AI may look intelligent from the outside. Behind that intelligence is a lot of very human work.
For a long time, when I heard people talk about artificial intelligence, I mostly thought about the tools.
ChatGPT. Image generators. AI assistants. Automation. The increasingly impressive things these systems can produce with a few words typed into a box.
What I didn’t fully appreciate was everything happening behind the scenes to make those systems useful.
Someone has to prepare the data.
Someone has to label it.
Someone has to determine whether an AI response is accurate, relevant, safe, or simply nonsense.
Someone has to compare two answers and decide which one is better.
Someone has to listen to an audio clip and determine whether the sound is speech, music, or a sound effect.
Someone has to watch a video and decide where one scene ends and another begins.
And sometimes, someone has to look at an AI-generated answer and essentially say:
“No. That’s not good enough.”
I became one of those people.
Not as a machine-learning engineer. Not as a data scientist.
As a contractor learning how humans help train, evaluate, and improve AI systems.
And the experience taught me much more than I expected.
AI doesn’t learn by magic
One of the biggest things I learned is that AI is not simply a machine that becomes intelligent on its own.
There is an enormous human component behind the systems we interact with.
In traditional machine-learning tasks, human workers may identify objects in images, label sounds, classify content, or draw bounding boxes around objects.
With large language models, the work can become much more nuanced.
People may evaluate responses, write prompts, compare answers, assess reasoning, check search results, identify inconsistencies, and judge whether a response actually follows the instructions it was given.
The distinction is fascinating.
Traditional machine-learning annotation often asks:
“What is this?”
LLM-related evaluation increasingly asks:
“How well did the model respond?”
That difference changes the kind of thinking required from the human evaluator.
And I learned that firsthand.
What I actually learned
My training exposed me to several different areas of AI data work.
There was audio annotation, where the instructions could be incredibly specific.
Is the person speaking English?
Is it actually speech, or is it singing?
Is that background music?
Is it a sound effect?
Does a mechanical sound deserve its own annotation?
Where exactly does the sound begin and end?
Then there was video segmentation.
A scene isn’t simply “whatever happens in the next few seconds.”
You have to understand the rule.
A scene changes when the physical setting changes—not merely because the camera angle changes or a person moves within the same environment.
Then there’s camera movement segmentation.
Static. Pan. Zoom. Dolly. Tracking.
When the camera behavior changes, the annotation may need to change with it.
And then there is action segmentation, which taught me another important lesson:
A continuous activity isn’t necessarily one annotation.
If a person places a skateboard, steps onto it, and then rides away, those can represent separate atomic, intentional actions.
The work requires you to stop thinking like a casual viewer and start thinking like someone documenting events for a machine.
That shift is harder than it sounds.
Then came the LLM side of the work
This was where things became even more interesting.
Prompt writing. Rationale evaluation. Pairwise evaluation. Search evaluation.
Multimodal evaluation. Reverse engineering. Image evaluation.
These aren’t necessarily tasks where there is always one obvious answer.
Sometimes you’re evaluating quality.
Sometimes you’re comparing two responses.
Sometimes you’re asking whether a model’s explanation actually supports its conclusion.
Sometimes you’re evaluating whether the visual and audio information in a video correspond correctly.
And that requires judgment.
Not just clicking.
Not just typing.
Judgment.
That was probably one of the biggest lessons I took away from the experience.
The part people don’t always see: the contractor reality
But there’s another side of AI data work that I think deserves to be discussed honestly.
The work can be interesting.
It can expose you to technologies you might never otherwise encounter.
It can give someone without a traditional AI or computer-science background a way to enter the AI ecosystem.
But it can also be highly uncertain, particularly when you’re working as a contractor.
My experience taught me not to confuse being accepted into an AI project with having a stable AI career.
Project-based work can mean that the availability of tasks changes.
You can spend considerable time learning guidelines, completing assessments, earning certifications, and preparing yourself for production work—only to find that the actual volume of available tasks is inconsistent.
Sometimes the work is there.
Sometimes it isn’t.
And when you’re paid based on completed tasks or project availability, that uncertainty matters.
You can have the skills and still have nothing to work on.
That was one of the hardest lessons for me.
The hours can be longer than they look
From the outside, data annotation can sound simple.
“Just label the data.”
But good annotation isn’t always fast.
You have to read the guidelines carefully.
You have to understand exceptions.
You have to make consistent decisions.
You may have to review examples.
You may have to complete certifications before being allowed onto a particular task.
And when quality standards are strict, rushing can be counterproductive.
There is a strange tension in this kind of work:
You are expected to be accurate, but efficiency matters too.
That means you are constantly balancing speed with quality.
And when the compensation is modest, spending significant amounts of time on training, qualification, review, and task preparation can become difficult to justify financially—especially for contractors who don’t receive the same benefits and protections as regular employees.
That’s something people considering this type of work should understand before jumping in.
The compensation question
I also think we need to have a more honest conversation about money.
AI is a multi-billion-dollar industry.
That doesn’t mean every person doing the human work behind AI is highly compensated.
There is a very large range in this field depending on the role, specialization, employer, geography, project, and whether someone is an employee or contractor.
The Philippines is attractive to companies looking for AI and data talent partly because of its strong BPO ecosystem, English proficiency, and growing technical workforce. Industry sources increasingly describe the country as an important destination for AI annotation and related services.
But workers need to look beyond the headline:
“AI jobs are growing.”
Ask instead:
- How much does the work actually pay?
- Is it hourly, task-based, or project-based?
- Are training and qualification periods compensated?
- How predictable is the workload?
- Are there employee benefits?
- What happens when a project ends?
- Can the skills transfer to other roles?
- Is there a pathway to more specialized work?
Those questions matter.
And then there’s instability
This is probably the biggest warning I would give someone considering AI annotation as their only source of income.
Don’t assume that because AI is growing, every AI annotation project will remain stable.
The industry changes quickly.
Models change.
Client requirements change.
Projects end.
Guidelines change.
Companies adjust their workflows.
Some work becomes automated.
Other work becomes more specialized.
Recent reporting on AI training work has also highlighted inconsistent project availability, changing guidelines, pay changes, and the uncertainty faced by freelance AI trainers and annotators.
That doesn’t mean the field is bad.
It means you need to understand what kind of work you’re actually entering.
For me, the lesson wasn’t:
“Don’t do AI annotation.”
It was:
“Don’t build your entire financial future around one project-based AI contract.”
There’s a big difference.
Why I still think the opportunity is worth exploring
Despite everything I’ve said, I don’t regret doing it.
Not at all.
In fact, I’m glad I did.
Because I now understand something I didn’t understand before:
AI isn’t just a tool I use.
I have now seen a little bit of what happens behind the tool.
I understand why data quality matters.
I understand why annotation guidelines can be extremely precise.
I understand why human judgment remains important.
I understand why evaluation is necessary.
And I understand that the future of AI isn’t going to be built entirely by engineers sitting behind powerful computers.
There is a much larger ecosystem behind it.
The Philippines is already positioning itself for that ecosystem. The government has been actively promoting data-center and AI infrastructure investments, while agencies such as the Board of Investments have highlighted growing demand for digital infrastructure driven by AI and cloud computing.
The Department of Labor and Employment’s Institute for Labor Studies has also noted that AI adoption is reshaping the Philippine labor market and creating demand for new skills.
And here in Western Visayas, DOST is already implementing regional high-performance computing facilities at institutions including ISAT U, Aklan State University, Northern Iloilo State University, WVSU, and DOST Western Visayas as part of the region’s AI development plans.
So yes, something is happening.
Even Iloilo is becoming part of the conversation
This became particularly real for me when a proposed hyperscale AI data center was announced for Oton, Iloilo.
The proposal generated significant local discussion around infrastructure, water, electricity, land, and environmental impact. The proposed facility was described as a 50-MW hyperscale AI data center on a large site in Oton.
There is an important footnote, though.
As of August 2026, that particular proposal was withdrawn by its proponent after the public consultation.
So I wouldn’t use the Oton proposal as evidence that an AI data center is definitely coming to Iloilo.
I would use it as evidence of something else:
AI infrastructure is no longer a distant conversation happening somewhere in Silicon Valley.
We’re talking about it here.
We’re asking what it means for our communities, our electricity, our water, our land, our jobs, and our future.
And that conversation is important.
What I would tell someone who wants to enter AI annotation
If you’re curious about this field, I wouldn’t tell you to stay away.
I’d tell you to enter with your eyes open.
Learn the fundamentals.
Understand annotation guidelines.
Improve your English and communication skills.
Develop attention to detail.
Learn how AI evaluation works.
Practice prompt writing.
Learn how to evaluate AI responses rather than simply generate them.
Explore multimodal tasks.
And most importantly:
Don’t stop at basic annotation if you discover that you enjoy this kind of work.
Annotation can be an entry point.
From there, you can explore areas such as:
- AI evaluation
- Quality assurance
- Data quality
- Prompt evaluation
- LLM response evaluation
- Search evaluation
- AI operations
- AI training
- Workflow automation
- AI-assisted business operations
The more specialized your judgment becomes, the more valuable your contribution can potentially become.
What I took away from the experience
I intentionally expanded into AI operations and annotation while continuing my executive support career.
I didn’t abandon the skills I’d spent years developing.
I added something to them.
My background taught me how to organize information, follow processes, communicate clearly, notice details, manage competing priorities, and support people.
AI annotation taught me to apply those abilities in a completely different environment.
And perhaps that is the most important lesson I took from the experience.
You don’t always have to start over to move into a new industry.
Sometimes you take what you already know and build another layer on top of it.
That’s what I’m doing.
I invested time learning how modern AI systems are evaluated, improved, and supported through high-quality human annotation and review.
And even though the contractor experience wasn’t always stable, the knowledge stayed with me.
That matters.
So, is AI data annotation worth it?
My answer is:
It can be. But know what you’re signing up for.
It can be a fascinating way to enter the AI ecosystem.
It can give you practical exposure to technologies that are shaping the future of work.
It can help you develop skills that transfer into other AI-related roles.
But it may also involve long hours, repetitive or highly detailed work, strict quality requirements, modest compensation depending on the project, limited benefits for contractors, and periods when there simply isn’t enough work.
For that reason, I wouldn’t recommend treating project-based annotation as a guaranteed job.
I’d treat it as one piece of a larger career strategy.
Learn from it.
Build your portfolio.
Document what you know.
Collect transferable skills.
Keep looking for opportunities.
And don’t put all your eggs in one AI-shaped basket.
Because the AI industry is moving fast.
And the safest place to be isn’t necessarily at the center of one particular project.
It’s having enough skills to move when the project does.
The human work behind AI is real.
So are the opportunities.
But so are the risks.
And I think both sides of that story deserve to be told.
This article reflects my personal experience as a project-based AI data annotation contractor and is not intended to represent the policies, compensation, working conditions, or practices of any specific company or platform.
If you’re considering AI annotation or evaluation work, I hope this gives you a clearer picture—not to discourage you, but to help you enter the field with realistic expectations.
About The Mari Effect
The Mari Effect is about making sense of change—building clearer systems, learning new tools, and finding practical ways to work better in a world that’s changing fast.
Because technology keeps evolving.
But people still make the difference.
Have You Ever Wondered What Happens Behind the AI We Use Every Day?
I’d love to hear your perspective.
AI can look almost effortless from the outside. But behind many of the systems we use are people reviewing data, evaluating responses, checking quality, and helping machines learn what good looks like.
I’ve had the opportunity to experience a small part of that world firsthand.
If you’re exploring AI work, navigating a career transition, or simply curious about what happens behind the technology, I hope this story gives you something useful to take with you.
Maybe you’re discovering a new skill yourself. Maybe you’re wondering where all this learning is leading.
Sometimes the pieces make more sense when we stop and look at what we’ve already built.
