Characteristics of Fake Reviews

Unnatural reviews have several common patterns:

1. Contradiction between playtime and content

"This game changed my life! Masterpiece!" Playtime: 0.3 hours

Reviews that give excessive praise with extremely short playtime like this are likely fake.

2. Reuse of boilerplate text

If the exact same phrases are used in multiple reviews, there is a suspicion of organized manipulation.

3. Unnatural language

Reviews with unnatural expressions typical of machine translation or misuse of particles should also be noted.

AI Detection Logic

This tool detects fake reviews in the following steps:

graph TD
    A[Collect Reviews] --> B[Metadata Analysis]
    B --> C[Gemini AI Context Analysis]
    C --> D[External Source Check]
    D --> E[Calculate Score]

Step 1: Metadata Analysis

• Playtime
• Account creation date
• Number of games owned
• Number of reviews posted

From these numbers, a heuristic score is calculated.

Step 2: AI Context Analysis

Give the following instructions to Gemini AI:

"Detect 'Sakura' (Fake/Paid/Manipulated) reviews from both provided Steam reviews AND external sources found via Google Search."

The AI deeply understands the context of the review text and detects unnatural patterns.

Step 3: External Source Check

Compare with ratings on Reddit, Metacritic, and major game media to see if the evaluation within Steam is unnaturally divergent.

Ingenuity for Accuracy Improvement

In the initial version, there were many false positives. In particular, there were cases where legitimate high ratings by enthusiastic fans were mistaken for fakes.

Therefore, the following improvements were made:

1. STRICT FILTER: Ignore general praise like just "Best!"
2. Check for specificity: Check if specific elements of the game (story, mechanics) are mentioned
3. Consistency with external evaluations: Check if there are similar evaluations on other platforms

As a result, we were able to reduce the false positive rate by about 60%.