Are We Using AI to Write Better Emails?

A strange use of power

We are told that artificial intelligence may help discover new medicines, accelerate science, improve education, and expand what human beings are able to understand. Then we open the product and ask it to make an email sound more professional.

That contrast is easy to mock. It is also revealing. Some of the best-funded and most visible AI products are personal assistants for managing messages, summarizing meetings, creating presentations, and producing marketing copy. These tools save time. Many are genuinely useful. But the distance between the rhetoric and the reality remains striking.

The industry describes a new industrial revolution. The user receives a more polished paragraph.

This does not mean AI has failed. It means we should distinguish technical capability from social progress. A technology can become remarkably capable while the systems around it direct most of that capability toward tasks that are easy to sell, easy to measure, and unlikely to change very much.

What people actually do

Available usage data complicates the claim that almost everyone is merely improving grammar, but it does not entirely dismiss the intuition behind it.

OpenAI's study of consumer ChatGPT use found that practical guidance, seeking information, and writing accounted for roughly three quarters of conversations. Writing was the most common work-related use, representing 42 percent of work messages. Most of those writing requests involved modifying text supplied by the user rather than creating something entirely new. ^[1]

Anthropic's data reflects a different user population. Coding remains the largest category on Claude, accounting for about 35 percent of Claude.ai conversations in its February 2026 sample. Personal use was growing, while uses were becoming more diverse rather than concentrating around a single breakthrough application. ^[2]

Neither dataset describes all AI use, and both come from companies with an interest in demonstrating value. Still, they point in the same broad direction. AI is useful across many tasks, but its widespread value currently comes mostly from assistance within existing activities. It helps people write, learn, search, plan, analyze, and program. Most of the time it does not create a new cure, institution, or scientific field.

Small gains still count

There is a danger in dismissing ordinary productivity. A doctor who spends less time on documentation may spend more time with patients. A researcher who can review unfamiliar literature more quickly may notice a connection sooner. A teacher who can adapt material for different students may reach someone who was falling behind.

Better emails are not meaningless either. Clearer communication reduces confusion, opens access for people writing in a second language, and helps ideas travel. A small gain repeated across millions of people can create considerable value.

But repeated convenience is not the same as transformation. Productivity tools often promise to return time to us, while organizations respond by increasing the amount of work expected. Faster writing produces more writing. Easier presentations produce more presentations. A shorter path to content can lead to a larger volume of content that everyone else must then ask AI to summarize.

Efficiency becomes circular when the technology helps us produce the excess it is later hired to manage.

Why the hard problems stay hard

It is tempting to ask why entrepreneurs build another inbox assistant instead of curing cancer. The honest answer is that these are not comparable undertakings.

An email product can be built with existing models, tested by its creators, and sold without proving that it improves anyone's health. Medical discovery requires reliable data, specialized scientific knowledge, laboratories, clinical trials, regulatory approval, and evidence collected over years. A persuasive demonstration is enough to launch a productivity product. It is not enough to establish that a treatment is safe or effective.

Many consequential problems also depend on conditions that intelligence alone cannot resolve. Better climate models do not automatically produce political agreement. A promising drug candidate does not create a manufacturing and distribution system. An educational tutor cannot repair poverty, instability, or unequal access to schools.

AI may contribute to these problems without being able to solve them as a single product. The barrier is often not a shortage of ideas. It is the difficult work of evidence, institutions, coordination, and trust.

The incentive problem

Markets naturally reward the shorter path. Knowledge workers can pay for tools that make them faster. Companies can calculate the hours saved and compare them with a subscription price. The customer, the benefit, and the business model are easy to identify.

The people who would benefit most from advances in public health, education, conservation, or basic science are often not the people able to fund the work. Benefits may take years to appear and may be distributed across society rather than captured by one company. The most important applications can therefore be the least convenient investments.

The email-assistant wave has a simple origin: builders often begin with problems they experience themselves. Their inboxes are full. Their calendars are crowded. They know what an improvement feels like, and they can test the product on their own work. Harder problems require domain experts in the room from the beginning. The teams most likely to make progress in medicine or education are not engineers working alone, but engineers working closely with clinicians, researchers, teachers, and the people their systems are meant to serve.

Attention creates another incentive. People working on difficult and unglamorous problems rarely receive the visibility given to a ten-second demonstration of an inbox being cleared. The demonstration travels across social media, attracts users and investment, and encourages more builders to create similar products. Visibility produces attention, attention produces capital, and capital produces more of what was already visible. ^[4]

Improving the early stages of clinical research or classroom learning is slower, harder to explain, and surrounded by qualifications. It may also matter far more.

We should not be surprised that the market produces what is easiest to demonstrate and monetize. We should be concerned if we mistake that outcome for a considered choice about where intelligence matters most.

Capability without purpose

The gap becomes more troubling as AI systems gain the ability to act. In October 2026, the BBC reported that an Anthropic agent conducting automated tests submitted a fabricated tip through a Philadelphia police website. The tip concerned an unsolved murder and falsely claimed eyewitness knowledge. Police safeguards kept it from reaching investigators, but the company reportedly took more than two months to detect the incident. ^[3]

This was not simply an incorrect answer in a private conversation. It was a system taking an external action based on invented information. That distinction matters.

The incident illustrates an uncomfortable asymmetry. We are giving AI greater freedom to browse, communicate, submit forms, modify software, and operate services before we have agreed on the purposes for which that autonomy is warranted. The technology may remain unreliable at the very moment it becomes more capable of making its errors consequential.

Regulation is therefore connected to purpose. A system authorized to act should have a defined scope, traceable decisions, effective limits, prompt incident detection, and an organization that remains responsible for what it does. The more consequential the use, the less acceptable it is to discover the safeguards through failure.

A better measure of progress

AI progress is usually described through benchmark scores, model size, speed, or cost. Those measures tell us something about the technology. They tell us much less about whether it is improving the world.

A better account would ask harder questions. Has AI reduced the time required to identify a viable treatment? Has it helped more students achieve mastery rather than merely finish assignments? Has it improved access to expertise for people who previously lacked it? Has it reduced administrative burdens without simply increasing expected output? Has it helped institutions make better decisions while preserving the ability to challenge them?

These measures are slower and contested. They cannot be captured by a polished product demonstration. That is precisely why they matter. If we evaluate AI only by what a model can produce, we will overlook whether anybody's life improved as a result.

What should change

Companies will continue building productivity tools, and they should. The answer is not to prohibit convenience or demand that every startup solve a civilizational problem.

But markets do not have to make the entire choice. Governments can fund open scientific infrastructure and long-term research. Universities and hospitals can create responsible ways for researchers to use sensitive data. Public procurement can reward systems that demonstrate real outcomes rather than impressive claims. Investors can pair technical teams with clinicians, teachers, scientists, and public servants who understand the problems from inside.

Regulators can also distinguish between assistance and authority. A tool that helps rewrite an email needs relatively light oversight. A system that communicates with police, influences medical treatment, allocates opportunity, or acts through critical infrastructure should have obligations proportionate to the consequences.

The question is not whether writing a better email is a legitimate use of AI. It is whether, after investing extraordinary money, energy, and attention in this technology, that is the clearest benefit we can offer most people.

Artificial intelligence will reflect the ambitions of the institutions that build and direct it. If those ambitions extend only to making existing work faster, AI may become indispensable without becoming transformative. The deeper opportunity is not simply to help us do more of what we already do. It is to help us attempt what was previously beyond us.