In 2014, programmers posted about 200,000 questions a month on Stack Overflow. In December 2025 they posted 3,862. In July of this year they posted 1,304. People did not stop having questions. They stopped asking them where the rest of us could see.
It’s a trend that nobody decided. No vote, no policy, no announcement that the world’s largest public archive of programming knowledge would stop being written. Only a long series of individually sensible choices. A developer with a bug now has a faster, kinder, more patient option than posting to a forum where a stranger might close the question as a duplicate. The private answer got better than the public one. So the public one stopped being produced.
Some of the decline predates chatbots. Volume had been falling since 2014, and the site’s moderation culture drove people away long before anyone typed a prompt. But researchers who compared it with its Russian and Chinese counterparts, where ChatGPT access was limited, and with math forums where the model was weaker, found a 25 percent drop within six months of release, across experience levels, with no measurable change in the quality of what still got posted.
I keep thinking about that pattern inside organizations, where the shared things are harder to count than an archive.
An organization I worked with rolled out AI tools with enthusiasm and told people to build. People did. They built small solutions that stripped process and complexity out of their own daily work, the kind of thing that used to need a scarce skill set, a budget line, and a place in an engineering queue that never came. Now it took a problem statement and an afternoon. Each build was a good decision by a capable person who understood the problem better than anyone else.
The trouble started showing up downstream, where the builders could not see it. Small differences in how each tool handled the same underlying rules produced drift, and it accumulated. Compensation was the example that got attention, because nobody wanted an unknown number of tools nudging pay decisions in slightly different directions. They had beaten this problem before, when it took the form of shadow spreadsheets. Same problem, better tooling, faster clock.
The thing that used to happen by itself
Most organizations run on a set of things everyone knows that everyone else knows too. The same numbers in the same deck. The same story about why the strategy changed. The same sense of what “good” looks like when a manager writes a review.
Almost none of that was designed. It was a byproduct of how work used to be arranged.
One office, so you overheard things. One system, so the numbers came from one place whether or not you trusted them. One version of the deck, because making a second took a week and somebody’s budget. Onboarding matched because running it twice was not worth the time. The colleague who saw the problem differently was standing at the same coffee machine whether you wanted their opinion or not.
Constraint did the work. Few people had to want the shared layer, or maintain it, or notice it. It came free with the walls.
That is why it has so few defenders now. You do not staff a role for a side effect, or put a byproduct in the operating plan. When the constraints lift, one at a time, each with a business case attached, the side effect goes with them.
What my field has been measuring for decades
I work in industrial-organizational psychology, which studies how people behave at work. We have measured the shared layer for decades, under names almost nobody outside the field uses.
The main one is team cognition: how much the people on a team hold the same picture of the work. Who is doing what, what the goal is, what counts as done, what to do when things go sideways. It sounds like a soft variable. It is not. Pooling 128 independent studies, the correlation between team cognition and team performance is about .36, which puts it in the range of effects organizations spend serious money chasing.
Socialization is the technical term for how a newcomer becomes a functioning member of a place. In 1979, John Van Maanen and Edgar Schein laid out the choice an organization makes when someone arrives. You can run people through the same experience at the same time, or let each person find their own way. Meta-analytic work across dozens of studies finds the collective version is associated with less role ambiguity, better fit, higher commitment, and fewer people quitting.
The same research names the cost. People who find their own way are more likely to change the job than inherit it. The shared version buys clarity and retention, and gives up some of the invention. That is a genuine trade, and any organization is entitled to make it.
Few of them are making it on purpose.
Microsoft ran the natural experiment by accident. When the company moved its workforce remote in 2020, researchers tracked the collaboration patterns of 61,182 employees. The network became more siloed and more static, with fewer connections bridging distant parts of the company. The bridges were not a program. They were a byproduct of a building.
Knowing that you know
One distinction explains why smart organizations keep getting surprised by this.
Information is what I have in my head. Common knowledge is what I have in my head plus my knowledge that you have it too, and your knowledge that I have it, on up. The political scientist Michael Chwe wrote a book about the difference, and his central example is the Super Bowl ad. Advertisers pay a premium for that slot not because it reaches many people, but because everyone watching knows the others are watching too. If you are selling something whose value depends on other people also buying it, reaching a hundred million people privately is worth less than reaching thirty million publicly.
AI is unusually bad at this, and not because of a flaw in the technology. A model can give ten people the same answer in ten private conversations. What it cannot do is let them watch each other receive it. Ten people can hold identical information and have no idea they hold it in common, which leaves them where they started when it comes to acting together.
Coordination runs on the second kind of knowledge, and so does the ordinary business of holding each other to a standard, which works only if we both know we were looking at the same thing. Facts are the easy half. The hard half is a shared sense of what they mean and what good looks like, and that gets built by people arguing in front of each other.
Military writing has the crispest version. Situational awareness is knowing what is happening. Shared understanding is agreeing on what it means and what to do about it. A recent piece from the Modern War Institute argues that dashboards and analytics deliver the first while getting credited with the second, and names the failure mode: tempo without coherence. Speed without reconciled interpretation widens the seams instead of closing them.
Elizabeth Ayer, writing about meetings, comes at the same thing from the other side. The output of knowledge work is judgment about uncertain things, and the standard for what counts as “true enough to act on” is set socially. Not by an individual, and not by a document. By a group of people arguing until the standard becomes visible. Which means the meeting so many of us are trying to eliminate is often the only place that standard gets made.
Four different things get called “shared context,” and they are not the same:
The standard everyone is held to.
The source everyone works from.
The moment everyone attends.
The process that catches our errors.
One of those is usually somebody’s job. The other three are rarely anyone’s.
Asking the machine instead of the person
Network analysts have a name for what happens next: social substitution. A tie that used to run between two people now runs between a person and a system.
The optimistic version is easy to state, and the CEO of the analytics firm Worklytics makes it well. Agents become new nodes in the network. They summarize your work for the people who need it, connect teams that were siloed, and answer routine questions on your behalf. Instead of asking a colleague, you consult their digital counterpart. He asks the right question himself: are agents adding connections, or becoming gatekeepers?
The early evidence says both, in different places. A survey of more than 22,000 employees in ten regions, run in June and July, found that 74 percent of regular AI users now put questions to a chatbot they would previously have put to a colleague. Fifty-nine percent ask coworkers for a second opinion less often, and 38 percent say newer people have fewer natural chances to build relationships. That is self-report about your own behavior change, which is soft evidence.
The direction matches harder data. In a two-wave study of 561 employees at technology firms, heavier AI use predicted weaker instrumental ties, the who-do-I-ask-for-help connections, while expressive ties, the who-do-I-trust ones, held up or strengthened. Both kinds of tie predicted knowledge sharing.
That split is the interesting part. The friendship survives. What thins is the working tie, the one that used to carry a question from someone who had it to someone who could answer it.
Those exchanges did more than move information. Asking someone told you what they knew and told them what you were working on. It surfaced the disagreement early, when it was cheap. It taught the junior person what a good question looks like, and the senior person what the junior person had not been told yet. And it did the small, repeated work that builds trust: you learn who is careful, who is fast, who says “I don’t know,” and they learn the same about you. A chatbot answers the question. It does not do the rest, and the rest was rarely the visible part of the exchange.
The surprise in the AI research
The intuitive story is fragmentation: personalized tools split us into a thousand private realities, each of us marinating in a custom version of the world. That story is appealing, but the strong version has not held up. When researchers switched people’s Facebook and Instagram feeds from algorithmic to chronological for three months during the 2020 election, in a study published in Science, they found no detectable effect on polarization or political attitudes. When economists tracked affective polarization across twelve wealthy countries over four decades, the United States showed the largest increase while six of the twelve grew less polarized. If the internet were the cause, it did not reach Germany. The filter-bubble argument overreached.
What the AI research shows is stranger.
Individually, these tools make people more productive and often more creative. Collectively, they make output more alike. In one experiment, writers given AI-generated story ideas produced individually better stories, while the stories across writers became more similar to each other. In a study of roughly 2,200 college admissions essays, each additional human-written essay studies, human-written essays added several times more new ideas to the pool than AI-written ones, and each additional GPT-4 essay did, so the gap widened with each essay added. human pool kept expanding while the AI pool converged. A 2025 preprint compared 27 large language models across 155 topics and found every one of them returned a narrower range of claims than an ordinary web search.
So the direction is not a thousand private realities. It is closer to one narrow shared reality, delivered to each of us alone.
Picture two competitors in different cities whose strategy teams ask a model a nearly identical question in the same week. They get nearly identical answers, build nearly identical decks, and arrive at nearly identical plans. Neither team knows this. Ten years ago they would have converged too, through the same conferences and the same three business books. At least then they would have known who else was in the room. Convergence used to come with a witness list.
We are not losing agreement. We may be agreeing more than we have in years. What thins is the layer where I know that you know, because the answer came to each of us privately.
Meanwhile the artifacts that held the common version are thinning. Wikipedia reported that human pageviews fell about 8 percent year over year, which the foundation attributed partly to generative AI summaries and to search engines answering questions that once required a visit. When Pew tracked people’s own Google searches, users who saw an AI summary clicked through to a source in 8 percent of visits, against 15 percent when no summary appeared. Roughly half as often.
Those public artifacts are also the training material, which is the part that should bother us. A 2024 paper in Nature found that models trained repeatedly on model-generated content degrade, and the tails of the distribution vanish first. The unusual answer goes before the common one does.
Coherence is not an architecture problem
Intelligence got cheap and coherence did not. The advantage now belongs to organizations that can act as one, and I think that is right.
What I notice is where the argument stops. Most treatments move straight to architecture: decision rights, governance layers, an orchestration tier, a named owner for whether the system still means what it meant six months ago. All sensible, and all the kind of thing that shows up in an operating model review.
Few of them ask where coherence came from before anyone tried to build it.
It came from the conditions. The same room, the same document, the same onboarding, the same unavoidable colleague. A decision-rights framework does not rebuild a common reference point. An orchestration layer does not reconstitute an occasion. You cannot org-chart your way back to a shared standard, because a standard is a rule plus a group of people who have argued about it enough to apply it the same way.
The architecture answer misses something else. A pair of researchers writing in Frontiers in Artificial Intelligence argue that expertise inside an organization is not only knowledge. It is a position, granted socially, that includes the right to settle a question. A system can produce an expert-grade answer without holding that position, so it cannot end the argument. Anyone who has watched two functions turn up with two flawless and incompatible AI analyses has seen it. The analysis was not the scarce thing.
Adjustable seat, fixed instrument panel
This is not an argument for standardization. The case for personalization is strong, and I believe most of it.
In 1950, an Air Force researcher named Gilbert Daniels measured 4,063 pilots and compared each of them against the average on the ten dimensions used to design cockpits. The cockpits had been built to the average pilot of 1926. Not one of the 4,063 fit the average on all ten. On any three of them, fewer than 3.5 percent qualified. The plane was designed for a person who did not exist, and pilots were dying in it.
The fix was not a better average. It was adjustability: the seat, the pedals, the straps, all built to move. Performance improved and accidents fell. It is a good design story, and it now gets used to argue for personalizing everything.
Look at what they left alone. The altimeter reads the same for every pilot. So does the artificial horizon, the fuel gauge, the tower’s instruction. The instruments stayed fixed so a pilot could be trained, checked, and trusted, and so two pilots could talk about the same flight. Adjust the seat. Leave the altimeter.
Organizations are adjusting both, and only one of them is usually a deliberate decision. The seat is the easy part and the gains are measurable: pace, format, schedule, accommodation, the path someone takes to competence. The instrument panel is where the losses accumulate without anyone deciding.
A promotion bar that flexes per person cannot be calibrated, and calibration was rarely about the bar itself. It was the room where people argued until “good” meant close to the same thing to all of them. A performance standard that adapts to whatever tool someone used cannot be evaluated. Apprenticeship works by watching someone more experienced do the work, and it does not survive people working from different scripts in different windows.
Grant the whole case for the seat. Plenty of people were served badly by the old common version, and personalization has been a genuine gift to them. Uniform standards have people they were not built for. The argument is not that the shared version was fair, but that we are dismantling the apparatus that could have told us.
The first generation that has to choose it
For most of the history of organized work, the shared layer was free. It came with the building, the schedule, the single system of record, and the cost of making a second version of anything. It was invisible because it was automatic, and automatic because we were constrained.
We are the first people who have to want it on purpose. That is a new managerial problem, and it does not look like the ones that have names and owners. It is not a communications problem or a tooling problem, and another all-hands will not fix it, since a broadcast is not a shared occasion.
The practical version is not complicated, which is different from saying it is easy. Personalize the path: the pace, the format, the schedule, the route someone takes to competence. Hold four things common on purpose: the bar people are held to, because you cannot calibrate a standard nobody shares; the source the numbers come from; one moment where the same thing is said to everyone at once, where they can see each other hearing it; and a process that catches errors, which needs a shared artifact to run on. Tailor the rest.
So the question I would put to a leadership team is smaller and more awkward than what are you personalizing.
What here is common only because nobody has gotten around to personalizing it yet? Ask it before the next efficiency gain, because the gain will be easy to measure and what it displaced will not be. Each of those decisions will be defensible on its own. That is what makes the sum of them worth watching.
Adjust the seat. Somebody still has to decide what counts as the altimeter.
References
Ayer, E. (2023, February 22). Meetings are the work. Medium. https://medium.com/@ElizAyer/meetings-are-the-work-9e429dde6aa3
Arkcoll, P. (2025, April 7). How AI agents will reshape organizational networks. Worklytics. https://www.worklytics.co/blog/how-ai-agents-will-reshape-organizational-networks
Billi, M., & Labraña, J. (2025). Is AI a functional equivalent to expertise in organizations and decision-making? Frontiers in Artificial Intelligence, 8, 1571698. https://doi.org/10.3389/frai.2025.1571698
Boxell, L., Gentzkow, M., & Shapiro, J. M. (2024). Cross-country trends in affective polarization. The Review of Economics and Statistics, 106(2), 557–565. https://doi.org/10.1162/rest_a_01160
Chwe, M. S.-Y. (2001). Rational ritual: Culture, coordination, and common knowledge. Princeton University Press. https://press.princeton.edu/books/paperback/9780691158280/rational-ritual
Daniels, G. S. (1952). The “average man”? (Technical Note WCRD 53-7). Wright Air Development Center. Account drawn from Rose, T. (2015). *The end of average*, excerpted at https://www.gse.harvard.edu/ideas/ed-magazine/15/08/beyond-average
del Rio-Chanona, M., Laurentsyeva, N., & Wachs, J. (2024). Large language models reduce public knowledge sharing on online Q&A platforms. PNAS Nexus, 3(9), pgae400. https://doi.org/10.1093/pnasnexus/pgae400
Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). https://doi.org/10.1126/sciadv.adn5290
Farnell, R. L. (2026, March 12). Shared understanding at machine speed: Preserving coherence in AI-enabled joint operations. Modern War Institute. https://mwi.westpoint.edu/shared-understanding-at-machine-speed-preserving-coherence-in-ai-enabled-joint-operations/
Guess, A. M., Malhotra, N., Pan, J., et al. (2023). How do social media feed algorithms affect attitudes and behavior in an election campaign? Science, 381(6656), 398–404. https://doi.org/10.1126/science.abp9364
Mesmer-Magnus, J., Niler, A. A., Plummer, G., Larson, L. E., & DeChurch, L. A. (2017). The cognitive underpinnings of effective teamwork: A continuation. Career Development International, 22(5), 507–519. https://doi.org/10.1108/CDI-08-2017-0140
MyIQ. (2026, August). Survey of 22,000 employees across ten regions, reported in FM Magazine: Employees increasingly consult AI, not coworkers. https://www.fm-magazine.com/news/2026/aug/employees-increasingly-consult-ai-not-coworkers/
Pew Research Center. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Saks, A. M., Uggerslev, K. L., & Fassina, N. E. (2007). Socialization tactics and newcomer adjustment: A meta-analytic review and test of a model. Journal of Vocational Behavior, 70(3), 413–446. https://doi.org/10.1016/j.jvb.2006.12.004
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y
Sourati, Z., et al. (2026). The homogenizing effect of large language models on human expression and thought. *Trends in Cognitive Sciences*. https://doi.org/10.1016/j.tics.2026.01.003
Understanding artificial intelligence usage and knowledge sharing: The social ties perspective. (2026). Two-wave survey of 561 employees at technology firms. https://pubmed.ncbi.nlm.nih.gov/42066534/
Van Maanen, J., & Schein, E. H. (1979). Toward a theory of organizational socialization. Research in Organizational Behavior, 1, 209–264.
Wikimedia Foundation. (2025, October 17). New user trends on Wikipedia. https://diff.wikimedia.org/2025/10/17/new-user-trends-on-wikipedia/
Wright, D., Masud, S., Moore, J., Yadav, S., Antoniak, M., Christensen, P. E., Park, C. Y., & Augenstein, I. (2025). What and whose knowledge? Measuring epistemic diversity and knowledge collapse in large language models [Preprint]. arXiv. https://arxiv.org/abs/2510.04226
Yang, L., Holtz, D., Jaffe, S., et al. (2022). The effects of remote work on collaboration among information workers. Nature Human Behaviour, 6, 43–54. https://doi.org/10.1038/s41562-021-01196-4
Stack Overflow question volumes: devclass, January 5, 2026 (https://www.devclass.com/ai-ml/2026/01/05/dramatic-drop-in-stack-overflow-questions-as-devs-look-elsewhere-for-help/4079575), and the Stack Exchange Data Explorer figures reported for July 2026.