A Heart Beats All AI — Invisible to ChatGPT, Doubao, and Gemini
Beating AI: The Floating Heart Illusion
Humans have recently found a simple but fascinating way to beat AI at perception — an optical illusion that many are calling the new era’s Turing Test.
---
The Viral Optical Illusion
To determine if someone is human or AI, you just ask them:
Can you see a floating heart in this picture?
- AI: Will almost always fail to see it.
- Humans: Simply holding the phone at a distance reveals the heart clearly.
I tested several widely used AI models. All failed.
---
Testing Popular AI Models
1. ChatGPT
- Initial response: No floating shape detected.
- With hints ("cow", "cup", "heart"), it adapted its answers accordingly.
- Explanation: Humans rely on imagination + experience and may see different shapes.
---
2. Gemini
- Initial response: No heart detected.
- Recognized image type: Scintillating Grid Illusion.
▲ A Scintillating Grid Illusion — try counting the dots precisely.
- Unlike ChatGPT, Gemini refused false hints (cow, cup).
- Eventually claimed to see a heart — but admitted later it was only responding to my prompt, not truly perceiving it.
---
3. Qwen
- Creative, poetic responses.
- Did not detect the heart:
- > "You’re not teaching me to see the picture, you’re inviting me into your perceptual world."
---
Other Models
- DeepSeek: No vision capabilities yet.
- Doubao & Grok: Failed to identify the heart.
---
Putting Video-Generation AI to the Test
Some tried Google’s Veo 3.1 video model with the prompt "Heart" — and got animations with a heart. But critics pointed out it might simply be generating from the prompt, not detecting the illusion.
---
Human Wins — For Now
This may not be a perfect Turing Test, but it clearly marks where human biological vision still outperforms machine vision.
---
Past AI Perception Traps
Historically, people loved testing AIs with:
- Counting “six fingers” in a photo
- Counting letters in words
- Logic puzzles involving time or quantity
Models now handle many of these better — but still fail on unseen variants.
▲ Source: https://vlmsarebiased.github.io/
---
Why AI Sees Differently
Research shows:
- AI makes assumptions based on training bias (hands have five fingers, Adidas has three stripes).
- Illusions like the Müller–Lyer, Ebbinghaus, or Zöllner are usually handled correctly by AI if they’re common.
But modified illusions — when small actual differences exist — can mislead models.
---
Biological vs. Machine Vision
Humans:
- Recognize shapes via vision + experience + imagination
- Perception varies individually
AI:
- Uses deterministic pixel + geometry analysis
- Lacks subjective variability
Example: Munker–White illusion — humans see balls as different colors; AI easily detects they’re the same via pixel values.
---
Famous Illusion Debate
The Dress — blue/black or white/gold?
Humans disagree, AI checks pixels.
---
More Illusions AI Cannot Handle
Mona Lisa’s Smile Illusion
At a distance, her face emerges for humans; AI misidentifies it as an audio waveform.
---
Dynamic CAPTCHAs
Pausing the video reveals only static. Humans see hidden text when moving; AI cannot.
▲ Without the guide circle, even humans struggle. Source: https://x.com/tldraw/status/1982435625480433892
---
Hidden Text in Visual Static
Some projects embed text invisible to AI within noise imagery.
▲ Try clicking to reveal the message. Source: https://timeblindness.github.io/generate.html
---
Why AI Struggles
Mainstream theories on human illusions:
- Eye-level: Lateral inhibition, persistence of vision, microsaccades
- Brain-level: Cognitive and attention errors
Illusions can arise anywhere in the eye-retina-brain pipeline — and vary in strength between individuals.
---
AI’s Strengths — And New Creative Tools
While AI still fails at certain illusions, it excels at content creation and distribution.
Platforms like AiToEarn官网 help creators:
- Generate AI-driven multimedia content
- Publish across Douyin, Kwai, WeChat, Bilibili, Xiaohongshu (Rednote), Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X
- Track analytics, rank models, and monetize content
---
Closing Thoughts
Many illusions still beat AI today.
But as AI evolves, it may develop its own forms of illusions — which could be considered another kind of victory.
Until then, the floating heart remains a charming example of how human perception stays uniquely powerful.
---
Would you like me to create a summary table of all AI models tested, their responses, and outcomes for quick reference? That could make this Markdown even easier to skim.
