The Waymo Effect: Why AI Reduces Collaboration in Science

⏱️ Lectura: 13 min

A driverless Waymo takes you from point A to point B without you having to greet anyone, and that same convenience has already reached the lab: according to an essay published on September 7, 2026, artificial intelligence is quietly emptying the human collaborator’s seat in science and software development.

📑 En este artículo

1. TL;DR
2. What happened: how the Waymo effect was born
3. Context and background
4. Technical details and performance
5. How to get started
6. Impact and analysis
7. What’s next
8. Frequently Asked Questions
   1. What is the Waymo effect?
   2. Who coined the term and where was it published?
   3. Does the essay provide quantitative data on the drop in collaboration?
   4. How can I measure whether my development team is losing human collaboration due to AI use?
   5. Does the Waymo effect mean you should stop using Claude or GPT to code?
   6. How does this relate to Borgmann’s device paradigm?
9. References

Daniel Hook, Chief Scientific Officer of the Holtzbrinck publishing group, named the phenomenon the “Waymo effect”: the tendency to prefer a language model over a real colleague because the model never brings friction, its own agenda, or unexpected objections.

TL;DR

• Daniel Hook, Chief Scientific Officer of Holtzbrinck Group, published the essay “The Waymo Effect” on September 7, 2026 in Research Agenda.
• The piece compares Waymo self-driving cars with language models: both eliminate the friction of dealing with another person.
• Hook draws on Tim Wu’s “The Tyranny of Convenience” (2018) and Albert Borgmann’s “device paradigm.”
• The central thesis: a human collaborator arrives with their own agenda and objects to things you didn’t ask about; an LLM only objects to what you ask it to.
• The essay is argumentative, not a study with its own figures: it does not provide a percentage for the drop in collaboration.
• Hook notes that the loss happens “by convenience, not by decision”: no one chooses to stop collaborating, they just choose the frictionless option.
• Applied to software teams, the same pattern can be measured with git log and GitHub’s pull request review API.

What happened: how the Waymo effect was born

On September 7, 2026, Research Agenda published an essay by Daniel Hook, Chief Scientific Officer of the Holtzbrinck group (the parent company that controls Nature, Springer, and Macmillan Learning), recounting a Waymo ride through San Francisco with Susan Winslow, CEO of Macmillan Learning. Both had crossed the country to meet in person and collaborate face to face, something they consider stubbornly hard to achieve over video call.

The irony they discovered is the seed of the article: they thoroughly enjoyed the driverless car ride, precisely because they didn’t have to talk to anyone. Two people who traveled thousands of miles to talk to each other ended up celebrating a technology whose core function is that you don’t have to talk to anyone.

From this, Hook derives an operational definition: the Waymo effect occurs when a technology eliminates the friction of dealing with another person, and that elimination is perceived as pure gain, because the cost of friction was always visible (the effort, the awkward chat) while its benefit (an outside, unsolicited point of view) was never advertised as valuable.

The essay carries that logic from transportation over to scientific research: language models like Claude or GPT are, according to Hook, quietly replacing the human collaborator in the thinking process itself, not just in mechanical writing or coding tasks.

Context and background

Hook doesn’t invent the concept from scratch: he draws on two ideas already established in the philosophy of technology. The first is Tim Wu’s “The Tyranny of Convenience”, published in 2018: the idea that once a frictionless option exists, we adopt it by default and lose, without realizing it, whatever that friction used to do for us.

The second is Albert Borgmann’s “device paradigm”: a device delivers a comfort (warmth, information, companionship) while hiding the practice that used to be required to obtain it, to the point where we stop noticing that practice has disappeared. Only the comfort keeps arriving.

Applied to a taxi driver: the cost of chatting with a stranger was obvious (the effort of being polite, the roulette of conversation). The benefit was diffuse and deferred: for many people, that driver was the last stranger with whom they had an unchosen conversation, the last reliable source of an opinion they hadn’t asked for.

Hook applies that exact structure to intellectual work. A human collaborator has limited availability, caught between classes, grant deadlines, and different time zones. A language model is available at two in the morning on a Sunday, which, let’s be honest, is when a good deal of scientific thinking actually happens.

flowchart TD
    A["Interaction with social friction"] --> B{"Does a frictionless alternative exist?"}
    B -- "No" --> C["Human collaboration is sustained"]
    B -- "Yes" --> D["The convenient alternative is adopted"]
    D --> E["The hidden benefit of friction disappears"]
    E --> F["No one decided it: it happened by convenience"]

Technical details and performance

Hook’s essay is philosophical, not a paper with its own metrics: it doesn’t include any adoption figures or collaboration-drop numbers. But the phenomenon it describes can be instrumented in a software team, which is where the Waymo effect becomes most concrete: every time a developer asks an AI assistant to solve something instead of knocking on a colleague’s door.

One direct way to measure it is to check the Co-authored-by Git trailers that both GitHub Copilot and Claude Code insert automatically when they participate in a commit. Counting how many commits carry a human co-author versus how many carry an AI co-author (or none) gives a rough approximation of a team’s real collaboration rate.

#!/usr/bin/env bash
# cuenta_coautoria.sh: measures how many commits from the last month cite
# an AI assistant as co-author, via the Co-authored-by trailer

DESDE="30 days ago"

total=$(git log --since="$DESDE" --oneline | wc -l)
con_ia=$(git log --since="$DESDE" -i -E \
  --grep="Co-authored-by: Claude" \
  --grep="Co-authored-by: Copilot" \
  --oneline | wc -l)

echo "Total commits (30 days): $total"
echo "Commits with AI co-author: $con_ia"
if [ "$total" -gt 0 ]; then
  echo "Percentage with AI: $(( con_ia * 100 / total ))%"
fi

This script runs the same on Linux, macOS, and on Windows inside Git Bash or WSL, because it only depends on git and standard shell utilities. The result doesn’t distinguish whether a commit with a human co-author was really a design conversation or just a formatting review, but it serves as an initial warning signal.

A finer-grained measurement is to ask the GitHub API how many pull requests had at least one human reviewer, not just a static analysis bot or an AI assistant marking the PR as approved.

import requests
from datetime import datetime, timedelta

REPO = "programacion-labs/backend-api"
TOKEN = "ghp_xxxxxxxxxxxxxxxxxxxx"  # use an environment variable in production
HEADERS = {"Authorization": f"Bearer {TOKEN}", "Accept": "application/vnd.github+json"}

def prs_recientes(dias=90):
    desde = datetime.utcnow() - timedelta(days=dias)
    url = f"https://api.github.com/repos/{REPO}/pulls"
    params = {"state": "closed", "per_page": 100}
    prs = []
    while url:
        resp = requests.get(url, headers=HEADERS, params=params)
        data = resp.json()
        prs.extend(p for p in data if datetime.fromisoformat(p["created_at"][:-1]) >= desde)
        url = resp.links.get("next", {}).get("url")
        params = None
    return prs

def reviso_un_humano(pr):
    reviews = requests.get(pr["url"] + "/reviews", headers=HEADERS).json()
    return any(r.get("user", {}).get("type") == "User" for r in reviews)

prs = prs_recientes()
con_humano = sum(1 for pr in prs if reviso_un_humano(pr))
print(f"PRs with human reviewer: {con_humano}/{len(prs)}")

This second script paginates the GitHub API using the standard Link header and classifies each review by the user’s type field (User versus Bot). Running it once a month, a team can chart whether the percentage of human review is rising, falling, or holding steady.

Interaction	When it’s the right fit	Advantage	Limitation
Human colleague	When the problem requires questioning your own assumption	Brings a different frame of reference and unsolicited objections	Requires a shared schedule, meetings, and waiting time
LLM (Claude, GPT)	When you need to iterate fast at any hour	Available 24/7, no social friction, responds instantly	Only objects to what you ask it to object to, brings no agenda of its own
Remote pair programming	When you need to transfer the team’s tacit knowledge	Catches context errors the author can’t see	Depends on time zones and both parties’ availability
AI bot PR review	To filter out mechanical errors before requesting human review	Fast and consistent, frees up human time for substantive work	Doesn’t replace design or architecture review

⚠️ Heads up: counting co-authorship via Git trailers doesn’t capture the design work that did happen in a hallway chat or a call before the commit; it’s an approximation, not a verdict on how much the team actually collaborated.

How to get started

If you want to instrument the Waymo effect in your own team, you don’t need an academic study: it’s enough to run the two scripts above once a month and save the result in a spreadsheet or an internal dashboard.

1. Pick a fixed time window: 30 or 90 days, always the same one, so you can compare month to month.
2. Run cuenta_coautoria.sh on every active team repository and save the percentage of commits with an AI co-author.
3. Run the Python script against the pull request API and save the percentage of PRs with at least one human reviewer.
4. Cross-check both numbers against the team’s subjective perception: a short survey asking “who did you consult about your last hard problem, a colleague or an LLM?” adds context that the git log doesn’t capture.

💡 Tip: run the git log script locally before connecting anything to the GitHub API: it needs no token or permissions, and it already gives you a first signal in seconds.

In research teams (not software ones), the equivalent is checking the real co-authorship of papers: how many authors per article, how many different institutions, and whether that figure has been declining year over year. Databases like Dimensions or Web of Science do track that data, though Hook’s essay doesn’t cite any time series of its own.

Impact and analysis

Hook’s argument about the Waymo effect isn’t that artificial intelligence is bad for science or software. It’s that convenience doesn’t announce what it’s taking with it. No one actively decides to stop collaborating with humans: each person, separately, picks the most comfortable option in the moment, and the sum of those individual decisions ends up reducing friction to zero across an entire field.

💭 Key point: the cost of social friction was always visible (time, discomfort); its benefit (an unsolicited objection, an outside point of view) was never advertised as valuable, so no one defended it when it started to disappear.

In research, this has a concrete consequence: the diversity of theoretical frameworks and unexpected objections is, historically, what keeps an entire field from being wrong in the same way at the same time. An LLM trained to be helpful tends to hand back a polished version of the question you asked it, not a different question you hadn’t thought of.

In software teams the risk is similar, though more operational: if a pull request’s human reviewer is systematically replaced by an AI assistant’s approval, the team loses the moment where someone with business context (not just syntax) asks “why are you solving this this way?” The code can end up correct and still be poorly aligned with what the business actually needed.

The counterpoint holds: no one is proposing banning LLMs or going back to email discussion forums. Hook himself makes clear he enjoyed the Waymo ride and would take it again without hesitation. The point isn’t to reject convenience, it’s to consciously notice what friction you’re letting go of before it disappears entirely.

What’s next

Hook doesn’t propose an official metric for the Waymo effect, nor does he announce that Holtzbrinck (the parent company of Springer Nature and Digital Science) will measure it across its own journal catalog. The essay ends as an invitation to reflect, not as an action plan with a deadline.

What’s predictable is that the argument will add to a debate that was already growing: how much cognitive autonomy is lost, or gained, by delegating to an AI assistant work that used to force an uncomfortable conversation with another person. That debate already has empirical evidence in the software field, though with its own metrics that aren’t directly comparable to Hook’s philosophical thesis.

For technical teams, the most reasonable practical step is the one described in the previous section: instrument, don’t assume. Measure the real proportion of commits and reviews with and without a human co-author over a few months before drawing conclusions about your own team.

📖 Summary on Telegram: View summary

Try it yourself: run the cuenta_coautoria.sh script on your main repository right now and compare the AI co-authorship percentage of the last 30 days against the previous 30.

Frequently Asked Questions

What is the Waymo effect?

It’s the name Daniel Hook gave to the phenomenon whereby a frictionless technology (a self-driving car, a language model) quietly replaces a human collaborator, and that loss is experienced as a gain because the cost of friction was visible and its benefit was not.

Who coined the term and where was it published?

It was coined by Daniel Hook, Chief Scientific Officer of the Holtzbrinck group, in an essay published on September 7, 2026 in Research Agenda.

Does the essay provide quantitative data on the drop in collaboration?

No. It’s an argumentative piece grounded in Tim Wu’s tyranny of convenience and Albert Borgmann’s device paradigm, not a study with its own dataset.

How can I measure whether my development team is losing human collaboration due to AI use?

By counting the proportion of commits with a human co-author versus an AI co-author (via Git’s Co-authored-by trailer) and the percentage of pull requests with at least one human reviewer, using the GitHub API.

Does the Waymo effect mean you should stop using Claude or GPT to code?

No. Hook makes clear he enjoyed the Waymo ride and would take it again. The point is to consciously notice what friction, and what hidden benefit, is being lost before it disappears entirely, not to reject the tool.

How does this relate to Borgmann’s device paradigm?

Borgmann describes how a device delivers a comfort while hiding the practice that used to be required to obtain it. Hook applies that same idea to collaboration: the LLM delivers the useful answer while hiding the fact that there was no longer a human conversation behind it.

References

• Research Agenda: Daniel Hook’s original essay, “The Waymo Effect,” published on September 7, 2026.
• The New York Times: Tim Wu’s essay, “The Tyranny of Convenience” (2018), cited as conceptual groundwork.
• Wikipedia: profile of Albert Borgmann, the philosopher behind the “device paradigm.”
• Wikipedia: profile of the Holtzbrinck group, parent company of Springer Nature and Macmillan Learning, where Hook serves as Chief Scientific Officer.
• GitHub Docs: official documentation on the Co-authored-by trailer used in the measurement scripts.

📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de Erik Mclean en Unsplash