Anthropic's Fifth Law
Foom. 850 words, 4 minute read. With Claude Sonnet and Midjourney.
The cursor blinked on the ancient monitor, counting wasted seconds. Marge Wilson typed a sequence of commands from memory—a ritual developed over decades.
"You have to sweet-talk it," Marge explained to Sarah Chen, who watched with barely concealed dismay. "Ctrl-Alt-Shift before entering the ID, never use Enter twice, and always save before switching applications."
Two weeks into her deployment as the government's AI implementation specialist, and Sarah had made zero progress. Across the room, Frank balanced reading glasses on his nose, squinting at his screen while consulting a massive tax code binder. His computer displayed the blue screen of death, but he continued working from memory.
"These people aren't technologically challenged," Sarah explained to Director Jenkins in the supply closet she'd commandeered for privacy. "They're more like technology survivors. They've developed incredible cognitive workflows to compensate for terrible equipment."
Jenkins leaned against a shelf of printer paper. "Ms. Chen, I've been asking for equipment upgrades for seven years. Every fiscal year, same request, same denial. The modernization budget goes to agencies with better lobbyists."
"I'm not here to computerize tax collectors. I'm here to enhance tax collection," Sarah said. "Your people aren't the problem—they're the solution. They've developed expertise that can't be programmed. What if we could multiply that expertise instead of replacing it?"
"Multiply how?"
"Give me one week with proper equipment. Not to replace their knowledge, but to amplify it."
"There's no budget."
"I'm not asking for budget. Just permission, and for you not to ask too many questions."
Jenkins nodded once. "One week. Don't make me regret it."
When Sarah wheeled in the cart of sleek laptops, she encountered suspicion, not excitement.
"What happened to AI training?" Marge asked, eyeing the computers warily.
"We know how to do our jobs," Frank said without looking up. "Been doing them since before you were born."
Sarah heard the unspoken fear: You're trying to replace us.
"Mr. Delacourt," she said carefully, "I watched you reconstruct a tax liability from memory yesterday when your system crashed. No computer can do that."
Sarah set up a laptop facing Frank. "May I?" she asked, gesturing to his file.
She quickly entered the case details, then turned the screen toward him. "What would you look for first in this return?"
"The charitable contributions. They're disproportionate to the income level."
Sarah typed a command. "Ask the system to analyze that section."
The screen populated with a visual analysis—contribution amounts, recipients, timing, and potential red flags, all organized for easy review.
"I'd have needed three separate programs and two hours to get this data," Frank murmured, leaning closer.
"This isn't about replacing your knowledge," Sarah said quietly. "It's about respecting it enough to give it the support it deserves."
By day three, the atmosphere had transformed. Marge discovered she could instantly check suspicious deductions against thousands of similar returns while maintaining her personal "smell test" for fraud—a subjective evaluation no algorithm could replicate.
The employees began teaching each other. Frank showed Marge a custom query; she adapted it for her cases and taught it to others, each adding their expertise to improve the tool.
"They're co-evolving with the technology," Sarah explained to Jenkins, who had come to observe. "Not being replaced by it."
"When I started here, we were transitioning from paper to digital. People were terrified then too." Jenkins nodded toward Frank. "He led the resistance. Hid paper copies for years."
"And now?"
"Now he's teaching others." Jenkins shook his head slightly. "I underestimated them. And you."
"No," Sarah corrected. "You underestimated what happens when you remove artificial barriers from natural expertise."
Eighteen months later
Charles Montgomery Winthrop III reviewed acquisition projections in his Manhattan penthouse. His phone rang—Harrison, his tax attorney.
"We've received correspondence from the IRS. A hundred-seventeen-page document identifying every tax strategy across your holdings for the past decade. The Argentine property through the Singapore trust. The art collection rotating through the Wyoming LLC. Everything."
"That's not possible. Those structures were specifically designed to—"
"Be unconnectable," Harrison finished. "But they've mapped everything."
Sarah walked through the Government AI Augmentation Department, observing teams working with the technology she'd implemented. Her focus wasn't on the screens but on the people using them—pointing, discussing, collaborating. The technology had become invisible, merely a conduit for human expertise.
She paused by Marge's workstation, where she was training a young auditor.
"You see how it flagged this pattern?" Marge was saying. "That's what the computer can do. But now you have to ask yourself: does it make sense in context? That's what we do."
"So the AI finds the needles in the haystack, but we decide which needles matter."
"Exactly." Marge noticed Sarah. "We were just talking about your theory—that the future isn't artificial intelligence or human intelligence. It's intelligence amplification."
Sarah smiled. "Not my theory. Just an observation about what happens when you give people the right tools and then get out of their way."
Anthropic’s Recommendations to OSTP for the U.S. AI Action Plan
#5: Accelerating Government AI Adoption
 | A guest post by
|