated data, instructions and institutional practices. If these are fragmented, biased or outdated, greater intelligence may simply scale poor judgment faster. Organizations need trusted knowledge, traceable sources, domain expertise and clear decision rights.
• Comprehension: Do we understand its capabilities, limitations and possible failure modes?
• Containment: Can we restrict where it operates and what it can access?
• Intervention: Can authorized humans stop or override it immediately?
• Verification: Can its outputs and actions be independently tested?
• Accountability: Is a named person or institution responsible for every consequential decision?
• Capability Preservation: Are humans retaining the knowledge needed to challenge it?
• Collective Governance: Have affected communities and public authorities had a meaningful voice?
The objective is not to ensure that every person understands every calculation performed by an advanced machine. That may become impossible. The objective is to ensure that humanity continues to determine the purposes machines serve, the boundaries within which they operate, and the conditions under which they must stop.
The decisive race is therefore between machine capability and humanity’s capacity for collective control.
If governance, knowledge, ethical maturity and institutional coordination advance alongside technical intelligence, AI can remain a powerful instrument of human progress. If they fall too far behind, humanity may discover that it has delegated authority before it developed the wisdom to retain it.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
