Opinion

The Waymo Effect: How AI Is Quietly Eroding Research Collaboration

A 177-point Hacker News thread surfaces an underappreciated problem: AI isn't just replacing drivers, it's silently stripping the human-to-human element out of scientific research.

Font size

Source: researchagenda.news
Hacker News thread: 177 points / 127 comments

What the Waymo Effect Is

Research Agenda recently coined a term: the Waymo effect.

The analogy is straightforward. Self-driving cars didn't eliminate the job of "driver," but they removed the need to chat with one when you hail a ride. You open Uber, pick a destination, the back seat is empty, and the door opens itself.

AI's infiltration into research is producing something similar.

You used to have to ask a senior PhD student, your advisor, or a postdoc in the lab next door: how to configure the environment, why the code won't run, where the dataset has hidden problems. Now you open ChatGPT or Claude and get an answer in 30 seconds.

The author's point isn't that "AI replaced research." It's that the social glue embedded in collaboration is being quietly extracted.

This isn't an efficiency problem. It's a structural problem for the research ecosystem.

What the HN Thread Argues About: Two Narratives Behind 177 Points

The piece, published on Research Agenda, pulled in 177 points and 127 comments on Hacker News (at time of access). The comments split roughly into two camps.

One camp leans efficiency, with typical remarks like:

"I didn't see this until my sixth year of PhD. AI freed me from grunt work, and the share of my time actually spent doing research went from 20% to 60%."

They frame "not bothering other people anymore" as liberation: they used to hesitate before interrupting a senior colleague, but now there's no social cost to asking an AI.

The other camp worries about collaboration, with more specific concerns:

"The 30 minutes you save by not asking someone is quietly breaking the mentorship chain in your lab. Students no longer knock on their advisor's door because they've already asked AI. Advisors no longer check in on students because students appear to be fine."

"Code reviews have thinned out. Review used to be a one-on-one teaching opportunity. Now it's just 'LGTM, I looked at the AI-generated code.'"

The underlying disagreement isn't "is AI useful." It's whether you account for the hidden cost of research output: the tacit knowledge generated through collaboration.

Three Things AI Is Quietly Stealing

Three concrete scenarios stand out.

1. Asking questions

The old path for an environment configuration problem: ask a senior student → they're offline → wait half a day → finally connect → problem solved in 30 minutes.

The new path: paste the error into Claude → get a working fix in 30 seconds.

The time savings are real. But you've also skipped building a relationship with that senior student, missed hearing what they're working on, and lost the chance to ask the question you wouldn't have thought to ask: "How did you come up with that approach?"

2. Code changes and PR review

Old workflow: a PR waits a week for review, and the reviewer, while reading your code, mentions "that pattern is deprecated in our lab, use X instead." A week of waiting, and what you learn isn't just how to write the code, but the engineering culture of the group.

New workflow: Copilot autocompletes in real time, you submit the PR yourself, you merge it yourself. Faster, sure. But will your code carry any of the lab's idioms? No. It'll just be "correct code."

3. Debugging

Debugging used to be an office activity: "Hey, this case won't run, anyone got ideas?" Now you paste it into an LLM and figure it out alone.

The side product of collaborative debugging: a bystander might offer an angle you never considered. The side product of AI debugging: you might miss a colleague's spark of inspiration.

Add them up: question-asking, review, debugging. These three mundane research routines are exactly where advisor-student and senior-junior trust gets built. The more capable AI gets, the emptier these spaces become.

Why This Hurts More in China's AI for Science Push

China has spent the past year aggressively pushing AI for Science. Major research institutes and AI labs have launched dedicated programs targeting materials, biology, astronomy, and earth sciences.

The logic chain is: AI makes research more efficient → output increases → national competitiveness grows.

But if the Waymo effect is real, the middle link in this chain is shaky.

Research output was never just "compute × data × algorithms." It's a byproduct of collaboration networks. You cite a paper because you heard the author speak at a conference. You debug a new method because someone in the next lab casually mentioned an idea.

Under the Waymo effect, collaboration networks are shrinking. A PhD student who leans heavily on AI might become a PI in three years, but with zero experience mentoring students, because their own training was entirely self-directed.

Efficiency metrics and collaboration costs have never been tallied on the same ledger.

How to Counterbalance It: Three Concrete Steps

No motivational talk. Process changes.

1. Mandate co-located writing sessions

Once a week or every two weeks, the lab blocks off 1–2 hours with AI turned off, and two or three people sit together to write the same chunk of code or the same section of a paper.

The goal isn't efficiency. The goal is restoring physical space for the kind of casual guidance that produces real learning.

The format can borrow from pair programming: one person types, one watches, switch on a timer. But this applies beyond engineering: it works for writing and experimental design too.

2. Run periodic "no-AI" code reviews

At least for internal lab PRs, reviewers shouldn't be allowed to use AI to render a "pass" verdict. Reviewers must read the code themselves and produce at least one piece of non-obvious feedback: a convention, a historical reason, a "I fell into this pit myself once."

GitHub PR review comments are a fossil record of collaboration. AI can draft them, but it shouldn't sign off on them.

3. Count "teaching others" in evaluations

This one is hardest.

Current research evaluation runs on papers, grants, and funding. "I answered 50 technical questions for the lab next door" doesn't count as output.

But if a lab head explicitly lists "mentorship hours" in member KPIs, say N hours per month of Q&A, review, and teaching, and counts it toward promotion files, the Waymo effect has a chance of being counterbalanced.

Collaboration without KPIs is always the first thing cut in a功利 environment.

Editorial Reflection

After writing this, the more I sit with it, the more I think the Waymo effect's most underestimated damage isn't to research efficiency. It's to the growth pipeline for early-career researchers.

A new hire used to spend their first two years "learning while getting yelled at." They'd ask questions, make rookie mistakes, get corrected by seniors, and gradually absorb the field's tacit knowledge: why this dataset has noise, why this method doesn't work in our subfield, why that paper's limitations aren't actually mentioned in the paper itself. None of it lives in the literature. All of it lives in the hallway.

The more capable AI becomes, the less chance newcomers have to ask these "dumb questions." Not because AI answers well, but because asking AI requires no social courage, while asking a senior does. A shy newcomer in the AI era gets precisely protected, and precisely isolated.

My own prediction, with room for error: in the next 5–10 years, we'll see a mid-career crisis among "AI-native researchers." They'll be in their 30s and 40s, with strong publication records and solid engineering skills, but noticeably weaker lab management, cross-disciplinary collaboration, and teaching abilities compared to the previous generation. At that point universities and research institutes will discover that the PhD training cycle may need to be extended, or some form of mentorship mechanism will need to be reinvented.

What's uncertain: can that mentorship mechanism be rebuilt through process changes alone? I'm pessimistic. Process can restore the form of mentorship, but not the density. Still, the form is worth establishing first.


References

Source: https://www.researchagenda.news/articles/the-waymo-effect.html