Model Benchmarking, Part 1
AI Ethics. 1,700 words, 8 minute read. With Gemini.
The Vance & Carter Literary Agency perpetually smelled of aging paper and strong, slightly burnt coffee. Sunlight, sliced thin by dusty blinds, spotlighted the motes dancing above Ben Carter’s side of the office – a landscape of teetering manuscript stacks and the faint blue glow of dual monitors. Across the divide, Eleanor Vance’s desk was an island of meticulous order: pens aligned, contracts squared, a single, cherished philodendron reaching towards the light. She was carefully re-reading a marked-up first chapter, the heft of the paper familiar and reassuring in her hands.
Ben’s rapid-fire typing ceased. The sudden silence made the hum of the server rack in the closet seem louder. “Well,” he murmured, leaning back, chair springs sighing. “Didn’t see that coming down the pipe quite so… definitively.”
Eleanor looked up, marking her place with a silver bookmark. “Henderson finally cave?”
“Bigger leagues,” Ben replied, rotating one monitor slightly. The reflection caught the sharp lines of a formal announcement. “SFWA. Worldcon. Joint statement.” He paused, letting her anticipate. “AI-assisted work. Ineligible. Hugos, Nebulas. Done.” A small frown touched his lips as he added, almost to himself, "Good luck policing that."
A flicker of something – satisfaction, vindication – crossed Eleanor’s face. She straightened the manuscript she held. “Good. It clarifies things.” Her fingers traced the embossed title. “These awards should celebrate human ingenuity, not… clever mimicry.”
Ben pushed his glasses up, peering over the rims. “And the mechanism for clarification is…?” He let the question hang. “Are they installing detectors at the slush pile gates? What’s the threshold for ‘assistance’? Does using an AI to check historical accuracy count? What about brainstorming prompts? Have they even defined 'substantial' yet?"
“Ben, please,” Eleanor said, her voice tight with a hint of impatience, though she avoided his gaze, focusing instead on aligning the manuscript perfectly with the edge of her desk blotter. “It’s about the writing itself. The voice. Surely, we can still recognize authenticity.”
“Can we?” Ben’s voice was soft, less argumentative than genuinely questioning. “Or are we just going to get really good at recognizing authors who sound like they didn’t use AI? Think about the debut novelists, the ones experimenting right at the edge. Who defends them when the first whisper campaign starts?”
Eleanor picked up her pen, its cool weight a familiar anchor. “We trust the process. We trust the committees.” But a faint line appeared between her brows as she looked towards the window, the city outside a vast network of invisible connections. Ben’s question, and his muttered doubt about enforcement, echoed in the quiet hum of the office.
Summer pressed down on the city, the office air conditioner waging a losing battle. Six months after the ban, the agency felt different. Less a hub of creative discovery, more a cautious listening post. Phone calls were quieter, doors sometimes closed. Ben forwarded Eleanor a chain email circulating among agents – conflicting interpretations of the ban’s fine print, rumors of committee members resigning over impossible workloads, whispers of retracted guidelines. He’d already demonstrated a new detection software by feeding it the opening paragraphs of Ulysses, only to watch it flag Joyce’s stream-of-consciousness with a confident ‘78% Probability AI Origin’. He’d logged off without comment, the faint click of the mouse echoing his frustration. Yesterday, he'd mentioned hearing it flagged a senator's own memoir as AI-generated.
The call about Maya Singh felt like the inevitable crack in the dam. Synaptic Drift, her brilliant, challenging debut, was on the Nebula reading list – a triumph Eleanor had celebrated with champagne just days before. Now, she found herself gripping the phone, listening to Maya’s thin, trembling voice recount the accusations swirling on a niche sci-fi blog.
“Statistical anomalies in thematic clustering?” Eleanor repeated, keeping her own voice level, professional, despite the anger tightening her chest. “Maya, that’s academic white noise, not evidence.” She paused, hearing the sharp intake of breath on the other end, a sound like tearing silk. “No, absolutely do not engage. Don’t reply, don’t even read it… Because it lends them legitimacy. Send me everything – notes, drafts, outlines. We fight this with facts.”
She hung up, the silence heavy. On the video call earlier, she’d focused on Maya's trembling lower lip as she tried to form words, the way she kept pushing her hair back from a face pale with stress. Ben held out his tablet, already open to the offending blog post. Anonymous commenters, emboldened by the ban, were dissecting Maya’s prose, calling her intricate world-building ‘too seamless for human error’.
“They’re citing parallels to obscure texts only available in digitized archives known to be training fodder for LLMs,” Ben said quietly, scrolling through the venomous comments. “It’s a reach, Eleanor. A hell of a reach. But it’s out there. And apparently, the committee is 'investigating' based on this." The word hung in the air, tainted.
“It’s character assassination disguised as literary criticism,” Eleanor retorted, standing abruptly. She paced the small space, the sharp click of her heels a counterpoint to the drone of the air conditioner. “She researched meticulously! She layered her themes! Now innovation is suspect?” She stopped, pressing her fingertips against the cool glass of the window. “How do we even prove she didn’t use an AI? Install spyware on her laptop?” The sheer, invasive impossibility – and the lack of any clear, fair process from the award bodies – felt suffocating.
Ben leaned forward, elbows on his knees. “And she’s not the only one. Agents are quietly asking clients for ‘process statements’. Everyone’s spooked.” The burnt-coffee smell seemed stronger today, mingling with an undercurrent of something else – systemic failure.
Weeks later, the storm around Maya had passed, leaving behind a residue of suspicion that clung like damp air. Her name remained on the list, but the joy was gone. Eleanor found herself staring at a new submission – a dazzlingly original piece of speculative fiction, unlike anything she’d read. The prose wasn't just flawless, it sang – imagery sharp as cut glass, emotional beats landing with uncanny precision. She read a paragraph describing a zero-gravity garden, felt a genuine pang of awe, and then, immediately, a cold wave of doubt. Is it too perfect? Could a human write this? The question, involuntary and unwelcome, soured the beauty of the words. She pushed the manuscript away, annoyed at her own conditioned skepticism.
“It’s the chilling effect,” she said aloud, swirling the dregs of her lukewarm tea. “We’re making writers afraid to be brilliant. Afraid their best work won’t seem… human enough.”
Ben looked up from a contract negotiation. “Or we’re asking committees to be infallible mind-readers on top of literary critics – using tools that think Joyce and senators are robots.” He recalled the Finnish manuscript incident without needing Eleanor to prompt him. “Remember that piece? Flagged ‘high probability’ because the author’s syntax wasn’t standard American English. We almost passed on genius because it didn’t fit the machine’s profile.” He shook his head. “What happens next year? And the year after? Do we just keep feeding the paranoia while the actual tech gets even harder to detect?”
Eleanor considered the manuscript she’d pushed aside. Its pages felt smooth, almost unnervingly uniform. What if the author had used AI, not as a crutch, but as a partner, to achieve this level of polish, this resonance? Did the how invalidate the what – the story's power, its beauty? The question felt heretical, yet unavoidable. “We wanted to protect the source,” she murmured, tracing the title with a fingertip. “But maybe we’ve just muddied the waters so much no one can see clearly anymore.”
“Perhaps the definition of ‘source’ is expanding,” Ben suggested gently. “If a story connects, if it moves someone, changes their perspective… does the reader care if the author had a collaborator made of silicon instead of just caffeine and neuroses?”
The email arrived mid-morning, its subject line stark against the usual clutter of queries and contracts: Introducing the Anthropies. Ben spotted it first. “Well now,” he said, a note of surprise in his voice. “Looks like someone’s building a new wing on the museum. Or maybe just acknowledging the existing architecture.”
Eleanor leaned over as Ben forwarded it. She read the announcement slowly, her posture straightening almost imperceptibly. A new award, backed by a coalition of publishers, tech groups, and authors known for pushing boundaries. Explicitly for human-AI collaborative fiction. Transparency required. Process statements mandatory. Judging based on the final work’s merit and the innovative nature of the collaboration. It was, she realized, a formal concession – an admission by a significant part of the industry that the old definition of singular human authorship was no longer sufficient to capture the landscape.
“The Anthropies,” Eleanor repeated the name, testing its weight. It felt… engineered. Necessary, perhaps. A pragmatic pressure valve for a system threatening to explode. It wouldn’t erase the suspicion surrounding the established awards overnight, but it created a legitimate, acknowledged space.
Ben watched her, waiting. “So? Progress or partition?”
Eleanor didn’t answer immediately. Her gaze drifted to the dazzling, potentially problematic manuscript still sitting where she'd pushed it. She picked it up, feeling its physical weight, the smooth texture of its cover. Instead of putting it back on the main pile, she placed it deliberately on the corner of her otherwise clear desk – a space reserved for things requiring further thought, things that didn't fit neatly into existing categories. It represented the future, in all its complicated, ethically ambiguous, undeniably compelling potential. The low hum of the server rack filled the quiet office, a steady digital pulse beneath the rustle of paper, indifferent to the shifting definitions of creation. The story was changing, whether they were ready or not, and the old maps no longer applied.