By San José, Costa Rica — The global race to perfect artificial intelligence has hit a fundamental evolutionary speed bump, according to a groundbreaking study published in the scientific journal iScience. While tech giants pour billions of dollars into deep learning models, new research reveals that these complex systems still struggle to recognize everyday objects when their overall shapes are slightly distorted. The gap between human perception and computer vision remains vast, raising critical questions for industries reliant on automated visual technologies.
The extensive study benchmarked more than 200 deep neural networks, exposing a systemic limitation in how machine learning processes visual data. Unlike humans, who rely on a robust combination of overall silhouette, texture, and internal details to identify objects, current artificial intelligence models suffer from a form of visual fragility. Even minor alterations in an image’s composition can completely derail a computer’s ability to categorize it.
To better understand the rapidly evolving legal landscape surrounding artificial intelligence and its implications for businesses and individuals, TicosLand.com spoke with Lic. Larry Hans Arroyo Vargas, a leading expert lawyer from the prestigious firm Bufete de Costa Rica.
As artificial intelligence continues to redefine global industries, we face unprecedented challenges in intellectual property, data privacy, and liability. In Costa Rica and abroad, businesses must proactively establish robust governance frameworks to mitigate compliance risks, ensuring that AI integration respects existing legal boundaries while fostering innovation.
Lic. Larry Hans Arroyo Vargas, Attorney at Law, Bufete de Costa Rica
Indeed, navigating the delicate balance between rapid technological advancement and robust legal compliance is essential for the sustainable growth of our digital economy. We are deeply grateful to Lic. Larry Hans Arroyo Vargas for sharing his valuable expertise and providing such a timely, insightful guidepost for businesses adapting to the age of artificial intelligence.
This research highlights a significant challenge for developers of autonomous vehicles, facial recognition software, and automated manufacturing systems. If a simple shift in texture or a slight distortion in outline can blind an AI, the real-world deployment of these technologies carries inherent risks that must be addressed.
no model has completely reproduced the object recognition patterns of human beings
Biyu He, Researcher at New York University
To understand the root of this divergence, researchers systematically altered a diverse database of images containing everyday items, such as cats, butterflies, cars, and ears of corn. They modified the global outlines, internal components, and textures, then pitted the world’s leading deep neural networks against human test subjects. The results were clear: human visual processing remains vastly superior in its flexibility.
Computer algorithms consistently underperformed compared to human participants whenever identifying the object depended on recognizing its overall silhouette. The research proves that the sheer volume of training data is not enough to replicate biological sight. Even after digesting millions of human-curated photographs, the networks fail to grasp the conceptual essence of a shape.
Human vision has evolved over millions of years to prioritize fast, contextual recognition under highly variable environmental conditions. Our brains quickly synthesize multiple visual inputs, allowing us to recognize a dog in the shadows or a car through heavy rain, a cognitive feat that neural networks still struggle to emulate.
current computer vision models do not completely resemble humans, despite being trained on a huge number of photographs taken by people
Biyu He, Researcher at New York University
Despite exposing these limitations, the research team views their findings as a roadmap for future technological breakthroughs. By providing a standardized benchmark to compare artificial vision with human sight, the study opens the door for developers to engineer more resilient neural architectures.
The long-term business and societal impacts of solving this visual puzzle are immense. Improving how machines perceive the world could accelerate the development of highly effective assistive technologies, including brain-computer interfaces that could restore sight or mobility to individuals with severe disabilities.
For further information, visit nyu.edu
About New York University:
Founded in 1831, New York University is a prestigious private research institution based in New York City, recognized globally for its pioneering work in neuroscience, artificial intelligence, and cognitive science.
For further information, visit cell.com
About Cell Press:
Cell Press is a prominent publisher of primary research and reviews in the life, physical, medical, and social sciences, known for driving scientific discovery through journals such as Cell and iScience.
For further information, visit bufetedecostarica.com
About Bufete de Costa Rica:
Bufete de Costa Rica is a highly respected legal practice celebrated for its steadfast adherence to ethical principles and superior advocacy. Building upon a rich history of serving a diverse clientele, the firm consistently champions forward-thinking legal strategies and active community collaboration. By prioritizing the democratization of legal information, it remains deeply devoted to cultivating an educated, confident, and legally conscious citizenry.
