Duolingo’s AI Output Surged, but Its CEO’s Jobs Warning Needs a Caveat

Duolingo’s use of artificial intelligence has moved well beyond the experiment Luis von Ahn discussed in 2024. The company now documents a sharp increase in course production alongside continued employee growth, reinforcing his warning that automation can remove particular assignments without eliminating whole professions.

What remains unresolved is who bears that disruption. New international evidence supports von Ahn’s concern that workers with less protection may suffer first, but it complicates his emphasis on poorer countries: wealthier economies have more exposed jobs, while developing economies may lose valuable digital work before they gain the technology’s productivity benefits.

The 2024 warning followed a real reduction in contract work

Forbes’ September 24, 2024 interview documented that Duolingo had declined to renew contracts for about 10% of its contracted workforce in late 2023, affecting translation and lesson-writing assignments that AI could perform in some cases; it also placed von Ahn’s comments about poorer and less-educated workers at Duo’s Taqueria in Pittsburgh, where he ate tacos and drank a margarita.

The restaurant setting supplied the striking contrast, but the economically important detail was the type of work removed. These were contractor assignments rather than layoffs of an equivalent share of full-time employees. The distinction does not make the lost income immaterial, but it limits what the episode can prove about AI replacing complete jobs.

Translation and lesson production consist of multiple tasks: drafting alternatives, checking meaning, judging cultural context, organizing content and reviewing errors. Automating some repetitive components can reduce the number of paid hours available without making every language professional, educator or editor redundant. The Duolingo case is evidence of task substitution, not evidence that an entire occupation disappeared.

Duolingo now has evidence of scale

Duolingo’s current strategy overview states that AI and automation support course construction and Video Call with Lily, and records 20,500 language-course skills published in the first quarter of 2026, compared with quarterly averages of 7,100 in 2025 and 1,800 in 2024; the same page lists more than 850 employees, annual retention above 90% and an expectation that the team will grow.

Those figures establish that Duolingo’s content operation has changed substantially, but they do not independently measure teaching quality or isolate how much of the production increase came from generative AI. They also provide no contractor headcount, making it impossible to determine from the public figures how many assignments were removed, redesigned or transferred to different workers.

The combination nevertheless matters for creators and other project-based specialists. A technology company can expand its permanent workforce while buying less routine writing, translation or content-production labor from contractors. Headcount growth and displacement can therefore happen at the same company at the same time because they affect different employment relationships and different bundles of tasks.

This is also why broad claims that Duolingo simply replaced people with AI are misleading. The available evidence instead shows a reallocation: automated systems increase the volume of certain outputs, while employees continue to handle product development, engineering, design, review and organizational decisions. Stable or rising employee numbers do not cancel the loss of contract opportunities, just as lost contracts do not demonstrate the disappearance of human work across the business.

Poorer countries face a different risk, not necessarily greater exposure

An ILO–World Bank working paper published on March 17, 2026 examined 135 countries covering roughly two-thirds of global employment and concluded that high-income economies have greater overall exposure to generative AI, while developing economies may experience disruption before productivity gains because connectivity and job content distribute the risks and benefits differently.

This complicates a simple prediction that poor countries will be hit first or hardest. Their economies contain fewer computer-intensive professional and clerical roles, which reduces aggregate exposure. Yet the exposed office jobs they do have can be relatively desirable routes into formal employment, particularly for younger workers and women.

Workers in automatable digital roles are often already connected to the systems through which AI is deployed. Other workers whose performance might improve with AI may lack dependable internet access, suitable software, training or an employer capable of reorganizing work around the tools. A country can consequently have fewer exposed jobs overall while being less prepared to protect the people who lose them or distribute the resulting gains.

Education also does not map neatly onto safety. Many highly exposed clerical and professional roles require more formal education than agricultural, transport or manual work. Von Ahn’s warning is strongest when interpreted as a claim about bargaining power and institutional protection, rather than as a universal rule that lower educational attainment always produces greater technical exposure.

Exposure is not the same as a lost job

Three separate stages are often collapsed into one prediction. A task may be technically exposed to AI; an employer must then decide that the system is reliable and economical enough to adopt; finally, the organization must determine whether to remove the task, redesign a role or ask the same workers to produce more. Each stage can lead to a different employment outcome.

For a contractor paid to deliver a narrow, repeatable output, task automation can quickly become lost income. A full-time employee whose role combines judgment, coordination, accountability and several changing responsibilities is harder to replace as a unit, although AI may still alter the role or reduce future hiring. That difference helps explain why contractor demand and employee headcount can move in opposite directions.

Von Ahn’s 2024 remarks therefore remain relevant, but not because they forecast a uniform wave of mass unemployment. Duolingo’s subsequent production figures demonstrate how quickly a company can scale work that lends itself to automation, while the newer labor evidence shows that exposure, displacement and access to productivity gains are distributed through different mechanisms. The central conflict is no longer whether AI can perform useful work; it is whether the workers whose assignments become unnecessary have realistic paths into the new work created around it.

Also read:

• Uber CEO Dara Khosrowshahi Just Said What Every Other Executive Is Too Scared to Admit: AI Will Replace 70-80% of Human Work
• Hasbro CEO Says AI Will Become Core Part of Dungeons & Dragons
• Authors Suing OpenAI Will Get to See Its Secret Training Data in Heavily Locked Down Room