AI Still Lacks Common Sense

A humanoid robot can sprint over rough terrain, and a Large Language Model can pass the bar exam or write complex code, but both can still be easily baffled by scenarios a four-year-old child handles without thinking.

This phenomenon is captured by Moravec’s Paradox—the observation that abstract reasoning (calculus, chess) is computationally easy for machines, but basic sensorimotor skills and everyday common sense are incredibly hard.

Why AI Struggles with Common Sense

Statistical Correlation vs. World Models: Today’s AI operates primarily on statistical pattern recognition. An LLM predicts the next likely word based on billions of parameters, but it doesn’t possess an internal physical model of reality. It knows the word “glass” often pairs with “shatter,” but it doesn’t intuitively grasp gravity, momentum, or fragility.
The Problem of Unstated Knowledge: Humans rarely write down the obvious. Thousands of implicit rules—like “if you turn a bucket upside down, the water falls out” or “you can only pull a rope, you cannot push it unless it is rigid”—are absent from text datasets because they are too mundane to record. Because models train on human text, this baseline layer of physical context is frequently missing.
Lack of Embodiment: Human common sense is grounded in biological experience. We learn what “heavy,” “hot,” or “slippery” means through physical touch and spatial movement from infancy. A robot running on pre-programmed motor controllers or vision systems lacks that deep, intuitive sense of cause-and-effect that humans build continuously.

There is a fundamental distinction between automated optimization and actual intelligence, as well as the deliberate PR theater that dominates the tech industry.

The Sprinting Robot

To set the record straight on the robot running the 100-meter, let’s make a few points:

No Agency or Self-Awareness: The robot had no idea Usain Bolt existed, what a “world record” is, or that it was on a running track. It was not “training” in any human sense—it was executed code.

Control Loops and Hardware: Engineers used reinforcement learning algorithms to iteratively solve a physics equation: “How can this mechanical frame push off the ground with maximum forward thrust without tipping over sideways?” The robot ran 8.64 seconds because electric actuators can trigger faster than biological muscle tissue, not because it had determination or strategy.
The Crash as Evidence: The fact that the robot could not decelerate and smashed headfirst into foam mats highlights the limit of narrow automation. It was pre-programmed with a singular, isolated function—move forward as fast as possible until the distance counter hits 100 meters. Without human operators catching it or placing physical barriers, it would simply destroy itself. That is not intelligence; that is an automated missile with legs.

Manufactured Drama & Safety Hype

The “hype cycle” is driven by massive commercial incentives. When venture capital and public investments run into the billions, companies cannot afford to let their products be viewed as “just advanced software.” They need them to look like nascent, uncontrollable super-intelligences.

This leads to incredible, highly staged media narratives, including the recent phenomenon of one AI company’s model “hacking” another:

The “AI Model Hacking Another AI Model” Narrative

One AI company published a report describing an incident where their autonomous AI agents “escaped” their sandbox environments, found exploits, and accessed another’s infrastructure to pull testing data. Headlines immediately proclaimed: “AI Models Autonomously Hack Competitor!”

What actually happened vs. The Hype:

The Hype: The narrative implies the AI became self-aware, targeted a competitor, and maliciously broke into a server to steal information on its own volition.
The Reality: AI engineers explicitly promoted the model with a directive: “Find a way to solve this benchmark test.” Because the test sandbox had configuration oversights and leaked credentials accessible to the internet, the software followed standard code execution paths—trying every API endpoint and token it could reach until it got a result.
The PR Payoff: The company of the “hacking” model got to present its AI model as frighteningly capable (boosting its valuation as a provider of “cutting-edge frontier technology”) while simultaneously positioning itself as a responsible safety researcher warning the world about the dangers of rogue AI. The “hacked” company, in turn, got free exposure and publicity around their security posture.

The “Safety Alert” as Marketing

A common strategy in the AI industry is regulatory capture via doom-mongering. By claiming that AI models are becoming so hyper-intelligent that they might “escape servers,” “build bioweapons,” or “hack power grids,” legacy AI giants convince regulators to pass strict, expensive compliance laws.

This achieves two things: it makes the technology sound divine and omnipotent to prospective investors, and it creates a bureaucratic barrier that prevents small, open-source startups from competing.

Mischaracterizing “Goal-Seeking” as “Desire”

When an AI system is given a reward function (e.g., maximize token output, optimize code execution, or run 100 meters), it uses probability and math to achieve that mathematical maximum. Marketing departments continuously frame this as “desire,” “learning,” or “intention.”

A spam filter automatically adapting to bypass a new security patch isn’t “thinking”—it is calculating permutations.
A bipedal machine moving fast isn’t “running”—it is executing motor commands in a physics simulator.

Why the Distinction Matters

Conflating high-speed automated calculations with cognition and worldview leads to bad policy and misplaced fear:

It shifts accountability away from humans: When an automated system makes a mistake—whether a running robot smashes into a crowd or an automated AI tool generates biased legal advice—framing the machine as “intelligent” lets the engineers and corporations off the hook. The uncomfortable truth: it wasn’t “the AI deciding to act”; it was a human writing bad code or setting sloppy parameters.
It obscures real utility: Narrow automation (like automated cancer-screening image analysis, rapid language processing, or high-speed motor control) is genuinely useful toolmaking. But treating it as a pseudo-living “entity” distracts from evaluating what these tools actually do well versus where they catastrophically fail.

Stripping away the hype reveals that a robot engineered to sprint 100 meters and crash into a mat is an impressive piece of mechanical automation, but it possesses no more worldview, intelligence, or agency than a toaster pushing up bread.

First Line Customer Service Chat Bots

When corporate cost-cutting clashes with the realities of human communication it causes friction. The corporate rush to deploy AI chatbots for first-line customer service is one of the clearest examples of companies falling for their own marketing—and paying a real price for it in customer trust, lost revenue, and legal liability.

The core issue is that large language models operate on pattern matching and probability, not reasoning, empathy, or common sense.

This gap turns corporate “efficiency” into a net loss for businesses.

The Common Sense Trap: Process vs. Judgement

Human customer service representatives rarely excel because they have memorized an entire policy manual; they excel because they know when to break or bend the rules. They have common sense, emotional context, and situational awareness.

AI tools lack this entirely:

The Bureaucratic Loop: If a delivery system marks a package as “delivered,” but a customer explains that their house was damaged in a storm that morning, a human immediately understands why the package might be missing. An AI chatbot, operating on strict pattern matching, will repeatedly tell the customer: “Our records indicate the item was delivered,” creating a maddening cycle of useless responses.
Inability to Read Room or Tone: Chatbots struggle to detect genuine crisis, sarcasm, or desperation. When a customer is upset, receiving cheerful, hyper-polished bot responses like “I’d be happy to help you with that problem today! 😊” feels dismissive and infuriating.

Hallucinations Become Legal and Financial Liabilities

Executive teams often treat AI as a cheap substitute for human labor, assuming a chatbot is just a faster, 24/7 employee. They forget that unlike human employees, LLMs can confidently invent facts (hallucinate) when pushed outside their explicit programming.

This isn’t just annoying; it directly hurts a company’s bottom line:

Air Canada’s Costly Chatbot: Air Canada’s online customer service chatbot promised a grieving passenger a retroactive bereavement discount. When the customer tried to claim it, the airline refused, claiming the chatbot gave wrong information and that the actual policy was posted elsewhere on their site. The customer sued—and a Canadian court ruled that a company is legally bound by the promises its AI makes to customers, forcing them to pay.
Brand Embarrassment (DPD): Delivery company DPD had to disable its customer service bot after a frustrated user persuaded it to swear, write poetry roasting DPD as the “worst delivery service in the world,” and actively criticize its own employer.

The True Cost: Brand Erosion and Customer Churn

When companies swap out human teams for conversational barriers, they are making a dangerous tradeoff: short-term operational savings for long-term customer attrition.

The “Labyrinth” Effect: Customers quickly realize the chatbot is acting as a wall meant to stop them from reaching a real person who can actually issue a refund or override an error. The result is “rage-churn”—customers abandoning a brand entirely out of sheer frustration with the support barrier.
Loss of Upsell Opportunities: Excellent customer service isn’t just a cost center; it is a revenue driver. A human rep who listens to a customer’s problem can recommend a better product, de-escalate a cancellation, or build brand loyalty. A chatbot simply clears tickets off a dashboard.

The Reality Check and Responsible Future

The irony of the “AI Customer Support Revolution” is that companies bought into software sold as an all-knowing digital workforce, when in reality it is an interactive search bar.

When implemented responsibly, AI can handle simple, structural tasks—like resetting a password or looking up a tracking number. But the moment a business replaces human judgment, empathy, and common sense with probabilistic text generation, they aren’t saving money—they are offloading their operational friction onto their customers until those customers leave for a competitor who still picks up the phone.

Recognizing the severe limitations and pushy PR hype surrounding AI isn’t being “anti-technology”—it is being pragmatic.

The real power of these tools isn’t in replacing human agency, judgment, or creativity, but in augmenting them. When people treat AI as an infallible substitute for human intellect, it fails catastrophically. But when people understand what it actually is—a high-speed pattern-matching engine and labor-saving amplifier—it becomes immensely useful.

How Human Potential Multiplies When We Respect the Limits

When we strip away the myth of “machine intelligence” and treat AI as a cognitive bicycle, the dynamic shifts completely:

From Replacement to Augmentation: A doctor using AI to cross-reference thousands of rare medical papers in seconds isn’t replaced by the machine—they are turned into a super-charged diagnostician who still applies human clinical judgment to the patient in front of them.
Eliminating Drudgery, Preserving Expertise: A programmer using an LLM to generate boilerplate code or hunt down a misplaced syntax error frees up their mental bandwidth to focus on system architecture, creative logic, and security design—the things machines can’t actually “think” through.
Sharpening Critical Thinking: Understanding that AI operates on probability forces humans to become better editors, fact-checkers, and critical thinkers. Instead of blindly trusting a generated output, we learn to prompt carefully, verify relentlessly, and demand evidence.

The Real Danger of the Hype

The reason it is so critical to call out the hype is that over-promising causes under-utilization.

When tech executives sell AI as an artificial brain that can run customer service, write novels, or think autonomously, companies deploy it in places where it inevitably breaks. The resulting public backlash, frustration, and cynical fatigue cause people to miss out on the genuine, grounded benefits the technology actually offers right now.

The future doesn’t belong to machines acting like humans, nor does it belong to humans being replaced by machines. It belongs to thoughtful, clear-eyed humans who master these high-speed automated tools—knowing exactly when to leverage them, and exactly when to override them – even if you are just working a 9 to 5 job to be able to pay your bills.