When some say “robots are incompetent,” DeepMind’s Chief Scientist counters: “Without a body, AI’s intelligence is merely an illusion.”

What is true intelligence? 

Is it the ability to speak, run, mimic human behavior, or truly understand the world and one’s own existence? In a recent interview, Murray Shanahan, Chief Scientist at Google DeepMind and Professor of Cognitive Robotics at Imperial College London, delivered a sobering perspective: *”AI without a body may exhibit certain cognitive functions, but it remains fundamentally distinct from embodied intelligence.”* 

This assertion gained real-world resonance last week in Beijing’s Yizhuang district, where 20 humanoid robots attempted the world’s first robotic half-marathon. They stumbled, got lost, and struggled—yet some crossed the finish line. While executing movements, they lacked comprehension of *”forward motion”*; running 21 kilometers left their cognition unchanged. Meanwhile, investor Zhu Xiaohu announced withdrawals from embodied AI projects, signaling a divergence between scientific conviction and capital pragmatism. 

Amidst the feverish enthusiasm for humanoid robots and embodied intelligence, leading scientists maintain: without a physical form, AI intelligence remains an illusion; while some investors have chosen to withdraw. Persistence and skepticism, idealism and pragmatism, begin to diverge on the long road toward embodied cognition. Here, guided by DeepMind’s redefined understanding of intelligence’s essence, we pierce through the veil of illusions to confront the fundamental question: what constitutes an AI that truly comprehends the world.

Section 1 | The Age of Illusion: Why Does AI Seem Intelligent?

When interacting with today’s large models, they respond fluently, analyze problems, and even craft novel expressions. When observing humanoid robots sprinting, they maintain balance, navigate obstacles, and execute complex maneuvers. Such phenomena create a collective illusion: AI appears profoundly intelligent.

Yet to Murray Shanahan, DeepMind’s chief scientist, these surface manifestations constitute an ontological illusion.

Why an Illusion?

Current large models—whether generating language or executing physical actions—fundamentally lack reasoning and comprehension. Their operations reduce to:

• Predicting the next plausible token

• Anticipating the next viable movement

• Retrieving responses statistically aligned with “correctness” in training data

Like actors memorizing thousands of scripts, they effortlessly switch emotional tones—mimicking love, fear, or triumph—without ever having loved, feared, or fought. They merely perform.

Linguistic Fluency ≠ Reasoning Capacity

Shanahan observes that human susceptibility to AI’s mimicry stems from language’s role as humanity’s supreme cognitive artifact. Speakers generally understand their utterances, but AI diverges: it is a statistical language-prediction engine, selecting contextually optimal responses to satisfy human interlocutors.

• It states “water is wet”

• Deduces “spring follows winter”

• Cites philosophical maxims

Yet when asked “Why did you answer this way?” it cannot provide self-aware justification.

Motor Coordination ≠ World Understanding

This logic extends to Beijing Yizhuang’s humanoid robot marathon:

• Running emerges from gait-pattern recognition and dynamic equilibrium optimization

• Obstacle avoidance derives from sensor-driven distance adaptation

But does the robot comprehend why it runs? Does it grasp competition rules, purposes, or goals? No. It mechanically executes programmed instructions, guided by parametric optimization. Its strides project algorithmic solutions, not volitional intent.

The Core: Illusion of Reasoning

Shanahan defines contemporary large models and humanoid intelligence as the Illusion of Reasoning:

They simulate reasoning through data-driven performance; they counterfeit understanding via statistical mimicry.

This illusion deceives not only the public but even tech insiders, as human brains instinctively equate linguistic fluency and motor precision with genuine cognition.

The Fundamental Disconnect

The DeepMind scientist’s thesis crystallizes a critical distinction:

• Language as performance, not comprehension

• Action as repetition, not awakening

• Intelligence as existential understanding, not mere content generation

Failure to pierce this epistemological veil, Shanahan warns, obscures humanity’s quest for the true origin of intelligent awakening.

Section 2 | The Mirage of Reasoning: Is AI Truly Thinking?

If today’s large models are performing reasoning through imitation, decades ago, the AI community pursued an alternative approach: hard-coded reasoning—the paradigm of Symbolic AI.

What is Symbolic AI?

From the 1960s to the 1990s, a seemingly rigorous methodology dominated AI:

• Encoding human knowledge and world rules into formal logic

• Embedding reality into computers via nested if-then rule systems

• Enabling machines to derive conclusions through deductive reasoning

Examples:

• If a person coughs, has a fever, and experiences dyspnea → likely influenza

• If water temperature exceeds 100°C → boiling occurs

• If morning traffic congestion → depart earlier

Proponents believed that exhaustively codifying the world would grant AI human-like understanding and reasoning.

Why Did Symbolic AI Collapse?

Initially successful in expert systems (medical diagnosis, logic engines), Symbolic AI encountered three insurmountable failures as systems scaled:

❶ Knowledge Acquisition Bottleneck

Human knowledge is inherently fuzzy, tacit, and context-dependent. Formalizing this into explicit rules proved:

• Prohibitively costly

• Conceptually impossible for ambiguous notions like “approximate,” “impending rain,” or “cautious operation”

❷ Rule Fragility

Symbolic systems relied on rule completeness and correctness. Real-world dynamics—unpredictable edge cases—caused systemic failures:

• Medical systems paralyzed by unencoded viral mutations

• Traffic planners incapacitated by extreme weather

The world is open; symbolic systems are closed.

❸ Commonsense Crisis

Even exhaustive rules couldn’t replicate humanity’s implicit commonsense. Consider:

• A book placed on water remains afloat; a stone sinks

Humans intuit this instantly. Symbolic AI required explicit encoding of:

• Weight, buoyancy, density

• Definitions of “submersion” and “floating”

This led to the Common Sense Knowledge Bottleneck—the cost of codifying trivial knowledge dwarfed system development itself.

The Fall of Symbolic AI

By the early 2000s, amid data explosions and machine learning’s rise, researchers gradually realized that hard-coding the entire world was a Sisyphean task. Symbolic AI faded from large-scale applications, plunging the field into a “AI winter.”

if you need the full report, please leave your email address we will send you by email