A nearly 76% drop in monthly Stack Overflow questions since ChatGPT arrived in late 2022 has, until now, been told as a slow platform collapse. A July 2026 University of Auckland study argues the real damage happens upstream: the highest-reputation contributors are the ones walking away, and the next generation of AI is being trained on a forum that no longer has their input.
Dr. Kenny Ching, the Auckland Business School researcher who published the work, calls the dynamic signal compression. In Ching’s framing, when a chatbot can mimic the output of a veteran programmer, the distinguishing feature of the expert, the ability to provide a high-quality answer, is neutralised.
A 76% Drop That Hides Where the Damage Sits
Stack Overflow has seen a nearly 76% decline in monthly questions posted since ChatGPT’s late 2022 arrival. An earlier peer-reviewed estimate of Stack Overflow participation decline placed the daily traffic drop closer to 12% by late March 2023, with the 50 most popular topics on the platform each losing more than 10% of weekly question volume. Both numbers describe a site asking fewer questions; neither describes who stopped answering.
Stack Overflow’s response has been to lean into the same shift. Earlier in 2026, the company launched AI Assist, advertised as a new way for users to access its 17 years of expert knowledge, and the 2024 partnership with OpenAI was framed around strengthening the world’s most popular large language models. Taken together, those moves put the platform’s archive on the AI side of the table while the community thins on the human side. Used by a forum, that is an arbitrage. Used by a community, that is a slow handoff.
The collapse is not being blamed on AI alone. Long-time users pointed to site moderation they called self-righteous and heavy-handed duplicate-question policing, problems that predate ChatGPT’s debut. Ching’s interest lies elsewhere: not when the activity fell, but who stopped doing the answering.
- 76% drop in monthly Stack Overflow questions since ChatGPT’s late-2022 arrival
- 12% drop in average daily web visits by late March 2023
- More than 10% drop in weekly questions on the platform’s 50 most popular topics
- 23 million registered developer accounts on Stack Overflow
How Signal Compression Reframes the AI Threat
Ching calls the underlying mechanism signal compression. His working definition, from the paper: it occurs when authentic performance loses its value because AI allows low-ability individuals to mimic the output of high-ability individuals.
The premise runs through every expert community AI now touches. In a forum where reputation is built on the visible effort behind a posted answer, a tool that produces a comparable answer in two seconds erases the signal that work was meant to send. Ching’s read: they are not leaving because they cannot compete with the technology; they are leaving because their hard-earned expertise is no longer distinct from a chatbot’s answer. That sentence is the argument: the forum does not lose because AI is better at coding; it loses because AI is good enough to make the best contributions look ordinary.
The reframing shifts attention from what users ask to who keeps answering. What it adds is a question that does not show up in a traffic chart: do the people who built the canonical answers still find a reason to publish them somewhere an AI can read?
Ching’s idea does not live only on Stack Overflow. Researchers at the University of Passau and Arizona State University published a study of organizational AI knowledge loss in the Academy of Management Review, summarized in February 2026. That paper warns that long-running machine-learning systems can lose human expertise in a creeping and unnoticed manner as employees forget, defer to, or hand work over to the model. The mechanic described is the same one Ching identifies on the forum. Once the model’s output looks as polished as the human’s, the human’s incentive to keep practising quietly evaporates. Ching’s contribution is to give that dynamic a name that travels beyond any one platform.
Why Top Contributors Are Walking Away First
Ching’s study analyzed 24,304 Stack Overflow contributors over 17 months and tracked how their activity changed as generative AI tools reached wider use. The pattern that emerged is consistent across the sample. Reputations at the top of the forum fell faster than activity at the bottom, the opposite of what a simple story in which AI just replaces the experts would predict. Bots do not out-argue veterans. Veterans stop arguing because the argument no longer pays.
They are not leaving because they cannot compete with the technology; they are leaving because their hard-earned expertise is no longer distinct from a chatbot’s answer.
Dr. Kenny Ching, business school researcher at the University of Auckland, made that point in a July 2026 interview about the signal-compression finding. He is quick to add that the loss sits on the supply side of the forum, not the demand side. Consumers of Stack Overflow answers have, in many cases, already moved on. The producers are also dwindling, and Ching’s distinction, from the paper, is that the loss is concentrated on the supply side. Any forum whose reputation system runs on visible expertise runs into the same wall the moment a competent free alternative lands.
From Coding Sites to Classrooms to Company Slack
Ching is explicit that the dynamic is not specific to coding. “I argue that the same dynamic is discouraging genuine effort in classrooms, corporate workplaces, and scientific communities,” he said. Ching’s point repeats across the settings, but the mechanism holds in each one.
In university classrooms, Ching hears from teachers that students with strong academic backgrounds often use AI to finish assignments quickly, while students who genuinely love the subject use it sparingly and responsibly. Their authentic work becomes harder to grade as AI-assisted work makes identical end products.
In technology companies, the dynamic shows up around workflow design. Managers Ching spoke with told him employees known to be capable quietly step back from AI-driven workflows, while weaker performers use AI to inflate their apparent productivity. That is the corporate version of signal compression: ability hides behind an even output, and the productive people absorb the cost.
The third ring sits at the platform level. “Long-standing experts, once the backbone of these communities, have reduced their contributions sharply since generative AI tools became widely available,” Ching said. The mechanic is consistent across the three rooms. Effort becomes invisible, so the people whose effort signals the most lose the most by spending it.
The Knowledge Reset and the Sources That Remain
If experts are the input, and experts leave, then the AI training pipeline is operating on a depleting input. Coverage of Ching’s study frames the next decade’s question bluntly: future AI models will not get dumber, so to speak, but they might turn to different avenues for training, such as Slack chats, Discord servers, or even users who currently ask them the same coding-related questions they once did on Stack Overflow. Each of those alternatives loosens the link between who is talking and what is being recorded.
A Slack channel can be archived and tokenized with relative ease, but it also carries the slang, side conversations, and half-finished thoughts of working engineers rather than their canonical Stack Overflow answers. A Discord server is even more loosely bounded. A user prompt to a chatbot closes the loop by asking the model to teach itself. None of those pipelines was designed to train a public foundation model; they were designed for working teams that now share an org chart with an AI. Ching’s study does not measure the output of any of these substitutes, and it is an open question whether they will.
- Slack chat archives pulled from engineering teams
- Discord servers run by programming communities
- The coding questions users feed back into chatbots
When Effort Stops Signaling Value
Ching’s deepest concern is the supply of expertise itself. His conclusion: expertise is not innate; it is formed through the accumulation of effort, practice, and the process of engaging with difficult problems.
Continue the logic past the headline chart. If AI discourages the people who would have built that expertise from spending the effort to build it, the next round of experts will be a smaller pool. Ching adds: if AI discourages people from exerting the effort required to signal their competence, we risk truncating the very process through which future experts are created. That is the feedback loop Ching is naming: AI compresses the expert signal, the expert has less reason to keep signaling, fewer new experts form, and the next round of AI trains on a thinner archive.
What Ching’s data does establish is that, on the forum he studied, contribution is moving in the wrong direction for the wrong reason. The reason sits on the supply side of effort, not on the capability side of AI. How much human knowledge the next round of AI gets to read depends on how much effort still carries visible weight on these forums.
Frequently Asked Questions
What does “signal compression” mean in the AI debate?
Signal compression is a term coined by Dr. Kenny Ching, a researcher at the University of Auckland Business School. Ching defines it as what happens when AI lets low-skill participants mimic the output of high-skill ones, so the visible signal that used to mark real expertise is compressed away. Contributors who spent the most effort on their answers are the ones who lose the most visible reward for doing so.
Why are top Stack Overflow contributors leaving?
Ching’s study of 24,304 Stack Overflow contributors over 17 months found that the highest-reputation users reduced their activity faster than less-skilled ones once generative AI tools were widely available. The contributors told Ching they were leaving because their expertise was no longer visibly distinct from a chatbot’s answer.
What is the “knowledge reset” in AI training?
The knowledge reset is the point at which an AI’s training data runs out of new expert contributions to learn from. Coverage of Ching’s research raises the prospect for Stack Overflow-style archives, where the best contributors are walking away faster than the platforms can replace them, and notes that future models may have to source from Slack chats, Discord servers, or user prompts instead.
Will AI get worse if experts keep walking away?
Ching’s argument is that the next round of training data will carry less expert input than the last, so any future stack of model weights would be built on a thinner foundation. A separate study by researchers at the University of Passau and Arizona State University, summarized in February 2026, reached a similar conclusion for companies: “Lost human expertise can impair the quality of AI models over time, potentially in a creeping and unnoticed manner.”
Are classrooms and workplaces affected the same way?
Ching says yes. In his words, “the same dynamic is discouraging genuine effort in classrooms, corporate workplaces, and scientific communities,” because effort becomes invisible when AI makes low-effort output indistinguishable from high-effort output. He has heard the same complaint from university lecturers and from managers running AI-driven workflows in technology firms.
