When AI Learns to Act in the World - Robotics, Ethics, and Society's Response

The Moment AI Reached Out to Touch the World

Last week, I was demonstrating a new robotics system to a group of students when one asked a simple but profound question: “What happens when the AI doesn’t just understand our commands, but can actually do things in the physical world?”

It was a question that revealed how we often think about AI advancement—focused on language models that chat or image generators that create pictures—while overlooking the perhaps more consequential frontier: AI that can perceive, reason, and act in our physical environment.

That same week, on March 13, 2025, DeepMind provided a compelling answer to that student’s question with the announcement of Gemini Robotics and Gemini Robotics-ER—AI models designed not just to understand the world, but to interact with it directly.

Gemini Robotics: AI That Can Perceive, Reason, and Act

Gemini Robotics represents what DeepMind calls a “vision-language-action model”—a system that integrates visual perception, language understanding, and physical action capabilities into a single unified architecture. Unlike earlier approaches that treated perception, cognition, and action as separate modules to be stitched together, Gemini Robotics was trained end-to-end to handle the complete loop: see the world, understand instructions, reason about what to do, and execute actions.

What makes this particularly powerful is its ability to generalize. In demonstrations, the model showed it could:

• Follow natural language instructions to manipulate objects it had never seen before
• Use tools in creative ways (like using a screwdriver as a lever when the proper tool wasn’t available)
• Adapt its grip and pressure based on tactile feedback when handling delicate items
• Recover from mistakes by adjusting its approach in real-time

The companion model, Gemini Robotics-ER (Embodied Reasoning), focuses specifically on the spatial understanding needed for complex physical tasks. It enables robots to reason about 3D relationships, imagine how objects might fit together, and plan sequences of actions—capabilities demonstrated through tasks like folding origami or packing irregular items into containers.

But as exciting as these technical advances are, they arrived amid a week that reminded us that progress in AI’s physical capabilities doesn’t exist in a vacuum—it unfolds within complex social, ethical, and legal frameworks that demand our attention.

The Week AI Forced Us to Ask Hard Questions

The same week that DeepMind unveiled robots that could fold origami, three other stories highlighted the societal challenges that accompany such advances:

First, Yale University made headlines when they suspended a scholar over concerns that an AI-generated article she published might have ties to terrorist organizations. The incident sparked intense debate about academic freedom versus security concerns in an era where AI can generate convincing content on any topic—including those that might be sensitive or dangerous. Critics argued the punishment was disproportionate, while supporters emphasized the need for vigilance against harmful content, however it’s produced.

Second, The Guardian reported on government proposals to use AI to replace certain civil service functions, promising increased efficiency and reduced backlogs. While the potential benefits are real—faster processing times, reduced human error, and 24/7 availability—the proposals also raised important questions about job displacement, transparency in algorithmic decision-making, and maintaining the human elements of public service that often require empathy, judgment, and moral reasoning that AI currently lacks.

Third, French publishers and authors filed a landmark lawsuit against Meta, alleging that the company used copyrighted works without permission to train its AI models. This legal challenge represents part of a growing global conversation about intellectual property rights in the age of generative AI, where models trained on vast datasets can produce content that closely resembles or derivative of existing works.

What This Means for Our Shared Future

These four stories from a single week in March 2025 paint a vivid picture of where we stand with artificial intelligence today:

On one hand, we have systems like Gemini Robotics that demonstrate astonishing technical prowess—AI that can perceive nuance, reason about complex physical interactions, and take meaningful action in the world. The potential applications are transformative: robots that could assist in eldercare, perform delicate surgical procedures, explore hazardous environments, or help with physical rehabilitation.

On the other hand, we’re grappling with fundamental questions about how such capabilities should be developed and deployed:

• How do we ensure AI systems that can act in the world are safe and reliable?
• What ethical frameworks should guide the development of embodied AI?
• How do we balance innovation with respect for intellectual property rights?
• What role should government play in regulating AI that impacts public services and safety?
• How do we prepare our workforce and society for changes that AI will inevitably bring?

The marriage of advanced perception (like that used to discover ancient desert civilizations) with sophisticated language understanding (as seen in GPT-4o’s native image generation) and now physical action capabilities (exemplified by Gemini Robotics) suggests we’re moving toward AI systems with unprecedented integrated capabilities.

But as these stories remind us, technological capability alone doesn’t determine outcomes. How we choose to develop, govern, and integrate these technologies will shape whether they amplify our best qualities as a species or exacerbate our worst tendencies.

If you work in robotics, AI ethics, policy, or simply care about how technology shapes our future, I encourage you to engage with these questions now—before the technology advances beyond our ability to shape its direction responsibly.

The robots are learning to act. The question is: what kind of partners will we choose to be?