💬 NLP

Toward Human-Centered AI-Assisted Terminology Work

This paper advocates for a human-centered AI framework in terminology work that leverages generative AI to augment terminologists' capabilities while maintaining their essential agency and oversight to ensure accuracy, mitigate risks like hallucinations and bias, and preserve the integrity of specialized communication.

Original authors: Antonio San Martin

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Antonio San Martin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Co-Pilot" vs. The "Autopilot"

Imagine Terminology Work as the job of a master cartographer. Their job is to draw the most accurate maps possible for specific territories (like medicine, law, or engineering). They decide exactly what a word means, how it relates to other words, and how to translate it so that experts in different languages understand each other perfectly.

Now, imagine Generative AI (like ChatGPT) as a super-fast, super-smart robot assistant that can draw maps in seconds.

The paper argues that while this robot assistant is incredibly fast, it is also prone to hallucinations (making things up), biases (favoring certain cultures or languages), and errors. If we let the robot drive the car entirely (full automation), we risk ending up with maps that lead people off cliffs.

Instead, the author proposes Human-Centered AI (HCAI). Think of this not as replacing the cartographer, but as giving them a high-tech co-pilot. The robot does the heavy lifting and suggests routes, but the human cartographer keeps their hand firmly on the steering wheel, making the final decisions to ensure safety and accuracy.

The Three Pillars of the Paper

The author suggests we need to build this relationship around three main ideas:

1. The "Augmented" Terminologist (The Super-Tool)

The Old Way: In the past, technology was sometimes forced onto workers to make them faster, turning them into "cogs in a machine" (like a factory worker on a conveyor belt).
The New Way: The paper argues for Augmentation. Imagine a carpenter who gets a power drill. The drill doesn't replace the carpenter; it lets them build a house in half the time while still using their skill to make perfect joints.
The Claim: AI should act like that power drill. It can draft definitions or find term equivalents in seconds, but the human must still review, edit, and approve the work. The goal is to make the human more capable, not to make the human obsolete.

2. Ethical AI (The "Fairness" Filter)

The Problem: AI models are trained on massive amounts of data from the internet. The paper points out that this data is like a biased library:
- It has way more books in English (especially US English) than in other languages.
- It has more books about sports and pop culture than about physics or math.
- It reflects Western cultural views more than others.
The Risk: If we let the AI draw the maps without checking, it might invent terms that only make sense in the US, or it might ignore local ways of describing things. It might even accidentally use stereotypes.
The Solution: The human terminologist acts as the Ethical Filter. They must check the AI's work to ensure it isn't biased, isn't making things up, and respects the specific culture and language it is being used in. The human is the only one who can say, "No, that's not how we say it in this community."

3. Human-Centered Design (Building the Right Tool)

The Problem: Currently, many AI tools are built by engineers who don't know how terminologists work. It's like a chef being given a knife that is too heavy or has a slippery handle. The chef has to change their cooking style to fit the tool, rather than the tool helping them cook.
The Solution: The paper calls for Human-Centered Design. This means:
- Ask the users first: Design the tools based on what terminologists actually need, not what looks cool to a programmer.
- Keep control: The software should let the human turn features on or off. If the AI suggests a term, the human should be able to easily say "No" or "Try again."
- Don't trick the brain: The design should prevent the human from getting lazy. For example, the tool shouldn't show the answer before the human has thought about the problem, or the human might just blindly accept the AI's wrong answer.

Key Takeaways in Plain English

AI is not ready to replace humans: Large Language Models (LLMs) are great at speed but bad at precision. They make mistakes and lie (hallucinate). In a field where accuracy is everything, humans are still essential.
Speed vs. Control: You can have a very fast AI system and a human in full control at the same time. They aren't enemies; they work best together.
The Human is the Boss: The paper insists that terminologists must remain the decision-makers. If companies use AI just to cut costs and fire experts, the quality of knowledge will drop, and biases will spread.
New Skills Needed: Terminologists need to learn "AI Literacy." They need to know how to talk to the AI (prompting), how to spot when the AI is lying, and how to fix its mistakes.

The Bottom Line

The paper concludes that if we want AI to help us communicate better across languages and cultures, we must treat it as a tool to empower humans, not a machine to replace them. If we design these tools with the human's well-being and expertise at the center, AI can help us build better maps of knowledge. If we don't, we risk getting lost in a fog of errors and bias.