As I settled into my internship at a major European bank, I was excited to explore the potential of AI assistants like Microsoft Copilot to streamline my work as an analyst. But what I discovered was both fascinating and frustrating.
On the one hand, Copilot produces well-formatted, professional-looking output that’s a breeze to read. However, upon closer inspection, I found that the factual accuracy is hit-or-miss, with summaries often containing incorrect figures or interpretations. What’s more, the analysis lacks depth and insight, rarely providing meaningful takeaways or conclusions.
This experience made me realize that current AI systems, including Copilot, ChatGPT, and Gemini, excel at writing but struggle to think, interpret, or connect ideas the way a human analyst would. It’s as if they’re exceptional at generating text but lack the ability to structure or connect information meaningfully.
I’ve faced similar challenges while building a learning app, QuizRevolt, which leverages AI to help users learn faster. The issue persists: AI often fails to form abstract, interconnected thoughts like humans. This raises a bigger question: Is this a limitation of current large language models (LLMs)? Are they inherently bad at forming abstract thoughts because of how they’re trained, or can fine-tuning, RAG pipelines, or multimodal inputs eventually bridge this gap?
I’d love to hear from others who have had similar experiences, especially in regulated industries or education tech. How are you working around these limitations? And what do you think is the fundamental gap that’s holding AI systems back from truly thinking like humans?
As I reflect on my experience, I’m reminded of the old adage that machines will never be the same as humans. While AI has come a long way, it’s clear that there’s still much to be learned about how to create systems that truly think and understand like us.