[I've added an update after watching the announcement.]

I've been on vacation, so I haven't been following the Stack Overflow moderator strike.¹ Not that there has been much progress. Negotiations stalled for a variety of reasons. Meanwhile Stack Overflow's CEO, Prashanth Chandrasekar, dug the company's hole a bit deeper during an interview with VentureBeat.

As I mentioned the last time I analyzed an interview, I assume the author accurately represented Prashanth. I'd prefer an unedited transcript because it allows the reader to interpret in the original context but we must play the hand we are dealt. I will pull out the quotes and offer my analysis below.²

There's definitely a question around how we leverage [generative AI] technology to deliver on our mission of helping build technology through collective knowledge. This intersection between the power of community on one side and AI on the other side—from my standpoint, human-generated community content has taken us to this level, we have a large impact, but there are also so many problems we can solve by leveraging this technology.

Before I jump in, please take a moment to read Jeff Atwood's and Joel Spolsky's introductions of Stack Overflow.³ Did you notice what I did? Both describe Stack Overflow as the solution to a very specific problem: programming knowledge locked up in minds and forums whence it's difficult to retrieve. They also explain how they intend to solve the problem including using their combined audience to form the core of the community of programmers to answer questions. As someone who was there at the time, it was easy to envision how the project might solve the problem even if that was an unlikely outcome.

By contrast Prashanth regularly talks about combining community and AI without going into detail about how that solves the problem at hand. Neither does he go into much detail about the problems the company intends to solve. I suspect one reason is that Prashanth, who has spent most of his career in management, has become something of an architecture astronaut. As Joel puts it, "architecture people are solving problems that they think they can solve, not problems which are useful to solve." Since there is overlap between a Q&A site and generative artificial intelligence, there must be a way of jamming them together.

But there's another factor. In May I wrote about Stack Overflow's business, which lost $42 million over 6 months and had just laid off 10% of its employees. Since then, the company's fiscal year-end results came out. Despite growing revenue, it lost $84 million over the year ending on March 31, 2023. In fact Prosus' entire education technology segment lost money despite growing income:

On an economic-interest basis, Edtech segment revenues grew by 28% (18%) to US$545m and trading losses increased to US$258m. Growth was affected by decreased demand in the macroeconomic downturn. Our portfolio companies have reacted quickly to changing market conditions and are rationalising their cost structures and investments. At the same time, our businesses are shifting resources to take advantage of new AI technologies which promise to transform the industry. By deploying GenAI technologies, our companies can better personalise the content and user feedback on their platforms.

In their most recent earnings call, Larry Illg, the CEO of the segment and Prashanth's boss, commented:

For Stack specifically, we think that Gen AI is going to be an important evolution to how developers work and learn in the future, helping them to be more efficient and also learn new techniques while they're in the flow of work. We think, and the Stack team believes the developer community can play a crucial role in how AI evolves and accelerates, ultimately, with the community being helpful in ensuring the quality of Gen AI offerings. And Stack itself sees its role as is bringing the power of its developer community—and it's important to remember that they have assembled one of the biggest developer communities in the world—bringing that community together with the technology power of AI with the goal of creating highly trusted solutions to technology problems. And the team is working and launching a variety of AI solutions and more to come here.

So it might be that the shift to allow AI answers has more to do with Stack Overflow's parent company than its CEO. Or maybe Prashanth has been pushing this as a solution to Stack Overflow's financial problems. Either way, it appears the company hopes the community will cooperate by vetting artificially generated content. It doesn't appear they have put much thought into why the community might want to do this, though it's important to remember these are comments directed to investors, not developers.

The call happened on June 27, a month after the policy allowing AI content. I can only speculate, but a reasonable interpretation would be that the company found it could not solve its financial problems by selling data and looked for other ways it could provide something to sell to LLM developers. Given Stack Overflow's voting system, it's natural to wonder if that data could be fed back into the models. (And even better if Stack Overflow can be paid for that data.) But you can't very well do that if machine-generated content is absolutely disallowed.

Let's pop the stack back to Prashanth's interview:

We want to do it responsibly and safely and have the right use case to solve specific user and customer problems. It's going to be very, very exciting—I can't wait for the world to know and for our community to use the things that we're about to announce.

This is clearly referencing Prashanth's keynote address on July 27 when he's planning on making an announcement. I predicted the announcement would be adding Prosus' LLM technology to Stack Overflow. That wouldn't help with revenue, but it's the easiest way for Stack Overflow to demonstrate the potential value of using voting data. In order to maximize press coverage on the day of the announcement, the company can only tease without breaking the news early. That's a pretty terrible communication strategy for a community, however. Unless your community trusts you without reservation (and Stack Overflow is nowhere near that blessed state) the teases will only increase anxiety.

We're probably talking about 30 million official developers based on our own data sets, but that number is probably going to at least triple because there are so many people that are going to come into the field of writing code because it's so much easier to begin to write a baseline level of quality of code. And I think the watermark has just sort of gone up in terms of the expectations for everyone.

Stack Overflow has 21 million users, so that's not where the 30 million comes from. A quick Google search puts the number of developers worldwide at 27 million. Since the world population stands at 8 billion, roughly 3 of every thousand people on Earth are developers. I don't know if that sounds like a large or a small number to you, but given there were no programmers before 1945 and not many more added in the 1950s, the profession has grown at a goodly rate.

In a blog earlier this year, Prashanth compared AI to the introduction of tractors in that they made the job easier. I pointed out that tractors reduced the number of farmers.⁴ Tripling the number of programmers would imply a rather dramatic change to the world economy. What sort of work would they be doing? As a lapsed programmer, I do find myself reaching for the old tools to help a colleague from time to time. But I don't envision my co-workers becoming programmers no matter how easy it becomes. If I weren't around to ask, they'd reach for other tools or just go without.

I think you want to definitely be adopting these tools to get more productive and efficient and learn faster in your role.

This is another way of expressing the well-worn idea: "AI will not replace you. A person using AI will." Remember there used to be an occupation called "computer". The people who had those jobs (often women) lost them to devices also called "computers". AI enthusiasts aim for a happy medium where the technology is disruptive enough to demand attention, but tame enough that everyone who adopts it will prosper.

Since the invention of writing, there were scribes who specialized in using (and creating) the tools required to make legible marks. They were hired to take dictation and copy manuscripts. It was technical and highly-specialized work. With improvements in technology (including the printing press and mass-produced metal nibs) more people could write for themselves and "scribe" took on a metaphorical meaning for someone who makes a living cranking out the written word. Technology dramatically increased the production of writing at the cost of people making a living doing that work.

After using ChatGPT for a few months to help me write code, I'm not sure it will have the same impact as metal nibs had on writing. For common questions, it's impressive but blind spots can be easily found. The primary advantage of LLMs (their output can be copied and used without modification) also makes them poor tools for effective learning. They create cargo cults from their training materials. They sometimes produce useful results if circumstances are favorable, but provide no help otherwise.

The decision to ban ChatGPT was the “right decision” back in December, said Chandrasekar. But now, he explained, the company has slowly been working with the community, doing research and getting input.

This is the primary spot I wish VentureBeat had provided a transcript. I think Prashanth has Stack Overflow Labs in mind here. If so, the feedback from the community was largely "please don't do this". It's hard to see how any of what happened after reversing the ban on ChatGPT is working with the community. "Against" seems the more appropriate preposition.

We have gotten feedback on how best to make sure that we will trust generative AI products, from a Stack Overflow standpoint, and that's what we're going to announce in about a month's time.

This is fairly vague, but despite the context I suspect the feedback came from people working with LLMs. It certainly hasn't come from the community since their elected moderators were told their judgment was unreliable. In that light "make sure that we will trust generative AI products" feels a lot more sinister than it first appears. Perhaps what Prashanth meant was simply "improve the trustworthiness of generative AI products", which once again signals using Stack Overflow voting to provide feedback to some model or other.

[False negatives are] causing a really negative impact because when a legitimate human being coming on the website wants to ask a question, we don't want to shoo them away. We want to share the question, we want to serve our purpose and mission.

[The decision to create a new AI policy was] not an easy decision by any means—but we just thought that the cost of dissuading people from asking a question on Stack Overflow was too high. And it was happening very alarmingly high rate based on the analysis that we looked at.

From what I can tell, most suspected machine-generated content comes in the form of an answer, not a question. Nevertheless, incorrectly flagged content does have a negative impact on the humans that produced it. I believe, however, the company is underestimating the number of answers copied from ChatGPT.

We've taken a public position around this subject, saying that we always want to be open and free with our data with our developer community. However, large AI companies building models—we absolutely want to work with them in a more formal way.

While this might be the company's public position, Prashanth privately wanted to limit who can access the data. On March 28, 2023, he ordered the data dump not be uploaded to Archive.org. The DBA who turned it off warned that the community would notice and it did. Rather than having an answer prepared, the company publicly struggled for an answer. Internal communication shows most of the company was as surprised as the rest of us.

The harsh reality is that Stack Overflow's CEO hasn't listened to the people who best understand Stack Overflow. He appears to have a surface-level grasp of the site, the community and the culture. As a result, he's made decisions that harm the value of Stack Overflow. This isn't necessarily malicious, but rather incompetence of some sort.

Prashanth has already demonstrated skill in selling Stack Overflow. That was an impressive feat. But nobody can excel at everything and he hasn't shown an aptitude for running a community-centered company. It requires much more than platitudes and boundless enthusiasm.

Update July 29

I've gotten some more information related to this post that I'd like to address in no particular order:

• Apparently the interview was conducted a month or so before the article was published. You can see evidence of that by noticing Prashanth said the announcement was "in about a month's time" when the article was published two weeks before the conference. This doesn't change my analysis much. It does mean there was less feedback from the community for Prashanth to incorporate into his answers.

• After listening to the announcement, I'm struck by how none of it actually requires machine-generated content be allowed on public Stack Overflow. From what I can tell Stack Overflow is using an LLM model (probably developed by Prosus) to index content that already exists. It doesn't use ordinary post voting to improve the model. It also doesn't seem to be feeding content into the sites either.⁵ What they are actually doing seems worthy of experimenting with, so provoking the moderators, your natural advocates, to strike seems particularly counterproductive.

• There was also a podcast episode from Stack Overflow⁶ in which Ellen Brandenberger, Director of Product Innovation, and Jody Bailey, CTO, discussed the new initiatives. One revelation is that Prosus did provide several of the models including a Slack chatbot called Plus One which is being used within the company. Employees are also dogfooding search improvements (which aren't AI-related necessarily). Both guests sound genuinely interested in getting feedback from the current users of public Q&A. More of this, please!

• The secret policy which sparked the moderator strike was finally made public. It focuses extensively on the unreliability of GPT detectors. Stack Overflow moderators seem to generally agree with this conclusion because they didn't use those detectors as a rule. Predictably it points out that users from Pakistan, Bangladesh and India are substantially more likely to be suspended for GPT usage and users from the US, Sweden, Britain and Australia are less likely than average. This lines up with Stack Overflow's own data that finds 55% of Indian developers trust output of AI tools compared to 42% of all respondents. I now believe Stack Overflow leadership genuinely believed most suspensions were false positives and simply communicated their concerns in a self-destructive way.

• Someone asked at what point incompetence becomes malicious. Well, my wife is a nurse. Her job is to take care of the patient's practical medical needs. She works with doctors who are tasked with addressing the patient's medical conditions. Since my wife is a pediatric nurse, her patients are children. Some children can't swallow pills and other children refuse to take medicine in suspension (liquids). From the doctor's point of view, these are identical for many medications as long as the dosing is calculated properly. From a nurse's perspective, one works and the other doesn't.

  Thankfully many doctors have learned to defer to the nursing staff when it comes to decisions such as "pill or suspension". Some will even write their orders to give nurses some discretion for this sort of thing. But other doctors (younger ones, generally) focus on their own priorities without taking into account the nursing staff. They are probably very good doctors when it comes to diseases, but their way of operating adds friction to the care process.

  So are these doctors malicious or incompetent? Your answer to that question should be similar to your answer to whether a CEO of a community-centered business adds friction to the operations of his company because he's focused on his own specialty. Given ongoing pain suffered by people in their wake, I can see how you might call this behavior malicious. It's certainly incompetent, though not in the primary duty of a doctor. So maybe this is the wrong framing?

  The core of the problem, in my estimation, is that doctors and CEOs aren't held accountable for the problems they sometimes cause in their organizations. This is also true of half a dozen other professions such as politicians, lawyers and pastors. People, even people who theoretically oversee these folks, just defer to the them by default. Unfortunately, these professions also attract certain personalities who project confidence that reinforces the perception that they don't need oversight.

  Practically I prefer to think of the problem as incompetence because that implies that experience could change the CEO's/doctor's behavior. But it might be better to attribute it to malice since the only reliable solution I've observed is for the person to move onto some other company.

• My blog has a comment section. You can see the comments on this post below or you can go directly to my Meta site where you can find all the comments.

---

1. Not to be confused with the Writers and Actors Guilds strike. There is some crossover however. As Brian Merchant wrote in the article I just linked to:

   The formula for seeking out that next multibillion-dollar "unicorn," in hindsight, was pretty simple: The next wave of start-ups had to promise that it would disrupt a stale industry with a newer, high-tech app-driven alternative, promise the potential for vast scale and promise that it could do so fast.

   Oddly the "stale industry" Stack Overflow intends to disrupt seems to be its own. ↩

2. This is a traditional fisking, though I try not to be nasty about it. ↩

3. It's dangerous to send people to really interesting articles since they might never come back. Still, you could do worse than just read what prompted Stack Overflow's founding and compare them to the current CEO's ideas. ↩

4. In a relative sense at least. Agricultural technology is a bit odd since it allowed for dramatically higher global population and thereby increased the number of farmers. AI seems unlikely to have the same effect. ↩

5. Though it's possible I didn't listen closely enough. ↩

6. I used to be a regular listener to the podcast. I stopped because I found I don't care about the topics which were rarely related to Stack Overflow (much less Stack Exchange). Listening to this episode was refreshing. Not sure if it's a regular feature of the show now, but I was pleasantly surpised with a discussion of a recent Lifeboat badge award. ↩