One of the latest big technological advances is AI. It has become part of our lives almost without us realizing it. ChatGPT, which was the first acknowledged contact with AI for many people, was launched on November 30, 2022, and since then the scene has drastically changed. Only 2 years have passed, but the advance is huge. Now hold on a second and imagine how it will be in 10 years… 

We’re at a new frontier. You can be prepared and ready for this new world, but you have to start learning now. So let’s begin with one of its building blocks: data labeling

That’s the first thing, right? Data labeling and annotation sounds boring and uninteresting, so why would you want to pay attention to it?

Well… Every app, social media platform, and website collects data, creating an ever-growing digital footprint. But without labeled data, AI is like a disassembled pair of scissors: an unfinished tool. 

Data labeling and annotation are the processes that make this data meaningful, teaching AI to see, hear, and understand the world around us. In other words, it’s the cornerstone that’s letting AIs transform our world.

Like Michael Cho, co-founder of FrodoBots, said: “In order to have genuinely useful robots in our life, whether at home to help us do our laundry, do deliveries, maybe be our personal butler to do stuff, we are a long way from it because of the lack of data and also the lack of real-world testing.”

So let’s define data labeling and annotation. It’s almost self-explanatory, but let’s do it anyway.

Data labeling is the process of categorizing data with “labels”. If you have a photo of a dog, it’s tagging them as “dog”. Have you ever completed a captcha that asked you to identify which pictures contained a bike or a train or some random object? Well, that’s basically it.

Data annotation goes a step further and lets you highlight other details. You're not just saying “this is a dog” but adding all the juicy details about it. It adds layers of information that make AIs more nuanced.

In the AI era, data is crude oil, but high-quality labeled data is like premium fuel. 

Accurate labeling directly impacts the performance of these systems, making them reliable and trustworthy.

The more accurately data is labeled, the better AI performs. And that means safer self-driving cars, more accurate medical diagnoses, and smarter personal assistants.

I know, I know… You want to get on to the fun stuff. But you need to understand a few concepts that will make things easier later on. I promise that I’ll talk about cool robots later, but hang on with me for a minute…

Different types of data present unique challenges. This means that each type require different labeling methods and tools. 

1. Image Labeling: It’s the dog photo example. It’s tagging what you see, like identifying animals in wildlife footage, pinpointing tumors in medical scans, or segmenting areas in autonomous vehicle footage.

2. Text Labeling: Text is data too. Labeling text implies things like analyzing emotions or language patterns to power chatbots and translation tools.

3. Audio Labeling: In audio recordings, you would have to distinguish speech from noise, identify specific speakers, or even transcribe it. In the end, it would help voice assistants recognize what you’re saying.

4. Video Labeling: Video adds a layer of complexity and includes tracking objects, analyzing behaviors, and detecting motion for security and autonomous driving applications.

Spoiler alert: this part is more interesting than it seems.

There are a few data annotation techniques depending on who's doing it. 

First, there’s manual labeling. This means that human annotators label data by hand. Assuming that they’re properly trained, this ensures high accuracy, but at a high cost and time investment.

On the other side, we have automated labeling. Algorithms automatically label data. This is faster, but also less accurate and prone to errors.

And in between them, we have semi-automated labeling. It’s a combination of human and machine effort, using algorithms to pre-label data and humans to refine it. Sometimes it’s called HITL for Human-In-The-Loop.

And this is really what we want. This combined effort gives us higher quality labeling in less time. But, of course, the skills needed to partner with an AI are slightly different from those needed for manual labeling. Which means current human labelers will need to upskill themselves.

So let me list some essential skills for modern data labelers:

1. Tool Proficiency

You’re going to use new tools, so the quicker you learn them, the more advantage you will have. Specifically:

• Learn popular labeling platforms (e.g., Labelbox, Scale AI, Appen)

• Understand keyboard shortcuts and efficient workflows

• Master quality assurance tools and validation techniques

2. Domain-Specific Knowledge

The old “pick a niche” advice. If you have a particular hobby or skill that makes you unique, you can double down on it and become an expert. For example:

• Medical labelers: Basic anatomy and medical terminology.

• Automotive labelers: Traffic rules and vehicle components.

• Text labelers: Grammar, context, and sentiment analysis.

• Computer vision labelers: Understanding perspective, lighting, and occlusion.

3. AI Collaboration Skills

And lastly, but not less important, understand your partner strengths and limitations:

• Understand confidence scores in pre-labeled data

• Identify common AI mistakes and biases

• Know when to trust or override AI suggestions

With all that said, I want to show you how working with an AI as a subject-matter expert would look like. It kinda looks like a loop, where each part improves the other one’s work.

1. AI first attempts to label the data.

2. Workers review AI-generated labels, focusing on edge cases and low-confidence predictions.

3. Humans should document patterns in AI errors for system improvement.

1. Use AI to flag inconsistencies in human labeling

2. Apply AI-suggested improvements while maintaining human judgment

3. If needed, cross-reference similar cases from the AI's knowledge base

Now, this is not a phase in itself, but something to keep in mind. The goal of working together with an AI is to improve speed while keeping a high quality. So there are a few things you can do to make maximize efficiency, like:

• Let AI handle routine, high-confidence cases.

• Focus human attention on complex scenarios.

• Use AI to batch similar items together.

Here’s where the excitement begins. As I said, data annotation might look like a boring topic at first glance. But it enables innovation across various industries, from healthcare to retail. Let's explore a few areas where it's making a big impact:

That means “text”, but you gotta learn the slang.

By labeling text data, companies improve chatbots, translation tools, and sentiment analysis systems, enabling machines to understand context, slang, and even sarcasm.

Elon Musk

@elonmusk

xAI’s Grok system is designed to have a little humor in its responses

136.7K

10:54 PM • Nov 3, 2023

The healthcare niche is huge. I remember Malariaspot, a game from 2012 where, as part of the game, players had to find and identify malaria parasites. It was shown that 22 players together were as reliable as an expert microscopist.

Now imagine how this could look like with AI and trained human labelers. To show a more recent example, in April 2024, iMerit launched advanced annotation tools for medical images, using AI to assist in accurate labelling for diagnostics and treatment.

This kind of data is invaluable for training AIs in medical imaging and diagnostics. It would assist doctors in making faster, more accurate diagnoses and make them available in remote places where there might not be enough health professionals.

Autonomous vehicles rely on labeled image and video data to "see" the world around them, helping, for example, cars make safer driving decisions.

Another example is FrodoBots, a robotics company best known for its Earth Rovers program, where participants are tasked to remotely operate its sidewalk robots in cities worldwide. The data from these test drives helps advance research in Embodied AI.

Another big one. Video and audio labeling are crucial for security applications, from identifying suspicious behaviors in surveillance footage to recognizing potential threats in audio recordings.

And yes, it works in retail too. Data annotation can enhance customer experiences, improve inventory management, and personalize recommendations, optimizing product placements and promotions.

Not everything is bright and shiny in data labeling and annotation. While it is essential for future development, it comes with challenges that can affect both quality and cost. Addressing these obstacles is crucial for creating reliable and trustworthy AIs.

The obvious one is quality and consistency. I’ve already mentioned it in previous sections, but ensuring high-quality, consistent labeling is critical. Human errors and inconsistencies, as well as imperfect algorithms, can reduce data quality. 

Semi-automated data annotation with rigorous quality assurance processes are necessary to maintain accuracy.

Bias in labeled data can lead to biased AI systems, which may make unfair decisions or reinforce stereotypes. If we feed AI biased data, it'll pick up those bad habits. Addressing these biases and adhering to ethical guidelines is essential for creating fair, inclusive AI.

As data needs grow, labeling projects require more resources, people, and technology. 

And let me be clear. There’s huge demand for data annotation right now, and it won’t stop any time soon. Each day, over 4 billion people use the internet, generating about 3 quintillion bytes of data. That’s raw unstructured data. 

Data annotation is a market projected to reach $5,842 million by 2033, growing at a CAGR of 34.4% from 2024 to 2033.

Automation can help, but finding a scalable solution that maintains quality is still an ongoing challenge.

When you're dealing with this much data, you've got to find smart ways to get things done faster. Efficiency is a priority. Data labeling can be time-consuming and expensive. Careful planning, properly trained workers, and choosing efficient tools is key to manage these costs.

This is a special challenge. It might not directly apply to data labeling, but it does to its source: data itself.

I mean, to annotate data, you need to get data before. And sometimes such data comes from you.

Labeled data often includes sensitive information that must be protected to respect individuals' privacy, and maintain public trust. So with justifiable growing concerns about privacy, data annotation projects must prioritize data security and regulatory compliance.

Legal frameworks like GDPR in Europe, CCPA in the USA, or PIPEDA in Canada mandate strict guidelines on data privacy. And data labeling and annotation projects must obviously comply with these laws to avoid legal repercussions.

But there’s some friction. For example, a principle emphasized in the GDPR is data minimization. It involves collecting and retaining only the personal data necessary for the specified purpose. But for AI development and innovation, you want as much data as you can get. So where do you trace the line?

Other anonymization and data protection techniques that can be used when you do collect data include pseudonymization, encryption, data masking, tokenization and differential privacy. 

It’s an evolving field. Tech is zooming ahead at light speed, but rules to make it safe cannot keep up. Regulatory organizations take time to make decisions, and bureaucracy makes them even slower.

To have a bright and safe future, data projects and rule-makers will need to work together and iterate a few times to get to a solid and useful agreement.

AI opened a new frontier for data annotation, but it’s not the only new technology that affects it. Decentralization through blockchains promises a future where users have more control over their data.

Let me explain how this all ties together.

For the uninitiated, blockchains are secure, transparent, and (frequently) decentralized ledgers.

I’m simplifying here, but blockchains are like a database. One that it’s open. One that’s not kept in the vault of a big corporation, but distributed among many computers (known as nodes) around the world.

This decentralization enables services that are owned and controlled by users rather than corporations. Thanks to blockchains, we have private property on the internet.

As Chris Dixon says, we can see the internet in three acts. 

• In the first one, the “read era” (circa 1990-2005) anyone could read about almost any topic through websites.

• In the second act, the “read-write era” (roughly 2006-2020), anyone could write and publish to mass audiences through posts on social networks and other services.

• Now we’re in the “read-write-own era,” and you can also own digital assets on the internet.

In a way, it's like the background values for Data Privacy and for Web3 are cousins. Both are about ownership, about taking control… And we can only have it together, because due to network effects, together, your privacy is not shared, but strengthened individually. Each individual who joins does not take. It empowers you, but also everyone else before, and after you.

So web3 allows users (a.k.a. YOU) to have more control and ownership over their data. You can now do more than simply clicking a button that says “Yes, I accept those cookies”. You can decide what and how your data is used.

And you can be rewarded for it.

A new data economy will arise. One where people (and not just big data corporations) are incentivized to participate.

For example, Navigate is giving you a chance to start doing this. You can select which data sources you're comfortable sharing, and start gaining points passively while you browse like you normally do. Or you can play data quests that will reward you with even more points while having a bit of fun.  

Data labeling and annotation is evolving rapidly, and its future is promising.

There’s huge demand for it, and there are obvious applications within reach.

The use of AIs and blockchains in data annotation models are likely to streamline processes and enhance reliability across many different industries such as healthcare, finance, security or retail.

We need to prepare for the next wave of innovation. That means having proper legal frameworks for data management. But the impact of high-quality labeled data cannot be overstated, so choosing the right tools and platforms to optimize efficiency is also key.

Those who invest in themselves and understanding these processes will be better equipped to thrive in an AI-driven world. The age of data is here—let's ensure it's as powerful and ethical as possible.