AI Lies? Why We Continue to Use LLMs Anyway

Imagine this: You start your workday this morning and hand a multi-page draft contract you compiled yesterday to an AI. You say, “Pull out the major risk items.” The AI provides the answer in an instant. But suddenly, a thought crosses your mind: ‘Can I really trust this? Did the AI just make things up on its own?’

Recently, public opinion surrounding generative AI—especially Large Language Models (LLMs)—has been sharply divided. Some cheer, claiming AI will completely transform our lives, while others strongly criticize it, pointing to hallucinations (the phenomenon where AI presents false information as if it were true), environmental impacts, and reliability issues. Some open-source software projects, such as Zig and Gentoo, have even begun rejecting AI-generated code contributions. [Source: The LLM Critics Are Right. I Use LLMs Anyway.

Jeremy Theocharis](https://www.theocharis.dev/blog/llm-critics-are-right-i-use-llms-anyway/)

However, despite mounting criticism, countless users around the world are still paying monthly fees to use LLMs. [Source: 2025: The Year in LLMs

Hacker News](https://news.ycombinator.com/item?id=46449643) Why exactly do criticism and usage coexist?

Why Does This Matter?

The era of using AI simply because it is “amazing” is over. AI is now establishing itself as an essential tool that determines work efficiency. For developers, LLMs serve as excellent partners when designing complex function structures. Source: Here’s how I use LLMs to help me write code However, if the trust issue remains unresolved, we cannot safely use AI in business settings where critical decisions must be made.

The voices of critics act as “guardrails” that make AI safer. The hallucinations they point to are, in fact, byproducts of the fact that the internet data LLMs are trained on is full of contradictions and biases, and that models are trained to provide “confident answers the user will like” rather than accuracy. Source: It’s 2026. Why Are LLMs Still Hallucinating? - Duke University Libraries As we face these problems head-on, the technology becomes more refined.

Easy Understanding: The Superpowered Librarian

You can easily think of an LLM as a “superpowered librarian” who has read every book and internet post in the world. This librarian has learned countless patterns and quickly finds the most plausible answers when we ask questions. Source: Part 1: How to Apply LLMs and AI to Contracts

Sometimes, however, it mixes in bizarre information from that vast knowledge. It’s similar to a librarian who has read so many books that they mistake the contents of a fictional novel for actual historical facts. This is why a method of having a “critic” watch over the librarian’s answers has recently been introduced.

“Critic Models” are LLMs that read code written by another AI and provide feedback—just like a fellow developer—on whether there are logical errors or security risks. [Source: LLM Critics Help Catch LLM Bugs

Athina AI](https://blog.athina.ai/llm-critics-help-catch-llm-bugs) In fact, the critic model “CriticGPT” has sometimes performed better than humans at finding bugs in code. Source: LLM Critics: A New Weapon in the Fight Against AI Bugs

Where Do We Stand Now?

Today’s LLMs operate based on a “decoder-style transformer” structure that understands context to predict the next word, but they have become much smarter internally. They are increasing efficiency by using a “Mixture-of-Experts (MoE) structure”—reminiscent of groups of experts—activating the most suitable small models depending on the nature of the question. Source: The State Of LLMs 2025: Progress, Progress, and Predictions

However, clear limitations still exist. According to MIT research, “synthetic errors” occur where a model forms incorrect correlations during training, making it look perfect on the surface but failing logically. Source: Researchers discover a shortcoming that makes LLMs less reliable This proves that we must not blindly trust AI’s answers and that a “verifying user” is still absolutely necessary.

The Future of AI: The Collaborative Assistant

Next-generation LLMs will become much cheaper and more efficient than they are now. Source: LLMs+: 10 Things That Matter in AI Right Now Instead of just pouring in more data, their ability to find errors we worry about and correct them autonomously will be strengthened.

We need to move past the stage of thinking of AI as a “perfect answer sheet that knows everything” and shift our perspective to viewing it as a “collaborative, creative assistant.” Criticism is not a lamp that stops technology, but one that lights the way for it to move in the right direction. Only when we recognize the limitations of LLMs and critically utilize the tool atop that foundation can we truly become the masters of the AI era.

AI Opinion

Criticism is the fertilizer for technological growth. The emergence of ‘critic models’ that acknowledge and compensate for LLM limitations shows that we can coexist more healthily with AI.

References

1. How I use LLMs - YouTube
2. As an Experienced LLM User, I Actually Don’t Use Generative LLMs…

3. [LLM Critics Help Catch LLM Bugs

   Athina AI](https://blog.athina.ai/llm-critics-help-catch-llm-bugs)

4. Using ChatGPT is not bad for the environment
5. Bad (but common) LLM criticisms - Ritza Articles
6. Part 1: How to Apply LLMs and AI to Contracts: What are LLMs anyway? - Knowable
7. Here’s how I use LLMs to help me write code

8. [LLMs+: 10 Things That Matter in AI Right Now

   MIT Technology Review](https://www.technologyreview.com/2026/04/21/1135645/llm-large-language-models-ai/)

9. It’s 2026. Why Are LLMs Still Hallucinating? - Duke University Libraries
10. Assessing the Strengths and Weaknesses of Large Language Models
11. LLM Critics: A New Weapon in the Fight Against AI Bugs
12. LLM News, Updates and Articles
13. The State Of LLMs 2025: Progress, Progress, and Predictions

14. [The LLM Critics Are Right. I Use LLMs Anyway.

    Jeremy Theocharis](https://www.theocharis.dev/blog/llm-critics-are-right-i-use-llms-anyway/)

15. [Researchers discover a shortcoming that makes LLMs less reliable

    MIT News](https://news.mit.edu/2025/shortcoming-makes-llms-less-reliable-1126)

16. 2025 LLM Year in Review – karpathy

17. [2025: The Year in LLMs

    Hacker News](https://news.ycombinator.com/item?id=46449643)