With the recent advancements in LLMs, it’s increasingly common to hear discussions about automating annotation for machine learning (ML) models. It understandably raises a lot of questions among model developers—especially when coupled with the perception that human annotation is slow, expensive, and error-prone, whereas automation is cheaper and faster.

So which is the right choice for your model, based on where the technology stands today?

The TL;DR

For the majority of model developers, a combination of the two — human and automation — is where you’ll see the best balance between quality and accuracy versus lower costs and efficiency.

In this post, we’ll explore why humans still need to be in the loop today. In Part 2, we’ll discuss where automation works best today in annotation workflows.

What Are Today’s Model Limitations?

Most ML models today are “closed systems” that learn patterns, structures, and relationships from a fixed dataset during training and fine-tuning. These datasets are often enormous, pulled from publicly available information on the internet. Due to the inherent variability and inaccuracies in the data, it unsurprisingly results in a certain level of noise, which impacts model performance.

Improving these models requires new knowledge — embedding context, making corrections, etc. — that can only be added by human labelers and annotators today. 

Why is this the case? If humans are replaced by models in the labeling or annotation workflow, the information within the model may be reorganized more efficiently but the amount of knowledge or capabilities within that closed system would be mostly unchanged.

Example: If a teacher asked students to correct each other’s quizzes, without providing the correct answers, the students’ collective knowledge might become better distributed since some understood parts of the material better than others. However, no new knowledge is created here and concepts that were missed are still unlikely to be fully grasped. It’s not until the teacher steps in to correct the remaining errors and teach new concepts that real learning occurs. Think of the teacher as the human annotators and the students as the model.

What about Models as Agents?

It’s true that not all models function as closed systems — there is ongoing research focused on developing AI agents that can autonomously pull in external information to enrich their knowledge base or add onto existing capabilities using external tools (such as accessing a calculator). 

Currently, these types of ML models are at a developmental stage where they’re still being explored. They also don’t eliminate the need for human input, rather, they shift where human knowledge is required. Models will still need carefully curated examples to learn how to use these tools or data effectively and they’ll still need downstream human validation and editing to ensure outputs are high quality, accurate, and generating the right outcomes.

Other Human vs. Data Challenges

When improvements are made to a model, it directly correlates with how effectively human knowledge is transferred into the training data. As a result, it’s crucial that the information humans inject into the model is both accurate and relevant and addresses the weak spots where the model will benefit the most.

However, this comes with a couple of challenges. 

1. Unclear pathways for model improvements: Without a clear understanding of where and how to enhance the model, human efforts may be inefficient or misdirected. Labelers and annotators often lack detailed insight into specific aspects of the model’s performance, hampering their ability to target improvements effectively. Conversely, model designers (often on the client side) may struggle to articulate the model’s strengths and weaknesses with sufficient precision to provide clear direction to labelers and annotators (frequently external vendors). Even in-house labeling and annotating efforts can suffer from miscommunication between teams.
2. Cognitive friction: Often, packaging and delivering new information is more of a challenge than producing the information itself. In image captioning, for example, it’s easy for a human to mentally process the details of the scene but it’s significantly more taxing for them to write it down into a well-written, concise paragraph. Similarly, it’s easy for most people to recognize a partially hidden object in an image but for a machine to do the same thing, a human would have had to carefully segment the contour of the occluded object using editing tools — a cognitively taxing task.

‍

Given all of these challenges and limitations, how should models and humans work together to produce the best results? Read Part 2 for more.

‍

Download the full white paper: Machines Still Need Us

‍

Yutong Liu & Kingston School of Art  / Better Images of AI / Talking to AI 2.0 / Licenced by CC-BY 4.0