Duolingo went AI-first, then took part of it back

An Authority Report is where we read what people with real operating responsibility said in public, in their own words, and work out what it means for a company that isn’t them.

This one is unusual because it has three acts: a memo, a public correction, and, a year later, an actual policy reversal. The third act is the useful one, and it is the one almost nobody covered.

April 2025: the memo

Luis von Ahn announced that Duolingo would become an AI-first company. Three provisions carried the substance:

• the company would gradually stop using contractors for work that AI can handle
• headcount would only increase if a team could not automate more of its work
• AI use would factor into hiring and performance reviews

Read purely as operating policy, this is close to what Shopify did two weeks earlier. What made the reception different was tone. The memo was clinical about the people affected, and the contractors it named were people who had worked with the company for years.

The correction

Asked about the memo later, von Ahn told the New York Times:

I did not give enough context.

He added that the company had never laid off full-time employees and did not plan to.

Both statements are accurate and neither addresses the contractors, which is what the criticism was about. Observers noted at the time that the correction adjusted the framing rather than any provision, the policies stood.

That is worth naming plainly rather than scoring a point with it, because it is the standard shape of a public walk-back: an apology about clarity, offered in place of a change. It usually works, and it worked here.

2026: the part that actually changed

The reversal came later and quietly. In 2026 von Ahn said Duolingo had dropped the rule requiring every employee to be evaluated on AI usage, because people were asking whether the company wanted them to use AI for the sake of using AI.

That question is the whole finding.

A usage metric measures usage. It cannot tell the person who used AI to do something genuinely faster from the person who routed a two-minute task through a model to produce evidence. Once the metric exists, both behaviours are rational, and only one of them is what anyone wanted.

Von Ahn puts the productivity gain at four to five times, a claim about capability made by the same person who removed the mechanism built to enforce it.

Remarks to CNBC, September 2025

Those two facts sit together comfortably: the capability was real, the enforcement mechanism was not the way to get it.

The read: mandate the outcome, never the tool

The distinction that survives all three acts is between requiring a behaviour and requiring a result.

Requiring AI usage produces measurable AI usage. It is easy to instrument, easy to report, and almost impossible to connect to anything that matters, because the tool is not the point, and everyone in the company knows it.

Requiring a result (this process gets faster, this backlog gets smaller, this response time falls) leaves the method open. Some of it will be AI. Some will be deleting a step that never needed to exist, which is frequently the larger win and which no usage metric rewards.

Duolingo’s 2026 reversal is an admission of exactly this, made by a company that was more aggressive about AI than almost anyone. That makes it a stronger data point than any cautionary tale from a laggard.

What this means at 40 people instead of 900

Do not measure tool usage. At any size, but especially at small size, where the theatre is visible to everyone and corrodes trust faster than the metric produces value.

Be specific about who is affected. The Duolingo backlash was not about the policy, it was about the people the policy quietly named. “We will stop using contractors for work AI can handle” is a sentence about specific individuals who read it, and in a smaller company they will read it in the same room as you.

Expect the correction to be about tone, then do better than that. The standard move is to apologise for clarity and keep the policy. It defuses the news cycle. It does not answer the person whose contract ended, and internally, people notice which one happened.

Take the capability claim seriously and the enforcement claim not at all. A four-to-five-fold productivity claim on specific tasks is plausible. A company-wide productivity claim is not, because the constraint moves somewhere else, usually to review, decision-making or the queue in front of the work. This is the same reason most pilots produce nothing measurable: the tool got faster and the process didn’t.

The read in one line

Duolingo proved the mandate works, for producing mandated behaviour. What it did not produce was enough value to survive its own review, and the company that ran the experiment hardest is the one that said so.

Frequently asked questions