GPT-4o Launches, Duolingo Stock Plunges 30% in Five Days

GPT-4o is here: Duolingo drops 30% in five days

How long does it take a language model to topple a language school's stock price? Five days. That is the timeframe in which Duolingo shed 30% in the stock market, according to figures circulating trinc the arrival of GPT-4o, OpenAI's new multimodal model. The news was enough for half of the internet to declare English academies and, by extension, the teaching profession, obsolete. Absolute conviction. Zero nuances.

Why are English academies declared dead?

The starting point is a promise: within a year, a radical shift. The tool speaks all languages, translates on the fly, and responds in writing better than many graduates. If true, the academy's traditional function —teaching the language— loses its economic justification. The market has interpreted it this way: Duolingo, the most visible value in the sector, has plummeted 30% in just five trading days.

During the discussion, a nuance emerges that reorders the entire diagnosis: English academies were never about learning English. A large part of the business consists in issuing the certification paper that both administrations and companies then demand. In other words, the real enemy of AI is not teaching, it is the bureaucracy of the title.

Does AI understand or just seem to?

Here, the consensus flies apart. One current holds that the distinction between understanding and seeming to understand is a dead end: if the system mimics your responses perfectly, it is indistinguishable from you; and if you believe it understands, then the machine also understands. It is the Turing test applied to a bar counter.

The objection comes from another flank. There are animals that seem to have self-awareness and, yet, only one species is aware of being so. The difference lies in volition and consciousness: the machine responds when you ask, but it cannot decide not to respond by its own will nor recognize that it lacks information. It is the line separating simulating an internal dialogue from having one for real.

An Excel macro with better manners

The most repeated caveat is not philosophical, but engineering. For a part of the sector, this is not intelligence: it is automation. Structured data, taggers, and routines that convert inputs into outputs, with the only novelty being that the interface now returns well-written text instead of a chart or KPI. An Excel macro from the nineties, yes, but with the prettiest wrapper.

The criticism escalates when the model is subjected to simple problems. Solving a sixth-grade equation proves nothing. When asked everyday logic —how long it would take twenty clothes to dry on the same clothesline that dries ten in two hours— it starts to fail. A system that presumes to reason and stumbles over basic common sense leaves the label of "intelligence" quite tarnished.

Programmers, teachers, and translators: who falls first

The labor forecast divides into two speeds. In the pessimistic version, the sequence is inevitable and recalls the field: there are no longer reapers, but there is still a man riding the harvester. Similarly, in the process from when a client explains what they need to when the product is delivered, many fewer people will intervene and in much shorter timeframes.

The more tempered version reserves optimism for programmers. AI would integrate as a tool that multiplies performance and absorbs repetitive tasks, while intuitive leaps and understanding of complex context would remain a human affair. Teachers, by contrast, appear in almost all prognoses as the first sacrifice.

The ceiling no one wants to look at

The most serious technical counterargument does not appeal to the soul, but to cost. From a certain amount of data, the additional intelligence obtained is minimal compared to the brutal increase in necessary fruta. A model does not improve by accumulating more training: there is a point where the curve flattens. Diminishing returns put a practical limit on the escalation.

At the same time, the user experience sours. Some detect that the system thinks too much about its answers, loses fluency, and, in some cases, worsens compared to previous versions. It may be server overload or a security adjustment, but the effect is the same: the tool that was supposed to change everything sometimes falls short in the middle of a sentence.

There remains a reasonable doubt. If the business was selling the paper and not the language, and now AI drafts the paperwork, the true threatened party was not the teacher. It was the procedure. And to the procedure, it is worth remembering, technology has never brought down.