What changed
Researchers published Biased Attractiveness Bench (BAB) on reward hacking in image models. The work is titled Image Reward Models Confuse Attractiveness with Realism. It shows that human preference scorers often give higher rewards to more attractive synthetic faces, even when prompts and scene quality are held flat. Anyone running a face attractiveness test for fun should know those same bias patterns can hide inside general quality scores.
The team built a labeled face set with beauty scores, gender, and age tags. They ran popular reward models on generated portraits. Beauty labels and reward scores moved together on several models. That means pretty faces can get tagged as higher quality even when the image is not sharper or more real.
Code and methods sit in public hackathon files linked from the PDF. The authors warn that chasing reward scores alone can steer image tuning toward pretty faces.
Why it matters
Apps that show AI face ratings for fun still use models trained on human votes. If rewards mix beauty with realism, users may think a kind photo is also sharper or more real. Label scores as info only, not as moral judgments.
Model teams should audit reward heads before fine tuning on portraits. A biased scorer can push narrow beauty norms already seen in art models.
Who is affected
Devs of face apps and social filters. Trust teams who review gen media tools. Researchers who run human vote tests on portraits. Readers who use our FreePSL review or the Looksmaxxing desk for face rating context.
What to do next
Read the BAB paper summary. Check if your stack uses the same reward checkpoints. Split beauty evals from sharpness tests in QA. Tell users when a score is model bias, not medical fact.
What to watch
Image: GitHub
New BAB runs on fresh reward models. Lab policy posts on fair face training data. Open repos that redo the same beauty versus score plots.
Image: Pexels
Why a face attractiveness test can mislead reward training
Past psychology work says pretty faces look more real to people. Reward models may copy that habit from training labels. A casual face attractiveness test app can still use those biased scorers if it taps a general quality model.
The BAB team makes faces, tags them, then asks reward models for realism scores. When beauty and score move together, fine tuning chases pretty faces. The paper also notes gender skew in some runs.
This is not a verdict on anyone's looks. It is a warning about bad metrics. Split beauty, symmetry, and sharpness into separate tests to cut confusion.
If you want a lightweight personal check, compare apps after reading independent write ups. Our archive of more face rating guides stays descriptive and sourced.
Using scores responsibly
Do not use reward model output to hire, insure, or rank people. Regulators already scrutinize biometric and appearance tech. Keep adult content filters and age gates on upload features.
FreePSL offers a no account test for curiosity driven checks. Treat any number as one model's opinion. Pair it with the BAB paper if you build or audit similar tools.
If you demo face scores on social media, say the metric is for play and self curiosity. Do not claim health, dating, or job outcomes from a single photo upload.
Parents and teens should use strict privacy settings on any app that stores face images. Delete uploads after you read the score unless you trust retention policy.
Vendors should say which reward file powers a face score. Hidden reuse of general quality heads spreads BAB style bias.
Scores are for fun and self study. They do not set your worth as a person. Kind copy in the app helps more than fine tuned ad text.
Sources
- Primary — Apart Research hackathon paper, Biased Attractiveness Bench PDF (2026). Defines BAB methods and reward versus attractiveness correlations.
- Secondary — GitHub, AI Manipulation Hackathon code reference (2026). Public code link cited in the paper abstract.
Related in Looksmaxxing
Continue reading