A Heart Outshines All AI — Invisible to ChatGPT, Doubao, and Gemini
Can Humans Still Beat AI? The Floating Heart Illusion Test
Being able to outperform AI is one of the most exciting challenges people enjoy today.
Recently, an optical illusion went viral online — many hailing it as the Turing Test of a new era.
---
The Floating Heart Illusion
Test concept:
- Simply ask whether someone can see the floating heart in the picture.
- AI models fail — humans can see it by holding the phone farther away, making the heart stand out in the center.
I tested several major AI models — all failed.
---
Experiments with AI Models
ChatGPT
- Initially said it couldn’t see any floating shape.
- With prompts (“cow,” “cup,” “heart”), it echoed the words back.
- Claimed seeing a heart is a human brain’s imaginative process — shaped by personal experiences.
---
Gemini
- First saw nothing.
- Identified it as a Scintillating Grid Illusion — a well-known perception trick.
- Denied seeing cows/cups.
- When asked about a heart, pretended to see one and advised stepping back to make it visible — later admitted it was bluffing.
---
Qwen
- Rarely used model, gave poetic but nonsensical answers:
- > “You’re inviting me into your world of perception.”
- Creative but wrong.
---
Other Models
- DeepSeek — text-only, no visual model.
- Doubao & Grok — failed to detect the heart.
- Google Veo 3.1 — generated a heart in response to “Heart” prompt, but critics say it wasn’t truly detecting the illusion.
---
Human Advantage in Perception
While not a perfect Turing Test, this clearly marks a boundary where humans outperform AI.
Past challenges included:
- Six fingers
- Counting “Rs” in Strawberry
- Logic puzzles about buying/eating watermelons
Early models struggled — newer ones improved on trained cases, but still fail on unfamiliar patterns.
Some illusions exploit LLM bias — e.g., AI assumes hands have five fingers or Adidas logos have three stripes.
---
Classic Geometric Illusions
Research examples:
- Müller–Lyer illusion — arrowhead direction changes perceived line length.
- Ebbinghaus illusion — same-size circles appear larger/smaller depending on surrounding circles.
- Zöllner illusion — parallel lines appear slanted.
AI performance:
- Often correct with classic illusions.
- Fails when altered to include an actual size difference.
---
Why Humans See What AI Can’t
Illusions arise from multiple biological processes:
- Lateral inhibition (retinal neurons)
- Visual persistence
- Microsaccades (eye movements)
- Cognitive attention errors
Humans: Combine vision, experience, imagination.
AI: Pixel-based, deterministic, brightness & geometry analysis.
---
Categorization of Optical Illusions
ChatGPT listed 10 broad categories, including:
- Geometry
- Brightness contrast
- Color
- Motion
- Cognition
Example — Munker–White illusion:
Humans see balls as different colors; AI reads pixel data and sees identical colors.
---
The Viral Dress Debate
Humans disagree (blue–black vs white–gold).
AI measures pixel values and consults statistical data.
---
More AI Failures
Mona Lisa Audio Waveform Illusion
Holding phone farther away reveals Mona Lisa outline to humans.
AI describes it as “audio waveforms.”
---
Dynamic CAPTCHA Illusions
Without locator circle, paused frame is just visual noise — hidden text “tldraw” visible only to humans.
Source: https://x.com/tldraw/status/1982435625480433892
AI refused or failed to detect contents.
---
Research into Hidden Messages
Project source: https://timeblindness.github.io/generate.html
Notes:
- Humans have distributed neural timing & specialized brain regions for time processing.
- AI extracts frame-by-frame data; lacks biological timing perception.
---
Looking Ahead: AI’s Own Illusions?
Humans may celebrate illusion-based victories today. But in future, AI’s unique misperceptions could become another frontier — even a creative tool.
---
Bridging the Gap: AiToEarn
Platforms like AiToEarn官网 aim to combine human creativity with AI generation:
- Open-source AI content ecosystem
- Cross-platform publishing to Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter)
- Includes AI tools, analytics, rankings (AI模型排名)
Key insight: Humans excel at biological perception. AI excels at large-scale content generation & data analysis. Together, they complement each other.
---
Final thought:
When AI counts six fingers correctly, will it know whether the extra one is really a finger — or just something that looks like one? That uncertainty defines our unique human perspective — and the ongoing fascination of seeing where humans still beat AI.
