Everyone Is an AI Expert Now. But What Does “AI Expert” Actually Mean?

There is something strange happening around artificial intelligence.

The more powerful AI becomes, the more people seem to describe themselves as “AI experts.”

But what does that actually mean?

Are you a programmer?

Do you design machine-learning algorithms?

Do you train models?

Do you build datasets?

Do you work on GPUs and infrastructure?

Do you develop applications using existing models?

Do you decide which content a model can generate?

Do you work on safety and alignment?

Do you evaluate models?

Do you simply use ChatGPT every day?

These are not the same profession.

And yet they are increasingly placed under the same label.

That distinction matters because the AI economy is becoming much more sophisticated than the public debate suggests.

1. Using AI is not the same as building AI

There are at least four very different levels of participation in the AI economy.

Level 1 — AI User

A professional uses AI to research, write, analyse, code, summarize documents, automate tasks or accelerate decision-making.

This can be extremely valuable.

But it does not make someone an AI engineer or AI researcher.

Level 2 — AI Integrator

This person builds workflows and applications around existing models.

They may use APIs from OpenAI, Anthropic, Google, Meta, DeepSeek or other providers.

They can build sophisticated systems without having designed the underlying model.

Level 3 — AI Engineer

This is a much deeper technical category.

It includes machine-learning engineers, research engineers, data engineers, infrastructure engineers, GPU specialists and software engineers working directly on AI systems.

Level 4 — AI Researcher

This is the frontier.

Researchers work on architectures, training methods, optimization, reasoning, reinforcement learning, multimodality, evaluation, alignment and new machine-learning techniques.

These four categories are all important.

But they are not interchangeable.

Calling all of them “AI experts” without qualification tells us very little.

2. The AI industry is already highly specialized

The AI ecosystem includes a surprisingly large number of professions:

AI/ML Research Scientists

Develop new algorithms and approaches to machine learning.

Machine Learning Engineers

Turn models into reliable production systems.

Research Engineers

Bridge fundamental research and engineering.

Data Scientists

Develop statistical and predictive models and extract information from data.

Data Engineers

Build the pipelines and infrastructure that make data usable.

AI Infrastructure Engineers

Work on GPUs, networking, distributed computing, storage and inference infrastructure.

AI Safety and Alignment Researchers

Study how models can be made more reliable, controllable and safe.

Evaluation and Red-Team Specialists

Try to break models, identify weaknesses and measure their capabilities.

AI Product Managers

Turn technical capabilities into products.

AI Application Developers

Build applications around existing foundation models.

AI Governance, Legal and Policy Specialists

Deal with regulation, privacy, copyright, security and accountability.

And finally:

AI Power Users

People who use AI extremely effectively in their professional work.

The last category is growing enormously.

But it should not be confused with the first categories.

3. The frontier is being built by a very small part of the global workforce

The scale of AI adoption is enormous, but the number of people actually building frontier models is comparatively small.

Stanford’s 2026 AI Index estimates that industry produced 91.2% of notable AI models in 2025, compared with only a small number originating from academia. OpenAI produced 20 notable models, Google 14 and Alibaba 11 in that year’s dataset. (Stanford HAI⁠)

This tells us something important.

The AI revolution is not being produced by “everyone working in AI.”

It is being driven by a relatively concentrated group of companies, researchers, engineers, infrastructure providers and investors.

The rest of the economy is primarily becoming an AI adopter.

That distinction will become increasingly important.

4. The United States still dominates the industrial ecosystem

The United States remains the world’s leading AI industrial hub.

Stanford reports approximately $285.9 billion in private AI investment in the United States in 2025, compared with approximately $12.4 billion in China.

The U.S. also had 1,953 newly funded AI companies in 2025, more than ten times the next country. (Stanford HAI⁠)

The U.S. also hosts the world’s largest concentration of AI data-center infrastructure.

Stanford estimates 5,427 data centers in the United States — more than ten times the number in any other country. (Stanford HAI⁠)

But there is a potential strategic weakness.

The United States is still home to more AI researchers and developers than any other country, yet the flow of AI researchers moving into the country has fallen dramatically.

Stanford estimates that the number of AI researchers and developers moving to the U.S. has declined 89% since 2017. (Stanford HAI⁠)

That may matter enormously over the next decade.

Capital can build data centers.

Capital can buy GPUs.

Capital can finance companies.

Talent is much harder to manufacture.

5. China is no longer simply “behind”

The traditional narrative was simple:

America develops AI. China follows.

That narrative is increasingly obsolete.

Stanford’s 2026 AI Index states that the U.S.-China AI model performance gap has effectively closed, with models from the two countries repeatedly exchanging the lead since early 2025. (Stanford HAI⁠)

China leads the world in:

• AI publication volume;

• citations;

• patent output;

• industrial robot installations.

The United States remains ahead in:

• notable frontier model production;

• higher-impact patents;

• private investment;

• AI startup financing;

• major AI infrastructure. (Stanford HAI⁠)

The result is not an American monopoly.

It is increasingly a two-pole AI ecosystem.

6. Europe has a different problem

Europe has excellent universities, researchers, engineers and scientific institutions.

But scientific capability and industrial dominance are not the same thing.

In 2025, Stanford counted 59 notable AI models associated with the United States, 35 with China and only a handful across individual European countries. (Stanford HAI⁠)

Europe therefore faces a structural problem:

How do you turn scientific excellence into technological scale?

That requires more than researchers.

It requires:

• capital;

• GPUs;

• data centers;

• energy;

• startups;

• large markets;

• engineering talent;

• risk-taking;

• management;

• commercialization.

Europe has many of these components.

The problem is that they are less concentrated and less effectively connected than in the United States.

7. Asia is much larger than China

Asia should not be treated as a single AI market.

China, India, Japan, South Korea and Singapore have very different strategies.

China combines enormous industrial capacity with research scale.

India combines a huge software and engineering workforce with rapidly increasing AI adoption and technical specialization.

South Korea combines AI with semiconductors, manufacturing and robotics.

Japan has enormous industrial, robotics and engineering capabilities.

Singapore has one of the world’s strongest concentrations of AI talent per capita.

The Middle East is also becoming increasingly relevant.

The Stanford AI Index highlights particularly rapid growth in AI engineering skills in countries including the United Arab Emirates, Chile and South Africa. (Stanford HAI⁠)

The geography of AI is therefore changing.

8. Latin America and Africa should not be measured only by frontier models

There is another mistake in the discussion.

People often measure AI power by asking:

Who builds the largest model?

But economic value can also come from adoption.

A country does not need to build GPT-class infrastructure to use AI to transform its economy.

A company in Latin America can use an American, European or Chinese model.

A financial institution in Africa can use AI for risk analysis.

A manufacturer in Asia can use AI for predictive maintenance.

A professional in Europe can use AI to accelerate research.

The model can be developed somewhere else.

The productivity gain can occur somewhere completely different.

This distinction is critical:

AI development ≠ AI adoption.

9. There is no reliable global percentage of “workers in AI”

This is where many discussions become statistically misleading.

There is no single authoritative number showing:

“X% of the global workforce works in AI.”

Different datasets measure different things:

• AI researchers;

• AI developers;

• AI talent;

• AI-related occupations;

• AI adoption;

• exposure to generative AI;

• AI-related job postings.

They should not be combined into one number.

The International Labour Organization estimates that roughly one in four workers globally are in occupations with some exposure to generative AI.

But exposure does not mean replacement.

The ILO’s research suggests that transformation of jobs is considerably more likely than complete automation of occupations.

This is a crucial distinction.

10. AI exposure is dramatically different across economies

Higher-income economies are more exposed because their workers are disproportionately employed in occupations involving information processing, administration, professional services and knowledge work.

Lower-income economies have much lower exposure.

This means AI will not affect every country in the same way.

The first wave of disruption will probably be concentrated in knowledge-intensive economies.

The second question will be how quickly less-developed economies adopt AI tools and infrastructure.

This could create an unexpected opportunity.

AI may allow smaller economies to access sophisticated capabilities without first building the entire technological stack themselves.

11. Which AI jobs are actually growing?

The World Economic Forum’s Future of Jobs Report 2025 identifies several technology-related roles among the fastest-growing professions through 2030.

They include:

• Big Data Specialists;

• FinTech Engineers;

• AI and Machine Learning Specialists;

• Software and Application Developers;

• Security Management Specialists.

The WEF also identifies AI and big data among the fastest-growing skill areas. (World Economic Forum⁠)

This is revealing.

The future of AI is not simply about AI researchers.

It is about the combination of:

AI + data + software + cybersecurity + domain expertise.

That is much more important for the real economy.

12. Now comes the interesting question: which AI platform is actually better?

The answer is:

Better for what?

The five ecosystems I would compare are:

OpenAI

Strengths:

• general-purpose reasoning;

• coding;

• multimodality;

• tool use;

• agents;

• ecosystem integration.

Weaknesses:

• proprietary models;

• limited control over underlying weights;

• platform policies;

• dependence on provider infrastructure.

Anthropic / Claude

Strengths:

• reasoning;

• coding;

• long-form professional work;

• document analysis;

• strong emphasis on safety and reliability.

Weaknesses:

• proprietary models;

• strong policy restrictions;

• limited control over underlying model weights.

Google / Gemini

Strengths:

• multimodality;

• enormous context;

• Google ecosystem;

• Search;

• Cloud;

• Workspace;

• Maps;

• Android;

• enterprise infrastructure.

Gemini’s greatest strategic advantage may not be the model itself.

It may be Google’s ecosystem around the model.

Meta / Llama

Different philosophy.

Llama’s importance is largely related to open-weight distribution and the ability for developers to deploy and customize models more independently.

That creates a fundamentally different relationship between:

model → developer → infrastructure → application → user.

More control.

More flexibility.

But also more responsibility.

And “open-weight” does not automatically mean “completely unrestricted.”

DeepSeek

DeepSeek demonstrated another important possibility:

competitive performance at radically lower costs.

Its rise has put enormous pressure on the assumption that frontier AI necessarily requires the same cost structure used by the largest American laboratories.

13. Performance is only one dimension

When comparing AI platforms, I would use at least seven criteria:

1. Reasoning performance

Can it solve complex problems?

2. Coding performance

Can it actually build and debug software?

3. Context

How much information can it process reliably?

4. Multimodality

Can it work with text, images, audio, video and other inputs?

5. Cost

How expensive is it at scale?

6. Control

Can the user modify, fine-tune, host or deploy the model independently?

7. Content restrictions

What does the provider allow or refuse?

The last category is particularly misunderstood.

14. “Censorship” is not one thing

When someone says:

“This AI is censored.”

I would ask:

By whom?

There are several layers.

Layer 1 — The model

The training and post-training can influence behaviour.

Layer 2 — The provider

The company establishes policies.

Layer 3 — The product

The consumer application may impose restrictions that differ from an API.

Layer 4 — The jurisdiction

Local laws can impose additional restrictions.

Therefore:

model behaviour ≠ provider policy ≠ product policy ≠ national regulation.

And:

open-weight ≠ unrestricted.

This distinction is essential.

15. The five platforms represent five different philosophies

OpenAI is building an integrated AI platform.

Anthropic is emphasizing reasoning, coding and safety.

Google is integrating AI into one of the world’s largest technological ecosystems.

Meta is pushing open-weight models and decentralized deployment.

DeepSeek is putting enormous pressure on the economics of AI.

There is therefore no single “best AI.”

There is only:

the best AI for a particular task.

The correct question is not:

“Which model is the smartest?”

It is:

“Which model gives me the best combination of performance, cost, control, reliability and freedom for the task I need to perform?”

16. And now my own conclusion

I am increasingly skeptical of the idea that simply knowing how to use AI makes someone an “AI expert.”

AI is becoming an extraordinary professional tool.

But it remains a tool.

For me, its greatest value is practical:

research faster.

analyse more information.

compare sources.

automate repetitive processes.

generate first drafts.

write and debug code.

build scenarios.

accelerate workflows.

That is enormously powerful.

But the output still needs to be evaluated.

AI can produce an answer in seconds.

It does not automatically make the answer correct.

That requires knowledge.

Experience.

Critical thinking.

Source verification.

Professional judgement.

And ultimately, responsibility.

The real AI skill of the future

Perhaps the most important AI skill will not be:

“I know how to use AI.”

It will be:

“I know when to use it, which model to use, how to verify the result and how to integrate it into my professional expertise.”

That is a completely different concept.

A programmer is not automatically an AI researcher.

A prompt engineer is not automatically an AI engineer.

A manager who uses ChatGPT is not automatically an AI specialist.

A researcher is not automatically a successful AI entrepreneur.

And an AI user is not necessarily an AI expert.

The competitive advantage will increasingly belong to people who combine:

domain expertise + analytical ability + AI + human judgement.

That is where I believe the real economic value of AI will emerge.

AI can accelerate the process.

Human competence still determines where the process should go.

Roberto Balbino — CEO, Fintech SBB
More than 30 years of experience in finance, financial markets and international business.
Independent analyst of finance, economics, geopolitics and global markets.