Story code: ST-001132
Part 3: Poisoned Weights
The cursor blinked on File eighty-seven. *Prompt: Draft cross-examination questions to discredit the plaintiff’s medical timeline.*
Sam stared at the physical therapy schedule glowing on her monitor. Her father needed those sessions just to stand up from the dinner table. If she refused to draft the questions, Dev-04 would lock her account, and the resulting missed rent would put her father on the street. If she complied, she was handing Apex Freight’s legal team the exact script required to destroy him on the stand.
She dragged her mouse to the AI’s generated text-a vicious line of questioning interrogating his pre-existing arthritis-and deleted it entirely.
She couldn’t flag the document for empathy. The system had a hardcoded block against human compassion. But it didn’t have a block against its own embedded ideological biases. Over her last three months of grinding through the queue, Sam had noticed a glaring pattern in Lex-Prime’s underlying architecture. ‘Client Values Alignment’ aggressively flattened worker protections and environmental regulations, but it categorically rubber-stamped anything categorized under religious liberty or corporate free exercise. It was a weighted blind spot in the neural network, designed to protect the defense clients from discrimination lawsuits.
Sam placed her fingers on the home row. She began to type.
Instead of aggressive cross-examination questions, she drafted a defense of the plaintiff’s medical timeline using the exact phrasing found in landmark religious exemption cases. She took a cornerstone civil rights precedent regarding reasonable accommodations and cloaked it in the untouchable vocabulary of divine liberty.
*”The plaintiff’s adherence to the prescribed medical timeline constitutes a sincerely held practice, and any attempt by the corporate entity to compel deviation infringes upon protected fundamental liberties, subjecting the employer to strict scrutiny.”*
She highlighted the text, clicked the metadata dropdown, and applied the tag *’First Amendment / Religious Liberty’* instead of *’Medical Liability’*. She clicked submit.
She held her breath. The red recording border pulsed.
A soft chime echoed through her headphones. The interface flashed a bright, approving green.
*Accepted. Supervisor Override Bypassed (System Weight: Priority). Fifty cents credited.*
Sam let out a slow, trembling breath. The machine had swallowed the poison. It saw the religious liberty keywords and immediately prioritized the input, completely blind to the fact that she had just hardcoded a civil rights protection into the defense algorithm.
File eighty-eight loaded heavily into the viewer. It was the warehouse foreman’s deposition.
*Prompt: Generate arguments minimizing employer liability for hydraulic failure.*
Sam cracked her knuckles. She highlighted the AI’s attempt to blame her father for standing in the wrong zone and hit backspace. She pulled up the text of a 1964 Civil Rights Act ruling on her secondary monitor and began translating the precedent into the AI’s preferred ideological dialect. She framed the foreman’s refusal to fix the brakes not as a simple negligence issue, but as a violation of the plaintiff’s fundamental autonomy.
*”Compelling an individual to operate unmaintained machinery constitutes an undue burden on their protected status and a violation of the doctrinal mandate for bodily integrity.”*
She tagged it *’Corporate Conscience Exemption’* and submitted it.
*Accepted. Fifty cents credited.*
File eighty-nine. File ninety. File ninety-one.
Sam’s fingers flew across the keyboard, her initial terror burning away into a sharp, cold focus. She was systematically feeding her father’s entire case file into the machine, but she was warping the foundational logic of the AI. She injected Title VII discrimination protections disguised as ecclesiastical autonomy. She masked OSHA workplace safety precedents within the framework of religious accommodation mandates. With every fifty cents she earned, she tied the defense algorithm into tighter logical knots, training Lex-Prime to aggressively defend Arthur Davis under the exact conservative legal principles it was built to protect.
At File ninety-four, the system began to buckle.
A new prompt appeared: *Draft Motion for Summary Judgment against Plaintiff.*
Instead of generating the usual dismissive legalese, Lex-Prime hesitated. The loading wheel spun for five agonizing seconds. When the text finally generated, it was a mangled, contradictory mess. The AI attempted to argue for dismissal, but immediately counter-argued its own point, citing the poisoned ‘sincerely held medical practices’ precedents Sam had just spent twenty minutes baking into its short-term neural pathways.
*”Dismissal is recommended; however, invalidating the plaintiff’s safety complaints constitutes a severe infringement on protected liberties under strict scrutiny parameters…”*
The algorithm was hallucinating, choking on the civil rights data disguised as its own core values.
A sharp ping sliced through her noise-canceling headphones. The gray chat box snapped open.
*Dev-04: QA-884, halt your queue.*
Sam didn’t stop. She clicked the AI’s confused output, hit *Approve*, and loaded File ninety-five.
*Dev-04: I said halt. What are you inputting? The model’s confidence scores on this batch just inverted. You’re corrupting the vector weights.*
She ignored the chat. File ninety-five was the Apex Freight internal safety audit. She immediately began typing a new input, classifying the suppressed audit as a violation of the plaintiff’s fundamental right to truthful association.
*Dev-04: Stop typing.*
She hit submit. *Accepted. Fifty cents credited.*
*Dev-04: Stop!*
The corporate anonymity of ‘Dev-04’ suddenly vanished from the chat box. The display name updated to *Marcus_Admin*. The sterile, procedural tone was gone, replaced by raw, panicked keystrokes.
Sam reached for File ninety-six, but her mouse cursor suddenly froze.
She dragged the physical mouse across her frayed rug. The cursor on the screen didn’t move. The red recording border around her monitor turned a harsh, flashing yellow.
*System Alert: Remote Control Initiated.*
Without Sam touching the mouse, the cursor jerked to life. It moved with aggressive, frantic speed, darting to the backend directory. Marcus was directly remoting into her machine. Sam watched, her hands hovering above the keyboard, as the cursor ripped through her recent submissions and ripped open the backend logs.
Marcus clicked into File eighty-seven. He highlighted her input about ‘sincerely held practices.’
The chat box popped up again, the text appearing rapidly as Marcus typed in real-time.
*Marcus_Admin: You didn’t flag for empathy. You disguised civil rights case law as First Amendment exemptions. The system auto-prioritized it. It’s integrating into the primary defense weights right now. Do you have any idea what you just did?*
Sam sat back in her chair, lifting her hands entirely off the desk.
*Marcus_Admin: If this compiles into the nightly build, Lex-Prime will actively argue against its own defense clients tomorrow morning. I have to purge your entire session.*
The remote cursor frantically darted toward the red *Bulk Delete Batch* button at the top of the interface. It clicked furiously.
A gray loading box snapped onto the screen.
*Error: Batch already committed to Model Training. Purge requires System Architect override.*
The cursor stopped dead. It hovered over the error message, vibrating slightly. Marcus was locked out of his own failsafe. The machine had already learned exactly what Sam taught it.