Artificial intelligence is moving rapidly from being a tool that assists human beings to becoming a system that can increasingly reason, recommend, predict, generate, and act. This transition raises an important question that deserves greater attention: if AI capability continues to increase, how do we ensure that human behavioural capability keeps pace?
In August 2026, Bill Gates described the current period as a turbulent AI era and argued that AI is becoming capable of replacing and exceeding human cognition in some domains. He also highlighted the possibility that increasingly powerful AI systems could eventually act against human interests and that society could lose control over them. Gates identified the preservation of human control as one of the profound questions raised by AI.
This concern does not necessarily imply that AI development should be viewed as a contest between machines and humans. A different question may be more productive:
What happens to human behaviour when humans increasingly work through, with, and around intelligent machines?
The overlooked variable: human behavioural precision
Much of the discussion surrounding AI safety has understandably focused on the capabilities of AI systems themselves: model reliability, alignment, transparency, robustness, explainability, cybersecurity, governance and regulation.
Yet AI does not operate in isolation.
In many real-world settings, there is a human being somewhere within the system who must interpret an AI output, verify information, recognize an error, make a decision, override a recommendation, communicate a result, or execute an action.
That human action may occur within seconds.
The quality of the overall system may therefore depend not only on whether the AI produces a technically sophisticated output, but also on what the human does at the point of interaction with that output.
This is becoming an increasingly important area of research. Recent Springer Nature literature on human oversight has emphasized that merely placing a human “in the loop” does not automatically create meaningful oversight. Human overseers can experience automation bias, complacency and difficulties detecting inaccurate or inappropriate AI outputs.
A 2025 review in AI & Society, for example, examined automation bias in human–AI collaboration and identified over-reliance on automated recommendations as a significant challenge, while also highlighting factors such as AI literacy, professional expertise, trust and verification demands.
This suggests that the human component of AI systems cannot simply be assumed to function perfectly.
It needs to be studied.
And potentially, it needs to be measured.
From human oversight to human behavioural precision
This is where the concept of Precision Behaviour Score (PBS; Bhadran's Score) may offer a new direction for investigation.
Developed by Dr. Renjith Seela Bhadran within the broader Point-of-Generation Segregation Theory (PGST) and subsequent behavioural precision framework, PBS was originally conceptualized around the measurement of behavioural precision at the point where an individual actually performs an action.
The underlying proposition is simple:
Behaviour is not only an attitude, intention or knowledge state. It can be examined at the moment of execution.
This distinction becomes particularly interesting in an AI-enabled environment.
Consider a clinician receiving an AI-supported diagnostic recommendation.
The relevant question is not simply:
“Did the AI make the correct recommendation?”
There is another chain of questions:
-
Did the clinician correctly interpret the output?
-
Did the clinician verify the relevant information?
-
Did the clinician identify contextual information that the AI may not have incorporated?
-
Did the clinician recognize an inappropriate recommendation?
-
Did the clinician intervene when intervention was necessary?
-
Did the clinician avoid blindly accepting the recommendation?
-
Was the final action consistent with the available evidence and professional responsibility?
These are human behavioural events.
They occur at identifiable points in a human–AI system.
They could therefore potentially become measurable behavioural units.
AI dependency is not the same as AI use
An important distinction may be necessary here.
The objective should not be to oppose AI use.
AI can expand human capabilities, reduce workload, improve access to information and support complex decision-making. Bill Gates himself has emphasized both the potential benefits of AI and the need to manage its risks.
The concern is different:
What happens when assistance becomes dependency?
A human may gradually move from:
using AI as a decision-support tool
towards:
accepting AI as the default decision authority.
This distinction is important because recent research has documented automation bias and excessive reliance on automated recommendations.
The challenge, therefore, may not be to reduce human interaction with AI, but to ensure that human interaction remains behaviourally precise.
The future may require stronger humans, not weaker humans
This leads to a potentially different way of thinking about AI development.
As AI capability increases:
AI capability ↑
does not necessarily imply:
Human capability ↓
Instead, society could deliberately pursue:
AI capability ↑ + Human behavioural precision ↑
The objective would be complementary intelligence rather than substitution of human agency.
AI could provide computational capability.
Humans would retain contextual interpretation, responsibility, judgement and the authority to intervene.
Recent work on meaningful human oversight similarly distinguishes between AI's ability to generate solutions and the human capacity to evaluate, contest and override those solutions.
This provides an important conceptual space for behavioural science.
Could PBS become a measurement layer for human–AI interaction?
The proposition I am advancing is not that PBS has already solved AI dependency or AI control.
That would require empirical evidence.
Rather, PBS provides a testable hypothesis:
If human behaviour at critical points of AI interaction can be operationalized and measured as behavioural precision, then behavioural interventions may potentially strengthen human oversight and reduce inappropriate dependence on AI outputs.
This could be investigated experimentally.
For example, researchers could identify a series of human–AI execution points and measure whether individuals:
-
independently assess the situation;
-
correctly interpret AI outputs;
-
verify critical information;
-
detect erroneous or inappropriate recommendations;
-
appropriately override AI when required;
-
avoid unnecessary rejection of accurate AI recommendations;
-
execute the final action correctly.
Such measurements could potentially be combined into a behavioural precision framework.
The important scientific question would then become:
Can behavioural precision be increased through structured training, feedback and reinforcement—and does increased behavioural precision improve human oversight of AI?
That is an empirical question.
From AI safety to human–AI behavioural safety
This may also expand the way we conceptualize AI safety.
AI safety is often discussed primarily in terms of the machine:
Is the system accurate?
Is it robust?
Is it aligned?
Is it secure?
Is it explainable?
But a deployed AI system is usually a socio-technical system.
Therefore another question becomes relevant:
Is the human behaviour surrounding the system sufficiently precise?
A highly capable AI system combined with imprecise human behaviour may still produce poor outcomes.
Conversely, a capable AI system combined with appropriately trained, vigilant and behaviourally precise human oversight could potentially produce a different level of system reliability.
This is particularly important in healthcare, aviation, autonomous systems, industrial operations, public administration and other high-stakes environments.
The control problem may therefore have a behavioural dimension
The idea of maintaining human control over increasingly capable AI is often discussed as a technical or governance problem.
It may also have a behavioural dimension.
Human control requires more than formal authority.
A person may technically have the authority to override an AI system but fail to do so.
A person may technically be responsible for a decision but simply accept the AI recommendation.
A person may technically be “in the loop” while contributing almost nothing to the actual decision.
Recent research on human oversight has raised precisely this concern: human involvement can become little more than symbolic approval if humans lack the ability or willingness to meaningfully evaluate and challenge AI outputs.
Therefore:
Human control of AI requires not merely humans in the loop, but humans capable of acting precisely within the loop.
That distinction could be fundamental.
A possible future research architecture
This creates an intriguing research pathway for PBS.
A future Human–AI Behavioural Precision Framework could investigate:
AI capability
↓
Human–AI interaction
↓
Critical behavioural event
↓
Human perception and interpretation
↓
Verification
↓
Decision
↓
Intervention / acceptance / override
↓
Final human action
↓
Outcome
PBS could potentially be evaluated at selected behavioural nodes within this pathway.
The resulting research could examine whether behavioural precision changes with:
-
AI reliability;
-
task complexity;
-
user expertise;
-
AI explainability;
-
time pressure;
-
frequency of AI errors;
-
level of automation;
-
training;
-
feedback;
-
repeated exposure;
-
and degree of AI dependence.
This would move the discussion from philosophical concern to measurable behavioural science.
The human future in an AI world
Perhaps the most important question is not:
“How powerful will AI become?”
It is:
“How precise will human behaviour remain as AI becomes more powerful?”
The answer will not be known in advance.
It must be studied.
AI will almost certainly continue to evolve. Human beings therefore face a choice about how to respond to that evolution—not by necessarily resisting AI, but by strengthening the human capabilities required to work with increasingly capable systems.
This is where Precision Behaviour Score may have a future role.
The proposition is deliberately simple:
AI may scale machine intelligence. Behavioural precision may help scale human control.
If this proposition can be empirically demonstrated, PBS could potentially evolve beyond its original applications into a broader framework for studying human behavioural precision in increasingly AI-mediated environments.
The future challenge may therefore not be humans versus AI.
It may be:
humans becoming precise enough to remain meaningfully human while working with machines that become increasingly intelligent.
And perhaps the defining scientific question of the coming decades will be:
As AI becomes more capable, can human behavioural precision rise at least as fast as human dependence on AI?
That question is open.
And it is measurable.
References for the blog
- Gates, B. The turbulent AI era is here. The choices we make now are critical. Gates Notes, 26 August 2026.
- Zhu, L., Lu, Q., Ding, M., Lee, S.U., et al. Designing meaningful human oversight in AI. AI and Ethics (2026).
- Langer, M., Baum, K., & Schlicker, N. Effective Human Oversight of AI-Based Systems: A Signal Detection Perspective on the Detection of Inaccurate and Unfair Outputs. Minds and Machines 35, 1 (2025).
- Romeo, G., & Conti, D. Exploring automation bias in human–AI collaboration: a review and implications for explainable AI. AI & Society 41, 259–278 (2026).
- Jovchevski, P., Buijsman, S., & Neerincx, M. What is Wrong With Automation Bias? Philosophy & Technology (2026).
- Human control of AI systems: from supervision to teaming. AI and Ethics (2024).
- Bhadran, R.S., Vasudevan, D. Bhadran’s point of generation segregation theory for behavioral precision in biomedical waste management. Sci Rep 16, 2531 (2026). https://doi.org/10.1038/s41598-025-32195-4