Welcome to the first Read Ahead. Every Monday: three pieces worth your time on how staffs decide and what breaks when software meets a real command, each with our take, plus one from outside defense.
Every tool that makes a staff faster also makes it easier to stop checking the work. This issue sits in that tension: how fast people can master a system, and whether they still notice when it’s wrong.
The staff’s most reliable AI may be the most dangerous
What happened: Anthony Pfaff, who directs strategic research at the Army War College, argues in Parameters that the military is judging AI by the wrong standard. Reliability works for outputs you can check, like whether an object is an enemy vehicle. It breaks down for plans, predictions, and assessments of enemy intent, which can’t be verified in advance and are often still contestable afterward. He proposes judging the “cognitive fit” of the whole human-machine system instead: how well it represents the environment, reasons about it, and adapts when it’s wrong.
Why it matters: The tension is real: world-class machines can still produce a failed plan if the staff stops thinking. Reliability is easy to show when you can check the answer. But a plan is a bet on the future. How do you score accuracy on something that hasn’t happened yet and may not happen?
Pfaff’s most uncomfortable point, borrowed from AI researcher John Zerilli and colleagues, is that AI can be most dangerous when it’s most reliable. Systems that fail occasionally keep humans watching, but a system that rarely fails is easy to overlook.
Our take:
1) Pfaff’s argument implies that faster and better need to be balanced, and the balance tips towards better. A boat rowing faster in the wrong direction isn’t better off than a slower one heading the right way. Speed only counts when the information is right, the staff understands it, and it leads the commander to a sound decision.
2) If reliability breeds complacency, then evaluating the machine isn’t enough and the staff must also be evaluated. Not on their ability to plan, but on their ability to catch the machine’s mistakes.
───────────────────────
The clock should stop when the Soldier can catch the machine
What happened: Michael Woodhouse, a retired air defense officer now in business development at Leonardo DRS, argues that time to proficiency should be a formal requirement beside range, reliability, and lethality. He defines it as how many days it takes a representative operator to perform under realistic pressure without a trainer prompting the next move. Without that requirement, fast fielding only moves the delay to the gaining unit.
Why it matters: His definition already includes recognizing a problem, recovering from an error, and explaining the decision afterward. That’s the part worth protecting. Read next to Pfaff, an interface tuned to make Soldiers proficient fast can also make them trust it fast. Proficiency tests should include the night the system is wrong.
───────────────────────
AFRICOM’s innovation chief is a Lieutenant Lolonel
What happened: U.S. Africa Command named its first chief innovation officer, Lt. Col. Jared Bindl, to drive technology adoption as the command’s ground presence shrinks. One of his five priorities is AI-enabled command and control, aimed at better campaign planning, shorter decision cycles, and wider AI adoption across the headquarters.
Why it matters: His charter is to align the headquarters, component commands, and partners, a broad coordination mandate for a Lieutenant Colonel. Innovation offices tend to succeed or stall on access and authority, so how this role is resourced is worth watching. Also, every stated C2 objective is about speed, which is easy to measure. Better, as Pfaff notes, is not.
───────────────────────
THE OUTSIDE READ (The one your boss hasn’t read yet)
Make people check the work, not just wait
Researchers at the University of Michigan–Dearborn tested two ways to stop people from rubber-stamping AI recommendations. One forced a delay before people could decide while the other required them to engage with the content itself. The content requirement worked best. It decreased overreliance to 5.7%, compared to 8.6% with no intervention.
There are important caveats to note from this study. Only 35 participants, one task, and an accuracy gain that did not meet statistical significance. But the overall lesson carries over to a staff. A mandatory pause before accepting an AI-generated product is friction of time. A requirement to check one key assumption against the source data is friction of thought. Early evidence points to the second but needs further studies to verify.
The term to carry into your next meeting is active verification: making the human engage with the substance before accepting the machine’s answer. This is the difference between a staff member who tries to blame it on “the AI” and one who can confidently answer their General’s questions.
───────────────────────
Onebrief builds planning software for military staffs, which is why we care about the difference between faster and better.
What did I miss? Let me know. RunningEstimate@Onebrief.com
Running Estimate is published by Onebrief, which builds planning software for military staffs. A running estimate is what a staff keeps current so the next decision doesn’t start from zero. This is ours, in public.
