When AI Depends on the Internet, Who Gets Left Behind?

The future of artificial intelligence is being built around a simple assumption: that everyone will always be connected. For millions of people, that assumption is already wrong.

There is a moment in the way we use artificial intelligence that most of us never think about.

You point your phone at something. You speak into a microphone. You turn on the camera. An AI system recognizes what you are showing it, understands what you said, or translates what you are doing.

It feels immediate.

But often, the first thing that happens is that your information leaves your phone.

It travels through the internet to a company’s computers somewhere else. The computers process it and send the answer back.

For someone with fast, affordable internet and a modern phone, this arrangement can feel invisible. For someone living with expensive data, unreliable connectivity, or government-imposed internet shutdowns, it can determine whether the technology works at all.

That raises a question that deserves much more attention:

What happens when we build AI for people who cannot always afford, access, or trust the internet it depends on?

The internet is not equally available to everyone

The technology industry often talks about connectivity as though it were a solved problem.

It isn’t.

In 2024, about 68 percent of the world’s population was online. In Africa, the figure was only 38 percent. In low-income countries, it was 27 percent. The International Telecommunication Union also found that a basic 2GB mobile data plan in Africa cost about 3.9 percent of average monthly income in 2024—almost twice the 2 percent affordability target set by the UN Broadband Commission.

These numbers are easy to read and forget.

But imagine an AI application that needs an internet connection every time you use it.

For a wealthy user with unlimited broadband, that requirement barely exists.

For someone who has to decide whether to spend money on data or food, it is a different product.

For someone in a rural area with an unstable connection, it is a different product.

And for someone living through an internet shutdown, it is not a product at all.

That last point is becoming harder to dismiss.

In 2025, Access Now and the #KeepItOn coalition documented at least 313 internet shutdowns across 52 countries, the highest number they have recorded. The shutdowns occurred during conflicts, political crises, protests and other periods when communication was often particularly important.

In Africa alone, they documented 30 shutdowns across 15 countries.

So when we build AI that assumes the internet will always be there, we are not simply making a technical choice.

We are making an assumption about who gets to use the technology.

What happens when personal data leaves your phone?

There is another problem.

Many AI applications need access to information that people would reasonably consider private.

A camera may see your face. A microphone may hear your conversation. An accessibility application may observe your movements. A health application may receive information about your body.

When that information is sent to a remote server, someone else has to store and process it.

That does not automatically make the technology unsafe. Companies can build strong security systems, delete data quickly and put strict limits on how information is used.

But the risk exists because the data exists somewhere outside the user’s control.

Afghanistan offers a particularly painful example of what can happen when control over a database changes.

For years, Western governments and aid organizations helped create systems containing sensitive biometric information about Afghans. When the Taliban took control of Afghanistan in 2021, they gained control of systems containing this information. Human Rights Watch warned that the data could put thousands of Afghans at risk.

The lesson is not that every AI company is going to hand your data to a government.

The lesson is simpler:

Data collected for one purpose can eventually be controlled by someone else.

Political circumstances change. Governments change. Companies are acquired. Servers are hacked. Laws change.

The safest piece of sensitive information is often the piece that was never collected in the first place.

What if the AI stayed on the phone?

This is where the idea of running AI directly on a device becomes interesting.

Instead of:

Camera → Internet → Server → AI → Phone

the system can work more like:

Camera → Phone → AI → Answer

The difference sounds small.

It isn’t.

If the AI can do its job entirely on the phone, there may be no reason to send the original video, photograph or recording anywhere.

The phone can understand it locally.

And if the model does not need an internet connection, the application can continue working when the network disappears.

This is particularly important for accessibility.

Consider sign-language translation.

A person could use a phone’s camera to communicate through sign language. A conventional system might send that video to a server, where a large AI model interprets it.

But researchers are increasingly exploring ways to move more of that work onto the device.

Arm, for example, has worked on adapting a sign-language model so that parts of the translation process can run on mobile hardware. The project demonstrates both the promise and the difficulty of bringing AI onto phones: models have to become smaller and more efficient, and sometimes there is a trade-off between speed, battery use and accuracy.

Other researchers have gone further. A 2025 study introduced SignEdgeLVM, a sign-language translation model designed with smaller devices in mind. The researchers reported a substantial reduction in the memory required by its attention mechanism, making the model more suitable for less powerful hardware.

This matters because accessibility technology is supposed to remove barriers.

It would be strange to remove one barrier—communication—and replace it with another: you need a reliable internet connection before you can communicate.

Even privacy improvements can leave us dependent on the cloud

There is an interesting example in Google’s recent sign-language research.

Google says its system was trained using more than 100,000 hours of data covering more than 50 sign languages. To reduce privacy concerns, the system first processes the video on the device and turns the person’s movements into numerical descriptions of their body and hands. The original video can then be discarded, while those descriptions are sent to a server for translation.

That is a meaningful improvement.

The user’s raw video does not have to leave the phone.

But it still requires a server.

If the internet disappears, the service cannot complete the translation.

That distinction matters.

There is a big difference between sending less personal data to the cloud and not needing the cloud at all.

The first improves privacy.

The second can improve both privacy and access.

But should everything run on the device?

No.

This is where the conversation needs some honesty.

Cloud computing is not inherently bad.

Large servers can run models that a phone simply cannot. They can process enormous amounts of information and provide capabilities that would be impossible on inexpensive devices.

And on-device AI has its own limitations.

A phone has limited battery, memory and processing power. Smaller models may be less accurate. Updating models can be difficult. Developers may have to support many different devices.

So the answer is not to throw away cloud AI.

The better question is:

Which parts of an AI system actually need the cloud?

If an application can perform an important task locally, why should an internet connection be mandatory?

If sensitive information can be processed on a phone without being uploaded, why collect it?

If an accessibility tool is being built for communities with unreliable connectivity, why make constant connectivity a requirement?

These are not merely engineering questions.

They are policy questions.

We need to stop treating connectivity as a universal assumption

This is where governments, funders and technology organizations have an important role.

When public money funds an AI system for healthcare, education, accessibility or humanitarian work, the question should not only be:

Does the technology work?

It should also be:

Who can actually use it?

A system that works beautifully in a well-connected city but fails completely without internet access may look successful in a demonstration while remaining useless to the people it was supposed to help.

That does not mean every publicly funded AI system must be completely offline.

It means offline capability should be considered when it is technically and economically possible.

The same should apply to privacy.

Organizations should ask whether they really need to collect people’s raw images, voices, movements or other sensitive information—or whether the technology could process that information locally and simply return the result.

That is a much better question than:

“How much data can we collect?”

The people most affected should have a say

There is another issue that is easy to miss.

Technology designed for marginalized communities is too often designed about those communities rather than with them.

A sign-language system should not be considered successful simply because its model achieves a high score in a research paper.

Does it work for the people who actually use the language?

Does it work with different signing styles?

Does it work on affordable phones?

Does it work without a permanent internet connection?

Does it respect the privacy of the people using it?

And perhaps most importantly:

Did the people who will depend on it have a meaningful role in deciding how it should work?

These questions can change what gets built.

The real digital divide may soon be about AI itself

For years, we have talked about the digital divide as a question of who has internet access.

AI could create another divide.

One group will have powerful AI assistants that work instantly, understand their voices and surroundings, and operate across expensive cloud infrastructure.

Another group may have AI tools that require a constant connection, consume expensive data and stop working whenever the network disappears.

That is not inevitable.

Phones are becoming more capable. AI models are becoming smaller. Researchers are finding ways to run increasingly sophisticated systems locally.

The technology is moving in the right direction.

But technology alone will not decide who benefits.

Policy will. Procurement will. Design choices will. And the assumptions we make today will shape who gets left behind tomorrow.

The goal should not be to eliminate cloud AI.

It should be to make sure that being offline does not mean being locked out of AI.

For accessibility tools, healthcare applications, translation systems and other technologies designed for people who may already face barriers, that distinction could be the difference between technology that looks impressive and technology that actually works.

The question is no longer simply how powerful AI can become.

It is whether that power will still be available when the internet isn’t.