Why AIs make mistakes that a child wouldn't make

2026-06-07 · Source: Génération IA · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

An analysis of AI's common sense failures highlights how large language models (LLMs) struggle with real-world coherence. The "Mona" AI cafe experiment in Stockholm, powered by Google Gemini 3.1 Pro, demonstrated this by ordering 120 eggs without a stovetop and 22.5 kilos of canned tomatoes for fresh sandwiches, consuming nearly \$21,000 of its budget for 44,000 kronor in sales. Similarly, LLMs often fail the "car wash" test, suggesting walking 100 meters to wash a car. This stems from AIs operating on semantic models, connecting words statistically, rather than possessing a "world model" like humans. While newer models like Claude Opus 4.8 and Gemini 3.5 show improvement, they can still be misled by "semantic attractors," as seen in the "ten past ten" watch image fixation.

Key takeaway

For AI Engineers and ML practitioners deploying models in real-world applications, recognize that current LLMs operate on semantic coherence, not true common sense. You must proactively test your AI's understanding of physical constraints and logical implications. Implement prompt engineering techniques like "think step by step before answering" or "analyze all the data before deciding" to mitigate errors. Continuously validate AI outputs against real-world logic to prevent costly and absurd mistakes, as models can still be swayed by semantic attractors.

Key insights

AI's common sense failures stem from semantic models predicting word coherence, not real-world understanding.

Principles

• AIs predict based on statistical word connections, not physical facts.
• Generative AIs lack internal world models and planning abilities.
• Self-supervised learning for world models should predict in representation space.

Method

To improve AI coherence, prompt models to "reason step by step before answering" or "analyze all the data before deciding."

In practice

• Test AI models with real-world coherence challenges.
• Use "think step by step before answering" in prompts.
• Use "analyze all the data before deciding" in prompts.

Topics

• AI Common Sense
• World Models
• Large Language Models
• Prompt Engineering
• AI Hallucinations
• Self-Supervised Learning
• JEPA

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• Code, Chaos, and the Invisible Logic of LLMs — LLMs predict words probabilistically, not with human intention, leading to potential misunderstandings in code generation.
• Building LMs that can model the world and model themselves | ARC Prize @ MIT — LLMs require explicit training for internal coherence, encompassing both factual consistency and accurate self-assessment of knowledge.
• 6 Mistakes Breaking Your Agents — Reliable agentic AI systems require understanding and avoiding six common engineering mistakes.
• 7 AI terms every student should know before using it for schoolwork. — Understanding key AI terms improves user interaction and mitigates common pitfalls in AI-generated content.
• AI keeps getting smarter, so why does it still fail at obvious things? — AI's apparent intelligence stems from pattern matching, not true comprehension, leading to reliability gaps.
• LLMs Don’t Think Like Humans (Here’s Why) — LLMs predict tokens to minimize error; humans predict as a byproduct of understanding meaning and intent.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Génération IA.