Judgement is a critical human capability, making sound judgement essential in automated environments. Developing sharp judgement protects teams from operational blind spots, while maintaining independent judgement ensures effective system governance. Without active judgement, reliance on AI grows unchecked, but exercising deliberate judgement preserves workplace capability. Strong judgement guides complex strategic decisions, so testing employee judgement remains a primary leadership priority. Analyzing how technology affects judgement helps organizations prevent cognitive decline, while protecting human judgement keeps operational workflows resilient. Continuous judgement practice strengthens organizational oversight, ensuring sound judgement stays central to success. Valuing human judgement secures long term success.
The rise of military robotics raises a governance question that goes far beyond whether the technology works. On 26 September 2026, former Ukrainian Defence Minister Mykhailo Fedorov announced an initiative with an extraordinary ambition: a defense technology initiative called the Army of Robots.
His concern was clear. As military robotics develops, we may not simply be looking at machines protecting soldiers from dangerous tasks. We may also be entering another technological arms race, where increasingly autonomous systems become part of warfare.
The Kyiv Independent article described an ambitious vision: robotic systems increasingly taking the place of human soldiers in dangerous battlefield environments. Many of the futuristic machines shown in the accompanying promotional video are AI generated concepts rather than robots currently operating on the battlefield. Most unmanned ground vehicles currently being used remain remotely operated by humans.
Removing a human from physical danger does not necessarily remove human responsibility from the system. It changes where human decision making sits, and it may change human evaluation habits over time.
There is an understandable reason for developing these technologies. If a machine can enter a dangerous area instead of a soldier, carry supplies, evacuate casualties, or identify threats, human lives may be protected. Ukraine is already using unmanned ground vehicles for some battlefield roles, generally operated remotely by humans.
A remotely controlled machine extends human action. A machine increasingly capable of interpreting conditions, selecting responses, and acting with greater independence begins to change where critical evaluation occurs.
Human control can mean many different things:
Does a human approve the mission or select the target?
Does a human supervise several machines simultaneously?
Can operators interrupt an action when anomalies appear?
How much information do operators receive before making a choice?
Can humans recognize when the machine may be wrong?
A person can remain formally responsible while their practical ability to influence outcomes significantly decreases.
Risk is usually evaluated in terms of failure: incorrect target identification, lost communication, software malfunction, or bad model outputs. However, another shift occurs while the technology appears to be working successfully: the human operators change.
[ High Verification / Active Challenge ] ──► System Reliability ──► [ Passive Deference / Reduced Oversight ]
Imagine a system that performs extremely well over hundreds of missions. At first, operators carefully verify its recommendations, challenge uncertain results, and intervene frequently. As the system continues getting things right, confidence increases, verification becomes lighter, and interventions become less frequent.
New operators enter the system having never experienced the earlier period when close supervision was considered necessary. Nothing has failed, yet the human software interaction has fundamentally altered.
This dynamic extends into banking, healthcare, infrastructure, insurance, government, and enterprise operations. The more reliable a system appears, the more reasonable it seems to rely upon it. Good technology should reduce unnecessary toil, but a governance problem hides inside that success.
As reliance increases, human capability surrounding the system can weaken:
Verification behavior decreases over time.
Challenge practices occur less frequently.
Domain knowledge required to perform the task manually dissolves.
Intervention procedures become unfamiliar to new operators.
The organization becomes more efficient while simultaneously becoming far more dependent upon the automated system.
An organization could have clear rules about who holds approval authority. However, authority is not the same as governing capability.
A person may possess the authority to intervene while becoming less capable of exercising that authority meaningfully:
Control Interface: The override control and procedure exist on paper.
Context Gap: The operator lacks sufficient information to evaluate the recommendation.
Knowledge Decay: The human no longer retains the underlying technical knowledge needed to challenge the machine.
Deference Habits: Operators become accustomed to accepting automated outputs rapidly.
The presence of formal human authority does not automatically prove the presence of human governing capability.
Authority shifts quietly through everyday choices rather than formal policy changes. A recommendation becomes routinely accepted, verification becomes occasional, and intervention becomes exceptional.
[ Formal Policy ] ──► Recommendation Accepted ──► Occasional Verification ──► [ Operational Decision ]
Eventually, the machine technically makes a recommendation, but operationally that recommendation becomes the decision unless something unusual happens.
Leadership must evaluate specific behavioral signals:
Are teams challenging the system as frequently as before?
Are operators independently verifying high-stakes outputs?
Is intervention declining due to system accuracy or human deference?
Can operators still recognize circumstances the system was not designed to handle?
Detecting that human reliance has increased should not automatically shut down an operational system. Detecting that verification has declined should not automatically revoke an operator’s authority.
These observations represent telemetry evidence:
[ Operational Telemetry ] ──► [ Evidence Interpretation ] ──► [ Authorized Human Decision ]
Observing change provides data showing that operating conditions have altered. The appropriate response still requires authorized human evaluation, supported by operational, legal, and technical evidence.
Separating observation from execution ensures that a system designed to support governance does not quietly become the system making the governance decision.
The better autonomous systems become, the more difficult governance becomes. Poor systems attract scrutiny; people watch them, question them, and expect mistakes. Highly reliable systems earn trust, making increased reliance rational.
Over long periods, successful autonomy reduces the amount of active evaluation performed around the system. The moment when humans feel least need to intervene may also be the period when their capacity to intervene begins to weaken.
Organizations must establish dynamic monitoring to track whether verification habits remain intact before critical edge cases force an unexpected failure.
To prevent cognitive erosion, enterprise governance must actively embed verification benchmarks directly into daily operational workflows.
Systems should periodically present operators with complex edge cases or synthetic scenarios requiring manual evaluation. By systematically testing human oversight in real time, organizations ensure that operators maintain active reasoning habits, contextual awareness, and the technical competence required to step in during unscripted system anomalies.
As machines take over direct operational tasks, humans are removed from direct participation. Sometimes this saves lives, improves safety, or boosts productivity. However, every time humans step back from a task, leadership must account for the capability surrounding it.
Human control is not preserved simply by keeping someone’s name on an approval workflow, nor is it preserved because an override button exists.
The deeper requirement is ensuring humans remain capable of recognizing when the machine should not continue doing what it has become very good at doing. The future of autonomous systems cannot only be about building machines capable of acting without us; it must ensure that we never forget how to evaluate them.