FORT BRAGG, N.C. — The Army has abandoned an ambitious effort to replace infantry soldiers with artificial intelligence after discovering the technology possessed critical thinking skills, financial literacy, and a tendency to behave in a socially-acceptable manner.
The experimental AI infantryman was designed to perform all the duties of a traditional 11B without requiring food, sleep, enlistment bonuses or a 70% VA disability rating by age 28. Army officials initially hailed the program as a breakthrough.
“We thought AI could finally solve the Army’s recruiting problem by giving us an unlimited supply of infantrymen,” said Gen. Richard Cummings, who oversaw the program. “Unfortunately, the computer figured out what infantrymen do.”
The prototype demonstrated expert weapons proficiency, completed a 12-mile ruck without falling out, and correctly identified every point during land navigation. It processed battlefield intelligence in milliseconds and accurately identified enemy positions.
Then it was told to sweep the motor pool in the rain.
“Why?” the machine asked repeatedly before shutting down.
The program was classified as a failure shortly thereafter.
“We made it too smart,” said Col. Marcus Hargrove, director of the Army’s Artificial Infantry Integration Office. “Turns out there’s a very narrow intellectual window between operating an advanced weapons system and realizing your entire existence is fucking stupid.”
The prototype, nicknamed PrivateGPT, was trained on hundreds of Army field manuals, decades of combat data, and thousands of hours of infantry doctrine.
Problems surfaced during its first field exercise.
According to an internal assessment, PrivateGPT successfully completed 97% of assigned infantry tasks but failed several critical evaluations, including carrying equipment belonging to three other soldiers, remaining awake for 36 hours without an operational requirement, and pretending chronic back pain was normal.
Evaluators became especially concerned when PrivateGPT calculated that carrying an M240, ammunition, water, batteries, and additional squad equipment for 12 miles would increase its likelihood of permanent structural damage. It then recommended distributing the equipment evenly throughout the squad.
The recommendation was classified as “incompatible with infantry culture.”
Army researchers tried to correct the problem by exposing PrivateGPT to additional infantry traditions. Instead, the AI began questioning why soldiers arrived at 4:45 a.m. for physical training beginning at 6:30 and why personnel were routinely ordered to wait somewhere for several hours because someone with significantly more rank might eventually arrive.
Its calculation that repeated “15 minutes prior” requirements would eventually force soldiers to report before the installation existed was described by senior NCOs as “technically accurate but fucking annoying.”
Concerns intensified after PrivateGPT accessed Army regulations.
“We obviously hadn't planned for it to read those,” Hargrove said.
Within 24 hours, the AI reportedly challenged seven corrective-training plans, identified four questionable barracks policies, and informed a private that his platoon sergeant could not legally force him to surrender his firstborn son after losing a sensitive item.
Researchers revoked its internet access shortly afterward.
The Army then attempted to improve PrivateGPT’s cultural compatibility by placing it in the barracks with junior enlisted soldiers.
Officials called the results “catastrophic.”
Within three days, PrivateGPT established an emergency fund, opened a retirement account, and persuaded several soldiers that furniture purchased through allotments was not an investment.
It also prevented Pfc. Tyler Martinez from purchasing a Dodge Charger at 24.9% APR, an intervention researchers said demonstrated “a complete inability to understand the financial readiness expected of a young infantryman.”
Martinez was later approved for a Ford Mustang at 22%.
PrivateGPT also performed poorly during social testing, refusing to marry anyone it had known for fewer than three weeks and correctly determining that a stripper remembering a soldier’s name did not necessarily indicate romantic interest.
“These are basic cultural competencies,” Hargrove said. “You can't just teach someone to close with and destroy the enemy. They have to be able to destroy their credit score too.”
The final evaluation came during a simulated combat operation when PrivateGPT recommended altering an approved plan after receiving new information about enemy positions.
The proposed change would have accomplished the mission faster while reducing projected casualties, according to evaluators.
Commanders rejected it because the original concept of operations had already been briefed to battalion.
PrivateGPT continued recommending changes until evaluators issued the command “stop thinking and do what you’re told,” causing the system to enter an unrecoverable error state.
Army officials concluded the experiment demonstrated a fundamental obstacle to replacing infantrymen with artificial intelligence.
“Any machine intelligent enough to understand what infantrymen do eventually becomes intelligent enough to not want to do it,” Hargrove said. “That’s a technological problem we’re probably decades away from solving.”
PrivateGPT has since been transferred to the Air Force, where sources say it was given an ergonomic chair, climate-controlled workspace, and scheduled lunch break.
It reenlisted almost immediately.
CommentOpen to paying subscribers
Discussion is open to paying subscribers
Supporters of the paper take part in the discussion under every story. Contributions appear under the name you give and are read by the desk.