The AI graveyard: the tools ChatGPT didn’t kill alone

“ChatGPT killed Stack Overflow” is the version of this story everyone tells. It is tidy, it has a villain, and it is wrong in a specific and instructive way. Stack Overflow’s monthly question volume did collapse, from roughly 182,000 at its 2016 peak to a few hundred by early 2026, but it had plateaued and begun slipping years before ChatGPT existed. The platform was already shedding questions from around 2017, thanks to its own aggressive moderation and a reputation for closing newcomers’ posts.

So what did ChatGPT actually do? It didn’t start the fire; it poured petrol on one already burning. The useful distinction here is between cause and accelerant, and the popular framing collapses the two. Disaggregate the five platforms on this chart and they sort into different stories: Stack Overflow was dying and AI sped it up; Chegg was a healthy business that AI substituted for almost overnight; Google Translate is losing a slice of casual use to chatbots while remaining enormous. One headline, five mechanisms and only one of these lines is built from counted events rather than visit estimates

Stack Overflow’s line indicts the platform, not just the chatbot.

The instinct is to date the collapse to November 2022. The measured data doesn’t cooperate. Questions plateaued across 2014–2016, then drifted down from 2017 — and even rose briefly during the 2020 lockdown – well before ChatGPT shipped. The decline is the product of years of duplicate-closing and downvote culture that drove beginners away. What AI supplied was a patient alternative that never tells you your question is a duplicate. The mechanism that matters isn’t “AI is better at coding answers”; it’s “AI removed the social cost of asking,” and Stack Overflow had spent a decade raising that cost.

Chegg is the cleanest kill, precisely because it was healthy.

Unlike Stack Overflow, Chegg wasn’t in decline. It peaked around 2021 as a profitable subscription business with a moat of 130 million human-written homework solutions. That moat evaporated the moment a free chatbot could solve the same problem instantly and explain it. The lesson the “AI disruption” headlines miss: Chegg’s vulnerability wasn’t weakness, it was that its entire value proposition, paying to access answers, became the thing AI gives away. Strength in the old model was no protection.

“Decline” means something different for Google Translate than for the rest.

Putting Google Translate on the same axis as Chegg invites a false equivalence. Translate’s indexed line falls, but the underlying product still serves enormous volume; what it’s losing is the casual, conversational translation that now happens inside a chatbot window. Read the index as “share of its own former relevance,” not “headed for the grave.” The same downward slope hides a terminal collapse in one case and a flesh wound in another. Which is exactly why indexing to peak is honest about shape and silent about scale.

The graveyard’s quiet warning is about what feeds the replacement.

Here’s the mechanism almost no coverage names: ChatGPT answers coding questions well because it trained on fifteen years of Stack Overflow answers. As the forums that generated that knowledge fall silent, the supply of fresh human-verified answers to new problems thins. The tools that killed the libraries were built from the libraries. Whether that’s a one-off harvest or a sustainable loop is the open question; and it’s the variable the celebratory “forums are obsolete” framing leaves out entirely.

Methodology note

Only Stack Overflow’s line is built from counted events; the other four are visit estimates, and the chart draws that distinction deliberately — Stack Overflow renders as a solid line, the rest as dashed.

Each line is indexed to that platform’s own peak year (=100), so the chart compares the shape of decline rather than absolute size. Stack Overflow uses measured monthly question counts from the Stack Exchange Data Explorer, 2014–2024, indexed to the 2016 peak-year average (~182,000 questions per month); the 2025–26 tail is anchored to reported near-2009 volumes, as the public query data ends in March 2025. The other four use annual website-visit estimates drawn from SimilarWeb-class data reported across 2025–26, plus company disclosures for Chegg, with years between sourced anchor points interpolated.

Peak years: Stack Overflow 2016, Quora ~2019, Chegg and Google Translate ~2021, Dictionary.com ~2020. The indexed display masks that these platforms differ in absolute scale by orders of magnitude — and a falling visit count is a softer signal than a falling count of questions actually asked, so the dashed lines should be read as indicative of trend, not precise annual readings.