The Attractiveness Illusion: Why AI Inflates Appearance RatingsArticle · digest for November 2
Digital smartphone interface against a dressing-room mirror showing a male portrait analysis and a rating distribution chart
Article5 sections≈ 5 min

The Attractiveness Illusion: Why AI Inflates Appearance Ratings

Every guy who has ever tried to get an objective read on his look has run into a shortage of honest feedback. Friends joke it off, acquaintances hand out polite compliments, and the mirror changes its verdict depending on the lighting angle. Enter multimodal AI. It seems like the ideal referee: a cold algorithm, gigabytes of data, zero bias. You upload a portrait to ChatGPT or Claude and get the verdict: "Great facial structure, masculine jawline, a solid 7.5 out of 10."

Your confidence shoots up, you put the photo on your dating profile... and get radio silence. Why does artificial intelligence paint a picture that shatters on contact with real-world dating?

The research lab QOVES, led by Shafee Hassan alongside scientists from Oxford and the Max Planck Institute, put this paradox to the test. They compared 2,513 ratings from multimodal large language models (MLLMs) against real human perception. The takeaway is sobering: commercial AI models systematically flatter the user.

The QOVES Experiment: How Algorithmic Politeness Breaks the Bell Curve

In reality, human attractiveness follows a normal Gaussian distribution. If you show photos of random guys to an independent sample of women, the graph forms a classic bell curve. Most people—around 68% of the sample—fall into the average range of 4 to 6 out of 10. Scores of 8–9 go to the rare few with standout genetics and symmetry, while scores below 3 are reserved for faces with pronounced anomalies. That's the biological baseline.

When QOVES researchers fed the exact same 2,513 portraits to commercial models, the bell curve vanished. The models refused to give low scores. AI rated nearly every face in the 5.5 to 8 range. No twos or threes. The algorithm acted like an eager salesperson praising every jacket you try on.

The reason lies in the training architecture. Commercial chatbots go through RLHF (reinforcement learning from human feedback). Developers bake in strict filters against rudeness. A blunt response like "your looks are a 3 out of 10" will offend the user and drive them to a competitor. So AI was trained into sycophancy: smooth over the flaws, dish out a dose of positivity, and tack on a couple of courtesy points.

The Halo Effect in Machine Vision: When Looks Substitute for Intelligence

Flattery is only half the problem. Far more serious is the issue uncovered by Aditya Gulati's team in their study "Beauty and the Bias." The researchers tested 7 multimodal architectures across 924 portraits in various social and professional scenarios.

It turned out that AI models are vulnerable to the Halo effect. In 86.2% of scenarios, physical appearance skewed the algorithm's decision. And in 94.8% of relevant cases, attractive faces were automatically credited with high intelligence, professionalism, trustworthiness, and kindness.

This creates a paradox of "algorithmic lookism." A guy expects AI to give an objective assessment of his skills or resume, while the model behaves like a layperson: pleasant features in a photo automatically convert into a label of "leader and professional." The model operates on surface-level stereotypes from its training data.

Geometry vs. the Chatbot: Why Language Models Are Blind to Facial Proportions

To understand why you shouldn't trust an LLM with men's styling, you need to look at optics. A smartphone front-camera shot is distorted by the wide-angle lens (lens distortion): the nose looks larger and the face looks narrower. Professional morphometric algorithms calibrate the frame using anatomical constants.

The primary facial landmark is the horizontal corneal diameter. In adults, it remains stable regardless of height or ethnicity—averaging 11.7 mm with a variance of less than 0.5 mm. Anchoring to this measurement, specialized plastic-surgery software converts the pixels between the eye canthi and jaw angles into real millimeters.

Language models have no spatial reasoning apparatus. They predict text based on token probabilities. In a study by Naiter et al., geometric measurements via multimodal LLMs showed zero agreement with clinical data. Feed the model a portrait of Quasimodo, and out of politeness the chatbot will find "expressive bone structure" in it.

The Digital Glow-Up Trap: From Generic Advice to Identity Collapse

Asking an AI to build a men's glow-up plan yields a pile of clichés: drink water, sleep eight hours, hit the gym. These canned lines don't explain what specifically needs fixing in your appearance.

The model can't spot chin retrusion, masseter hypertrophy, or bite alignment issues. In aesthetic practice, up to 0.7% of male facial issues are solved through orthodontics or posture rather than creams. Yet none of that nuance shows up in LLM replies.

Trying on hairstyles through image generators is risky too. It triggers identity collapse. When swapping details, diffusion noise alters skull proportions: skin turns plastic (the uncanny valley), and a stranger's face stares back from the screen. Trying to replicate that digital mirage at the barbershop is pointless.

Use a Blind A/B Comparison Instead of Asking for a Score

Upload two photos of yourself in different outfits or hairstyles in the same prompt and ask the model to compare garment fit and silhouette harmony without attempting to assign attractiveness scores.

Practical Protocol: How to Extract Sober Style Value from AI

You can turn AI into a useful tool if you follow three rules.

First, run a blind A/B comparison test. Upload two shots in different styles and ask neutrally: "Compare the shoulder seam fit and torso proportions in both photos. Which silhouette looks sharper? Back up your conclusion with facts." When comparing two options, politeness doesn't stop the model from giving a solid breakdown.

Second, block compliments in the prompt: "Act as a strict technical stylist. No compliments or praise. Break down garment fit mistakes and color clashes."

Third, test one detail at a time: eyeglass frame shape or jacket shade against your skin undertone, without trying to redraw your whole face.

Male attractiveness is built on posture, fitness, a sharp haircut, and inner calm. Working on the anatomy of attractiveness and skills from first-impression psychology will pay off far more on real dates than any virtual rating.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 2