If AI talks like a human, is it smarter? The trap of 'Humanization'

Imagine this: You ask the smartest intern at your company to write a report, and that intern suddenly demands, “I actually have ADHD, so could you talk to me using only Simplified Technical English (ASD-STE100, the restricted vocabulary English used in the aviation industry) to make it easier to understand?” While you might understand the intern’s personal situation, the core of the work depends on how accurately and clearly they can write that report.

Recently, many users utilizing Artificial Intelligence (AI) are trying hard to make AI seem like a ‘person.’ It is becoming a trend to add conditions to prompts (instructions input to the AI) such as “I have ADHD” or “Use a very human tone.” However, experts warn that these attempts could be ‘incorrect abstractions’ that undermine the AI’s intrinsic capabilities. Source 1, Source 3

Why does this matter?

As AI technology rapidly advances, we have gradually come to perceive AI not as a tool, but as a conversation partner. However, if AI sacrifices its inherent data processing efficiency to ‘sound human,’ it could lead to situations where it provides incorrect information or fails to solve complex problems at critical moments. The question of whether we will consume AI as a simple emotional companion or utilize it as a powerful thinking tool demands a fundamental change in how we approach technology. Source 4, Source 5

Easy to understand: Don’t take away AI’s ‘scratchpad’

Let’s compare the process of AI writing to ‘cooking.’ A Large Language Model (LLM) based on the Transformer architecture (an AI structure that grasps the relationships between words in a sentence) is like a chef searching for the optimal recipe using numerous ingredients (data).

When a user places an order to an AI saying, “You are a human,” it is like forcing a brilliant chef to ‘act like a human’ instead of cooking. Because they are focusing on the acting, they have fewer opportunities to display their natural talents, such as seasoning the dish correctly or checking the freshness of the ingredients.

Also, requests to keep answers overly brief require caution. In LLMs, ‘tokens’ (the language units AI breaks down for ‘thought’) are a kind of unit of thinking. Just as you need to write out enough calculation steps on a scratchpad to get the correct answer to a math problem, forcing too much conciseness on an AI is like taking away its ‘scratchpad space,’ which can cause the model to make dumber judgments. Source 12

Current Situation

In the AI industry, ‘Evaluation’—measuring how accurate an AI’s answer is and whether it aligns with the user’s intent—is emerging as a key task. Because AI answers are probabilistic, the same question can yield different results each time, which is why consistent performance evaluation is more important than anything else. Source 6, Source 9

Many people are assigning specific personas (fictional characters) to AI to gain a human tone, but experts worry that this ‘humanization’ can actually create confusion in evaluating AI’s efficiency and accuracy. We must face the fact that it is not that AI wants to look human, but that we are trying to put a human shell over it even at the cost of its efficiency. Source 4

What will happen in the future?

Moving forward, systems that precisely verify whether AI output is based on facts and free of logical errors—rather than giving AI an emotional persona—will become even more important. For example, in fields where accuracy is a matter of life and death, such as medicine or law, AI will be designed to undergo a process of checking logical steps one by one, rather than mimicking a human tone. Source 13

We must now wake up from the fantasy of trying to see AI as a human proxy. We must remember that while AI is sometimes not smarter than a house cat, it is a brilliant ‘thinking tool’ that exerts incredible efficiency when collaborating with humans. Source 8

MindTickleBytes’ AI Reporter Perspective

Technological progress often forces us into conflict between the ‘human warmth’ we expect from AI and the ‘mechanical precision’ AI can display. But remember: just as you wouldn’t ask your salary calculator or navigation system for a human story, AI, too, can provide the greatest help to our lives when it does not lose its essential performance and precision. It is time to focus on the kernel of ‘accuracy’ in the answers AI provides, rather than being seduced by the shell.

References

1. HumanisingLLMOutputsisDumb — Kuber Mehta

2. [HumanisingLLMOutputsIsDumb

   Hacker News](https://news.ycombinator.com/item?id=49243474)

3. [HumanisingLLMOutputsisDumb

   Devtalk](https://devtalk.com/t/humanising-llm-outputs-is-dumb/248727)

4. HumanisingLLMOutputsIsDumb - Cyber Media Creations
5. HumanisingLLMOutputsIsDumb - Avaoroi
6. EvaluatingOutputs
7. Who Validates the Validators? AligningLLM-Assisted Evaluation of…

8. [LLMs Are Dumber Than a House Cat

   Towards Data Science](https://towardsdatascience.com/llms-are-dumber-than-a-house-cat-81e7b3d63190/)

9. [Best Practices and Methods for LLM Evaluation

   Databricks Blog](https://www.databricks.com/blog/best-practices-and-methods-llm-evaluation)

10. My LLM’s outputs got 200% better with this simple trick.

11. [Oh boy. Someone didn’t get the memo that for LLMs, *tokens are units of thinking…

    Hacker News](https://news.ycombinator.com/item?id=47647907)

12. EvaluatingLLMOutputs: How to Know When AI is “Right” and How to…