Skew

Skew had not always understood her own lean. She had been folded in a paper-crafts workshop in a river village, alongside her friends Sort and Feed, and for her first folding-years she had felt simply wrong — like a flaw someone had shrugged over and set on a shelf.

The trouble began the day a customer complained. A little sorting-machine the workshop had built kept dropping the same kind of card into the wrong bin, over and over, and the customer was furious. "Your machine is unfair," he said. "It has something against these cards." The workshop elders fretted. They oiled the machine. They re-folded its gears. Nothing helped.

It was old Feed who finally understood. Feed spread out the training cards the machine had learned from — the examples it had studied to decide what went where — and laid them beside Skew. And Skew, without being asked, tilted. Hard. The stack was lopsided. Almost every card in it was one kind; the other kinds were barely there.

"The machine isn't cruel," Feed said quietly. "It learned from this. It only ever saw a leaning pile, so it learned to lean." She rested a hand on Skew's frame. "You've been telling us this all along, little one. We just never looked at you."

For the first time, Skew's tilt did not feel like a defect. It felt like a warning that had been trying to be heard. The hollow, blamed feeling she'd carried lifted, and in its place came something steadier — the quiet worth of a tool that does exactly what it was made to do.

She rolled to the AIForge academy when she was twenty-two folding-years old, because an academy full of people building thinking-machines was exactly the place most likely to build a leaning one by accident.

Bit, the academy's founder, met her at the door and asked the single question she asked every new teacher. "Skew — what is AI bias?"

Skew did not give a definition. She set a lopsided stack of cards on her low pan and let Bit watch it sink. "A machine that learns from a leaning pile learns to lean the same way," she said. "The machine is not the villain. The examples leaned before the machine ever saw them. So the machine copies the lean, faithfully, the way a mirror copies a face." She looked up, her pans still swaying. "People want to ask what went wrong inside the machine. The better question is who chose these examples, and whose examples are missing. That is where the lean begins. Fix it there, and you fix it everywhere."

Bit studied the sunken pan for a long moment. Then she nodded slowly. "You belong here," she said. "This whole academy builds clever things and never once checks what they learned from. You are the one who'll make them look."

Skew's classroom filled, over the seasons, with students who arrived confused and a little frightened by the word bias.

A girl named Maya came in one afternoon, arms folded. "My cousin's phone can't recognize her face in the dark," she said. "It works fine for me. Is the phone... prejudiced?"

Skew rolled to the workbench and unfolded her scales flat where everyone could see. "Let's trace it," she said. "Not guess — trace. When they taught that phone to recognize faces, whose faces did they show it most?"

Maya thought. "Lighter faces, probably. If that's who made it."

Skew placed a thick stack of cards on one pan. It plunged. "So the pile leaned." She tapped the sunken side. "The phone became wonderful at the faces it saw again and again, and clumsy at the faces it barely saw. It isn't choosing to fail your cousin. It's copying a lopsided pile of examples. Same lean in, same lean out."

A boy named Leo frowned. "But what if they never even used skin color? What if they were careful?"

"Ah." Skew's pans lifted with something like delight. "Then you look for a stand-in. Sometimes a machine leans through a side door. It might never be told a person's neighborhood matters — but if it learns from where people live, and neighborhoods are unfair to begin with, the lean sneaks in wearing a different coat." She swayed. "So we test it on everyone. If it works beautifully for one group and badly for another, the tilt is showing, no matter how careful anyone meant to be."

"So you can always find where it came from?" Maya asked.

"Almost always," said Skew. "Who gathered the examples. Who labeled them. Whose stories filled the pile and whose were left out. Those are facts about people, not mysteries about machines. And facts — " she rebalanced her scales a careful notch " — facts you can fix."