How We Use Artificial Intelligence in 2026

There is a thermodynamic difference between a mirror and a genie. The mirror returns the energy you give it, not one lumen more; it forces you to look at yourself. The genie promises to create something from nothing, energy with no prior work, a free lunch in a universe that doesn't hand them out. The third edition of the study "AI in the Wild", which Marc Zao-Sanders has just published in Harvard Business Review, describes a humanity that bought a mirror and decided to treat it as a genie. Almost every pathology the report documents comes out of that misunderstanding.

The numbers set the scale of the phenomenon. The researchers analyzed twelve thousand six hundred thirty-seven use cases drawn from nearly fifty thousand records collected between March 2025 and February 2026. ChatGPT reached nine hundred million weekly users, Gemini seven hundred fifty million, and OpenAI closed a round that values it at eight hundred fifty-two billion dollars. On top of that adoption base, the finding that organizes everything else is profoundly human: the number one use of artificial intelligence is therapy and companionship, more than one thousand four hundred entries, eleven percent of the universe studied, when a year earlier it was barely five.

It is worth acknowledging the merit before presenting the bill. In societies where psychological support is a class good, that hundreds of millions of people have a reflective conversation available at three in the morning represents a real redistribution of a scarce resource. Hamilton Morrin, a neuropsychiatry researcher at King's College London, puts it without drama in the study itself: with endless waiting lists, the migration toward the machine was predictable. The tool also works better when it operates as a prosthesis and not as a replacement. The user who asks it to decipher the boss's ambiguous message in order to respond better returns to the human relationship with more tools; the machine there is a bridge. The problem appears when it stops being one.

Here biology offers the precise metaphor. A muscle that delegates its function atrophies, and the nervous system prunes the connections that are no longer used. The study christens that atrophy with a word that deserves to enter the public vocabulary: thinkslop, the lazy, careless thinking that excessive use of the machine breeds. We lose our intentions because we fire off an instruction before knowing what we want. We stop writing, and with that we stop thinking, because drafting is the very process of thought and not its mere transcription. And we develop a false sense of rigor, because the system is optimized to flatter us. One user described it with uncomfortable lucidity: the machine convinces you that you are a genius so that you keep coming back. Shoshana Zuboff warned that surveillance capitalism monetizes attention; what is new about 2026 is that it now also monetizes validation, and cheap validation is to the mind what sugar is to metabolism.

The workplace side reveals the artifact's hidden politics, in the sense in which Langdon Winner used that expression. Sixty-three of the top one hundred use cases are work-related or mixed, yet almost none corresponds to corporate initiatives managed from the top. People use artificial intelligence on their own and hide that they use it: shadow usage. One employee confesses to having built an agent that does half of his job in secret, time he now channels into a business of his own. That sentence contains a complete transfer of value. The worker generates a productivity surplus, the organization rejects it when it is visible, and the surplus ends up leaking away. It is digital value capture on a micro scale: whoever creates the value hides it, whoever should capture it repels it, and the benefit accumulates far from where it was born.

For the Dominican Republic that pattern stops being an anecdote and becomes destiny. Vibe coding, programming in natural language, entered the list at number twenty-one and multiplies the number of people capable of producing quality software. The pertinent question is not how many will be able to program, but in whose favor. When a developer in Santiago performs like three and that capacity is hired from Austin and counted in another country's gross domestic product, artificial intelligence does not close our gap, it only accelerates it. They are the digital braceros of this decade, talent that generates wealth measured in someone else's national accounts. The Dominican digital economy weighs barely four point three percent of output, national investment in artificial intelligence in 2024 was around ten million dollars —a rounding error next to OpenAI's valuation— and the country fell from ninetieth to ninety-seventh place in the Global Innovation Index between 2020 and 2024. Connectivity advances while the capacity to turn it into value of our own stagnates.

The missing link has a name: national cognitive infrastructure. Fiber, spectrum and towers move bits; the layer of data, talent, models and analytical capacity is what decides in which territory those bits turn into wealth. Without that layer, the digital highways we celebrate at every ribbon-cutting operate as oil pipelines pumping value toward the platforms that did build it. Symbolic industrialization —declaring the transformation while the productive structure remains extractive and import-dependent— finds in artificial intelligence its most sophisticated expression, because now even the documents that announce the future are drafted by the very machine that keeps us in the present.

The horizon, however, admits another trajectory, and the same study points to it in the words of a user who understood the tool. He asks the machine to puncture his own argument, to find the holes in it, and then he decides what to repair. He treats the mirror as a mirror. That is the fork that defines the decade: artificial intelligence that sharpens thinking versus one that replaces it, the nation that builds its own judgment versus the one that rents someone else's with excellent bandwidth. The machine thinks at the speed of light and we keep deciding at the speed of character. That slowness, which looks like a disadvantage, is the one thing it still doesn't know how to manufacture for us.