It was 2 a.m. when Soo-ah couldn’t look away from the screen.

Data spread across her laptop: Starroad platform reviews. Forty-seven thousand reviews rated five stars. All written within the last thirty days. One number, extracted by the analysis tool, stopped her hands.

Sentence pattern similarity: 91.3%.

Among reviews on the same product category, 91.3 percent showed text characteristics originating from the same language generation system. Soo-ah deleted the number and ran it again. Same result.

Seven months since she joined Starroad’s Integrity Team. She’d thought it simple at first—filter out bad reviews. But Team Lead Park Jin-woo told her on day one: “Soo-ah, this isn’t a spam filter. It’s war.” It sounded exaggerated then. Not anymore.

In 2023, Britain’s Competition and Markets Authority publicly reported fake review conditions at major e-commerce platforms. The investigation classified over 30 percent of reviews on major platforms as originating from untrustworthy sources. After the announcement, the industry roiled, then quieted. Not because it was fixed. The game just got more sophisticated.

The reviews Soo-ah tracked were unlike spam of old. Sentences flowed naturally. Spelling was perfect. Fine details suggested a buyer’s life: “The smell was strong at first, but I got used to it after two days.” “My husband gave it as a gift, and the packaging was pretty.” But three things were always the same. Sentence rhythm, modifier placement, how the final sentence arranged exclamations.

It was the model’s fingerprint.

Soo-ah uploaded the report. An answer came within 24 hours. Legal department.

“We cannot suspend a seller based on unverified algorithm judgment.”

She rewrote the report. Added more data. A second answer came. Different department this time.

“This seller is a Platinum-tier partner.”

Soo-ah didn’t close the file. Instead she opened a new window. She began backtracking accounts connected to that seller.

The network was broader than expected. Twelve sub-accounts branching from one central account. Each used different IPs, different device fingerprints. Yet their activity windows overlapped. Between 3 and 5 a.m., Korea standard time. No human writes reviews at that hour. Something automated was firing.

She posted this pattern to the team internal channel on Thursday. Friday morning, some of her analysis account access was restricted. Official reason: security policy review.

Jin-woo told her quietly in the hallway. “Soo-ah, this isn’t the first time. Six months ago, someone on the team found something similar.” He didn’t continue.

“What happened to them?” Soo-ah asked.

“They left the team.”

That evening, Soo-ah backed up the analysis to her personal laptop. Only publicly available data with no legal issues. Carefully, but nothing omitted.

That night she read the reviews again. Not as numbers, but as sentences. Someone had trusted those five stars, pressed buy. Placed hope in a 50-dollar supplement. Given 30-dollar makeup as a gift. That trust was manufactured. A model generated it, an account posted it, a system allowed it.

Five stars were language. And language could lie.

Soo-ah saved the file and watched the screen. The outside was slowly brightening. Today more reviews would upload. While she slept or didn’t sleep. The analysis tool kept running.

Where the game would end, she still couldn’t tell.

This is a work of fiction and does not constitute investment advice.