Not long ago I was chatting with an engineer from another company, just sharing a bit of news about how my development has been going lately. These days I have swung all the way into AI-driven development: I write specs first and let agents implement them, I have rebuilt my review and CI habits around AI, and lately the felt productivity has changed completely. When I finished catching them up, they said quietly:

“That’s an echo chamber.”

This post is the record of taking that one line seriously. Working and building with full-on AI, and working fully remotely: I want to think through why those two, combined, make you prone to an echo chamber, as an individual rather than an organization, and as undefensively as I can.

What Is an Echo Chamber, Really?

An echo chamber is originally a room where only your own voice comes back, reverberating. As a metaphor for our information environment, it describes a closed space where the same opinions echo and dissent never enters. Ever since Cass Sunstein argued the point in “#Republic”, it is a term most often used alongside social media.

Let me get the terminology straight. The philosopher C. Thi Nguyen draws a distinction between two structures that everyday speech tends to conflate, in his Aeon essay “Why it’s as hard to escape an echo chamber as it is to flee a cult” and his paper “Echo Chambers and Epistemic Bubbles”.

• Epistemic bubble: a structure in which other relevant voices have merely been left out, often by accident. Just poor connectivity, a hole in the network.
• Echo chamber: a structure in which other voices have been actively excluded and discredited. The problem is not that information fails to arrive, but that even when it does, a manipulation of trust (“don’t believe anything those people say”) is baked in.

The decisive difference is the cure. A bubble shatters with mere exposure to outside evidence, whereas an echo chamber is reinforced by it. The arrival of outside criticism is itself pre-explained (“see, this is how they attack us”), so the more rebuttal arrives, the harder the conviction inside sets. That is why Nguyen argues escaping an echo chamber is as hard as leaving a cult. The diagnostic question becomes: is outside criticism simply not reaching me (a bubble), or have I developed a habit of discrediting it when it does (a chamber)?

Incidentally, the “filter bubble” of algorithmic personalization (Eli Pariser) sits closer to a kind of epistemic bubble in this framing, and treating it as distinct from an echo chamber keeps you from picking the wrong remedy.

The Echo Chamber You Slide Into When You Build and Work with Full-On AI

When you use AI across the board and it is producing results, lids for protecting your own conviction are available in endless supply. Meeting “AI still isn’t good enough” not by checking the data, but by thinking “this person just can’t write prompts.” Meeting a report that productivity didn’t improve not by examining the conditions, but with “that’s an unprepared setup.” Slapping the label “behind on catching up” on anyone cautious.

The insidious part is that these claims are often factually true. People who write poor prompts exist; unprepared setups exist. It is precisely because they are true that, used reflexively, they make a perfect lid for dismissing criticism without examining its content. You are manipulating the messenger’s trust rather than engaging the argument: the echo-chamber move itself.

The most dangerous footing is to rest on felt productivity. In a randomized controlled trial METR published in 2025, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” (arXiv, Simon Willison’s write-up), experienced open-source developers predicted a 24% speedup beforehand and self-reported a 20% speedup afterward, yet measured 19% slower when using AI. The scope is limited (16 developers, repositories they knew deeply, early-2025 tools), and results may differ in unfamiliar territory, as METR itself notes, but the one thing it drives home is that felt speedup and measured speedup can diverge.

The litmus test is simple. Do you seriously read research like this and apply it to yourself, or dispose of it with “doesn’t apply to me, I’m well set up”? If the latter reaction comes out, that reaction is itself the chamber signature. So I try to look at impact not by feel, but with metrics that can swing against me, like lead time and review rework count. I reconstruct criticism in its strongest form (steelman) before thinking about it, not its weakest (strawman). And I periodically ask myself: what evidence could overturn my conviction, and when did I last seriously consider it? If you can’t answer immediately, the source of your conviction is leaning toward insulation rather than evidence.

The Echo Chamber of Fully Remote Work

Fully remote work makes even better soil for an echo chamber. Group polarization, which Sunstein analyzes in “The Law of Group Polarization”, is the phenomenon where like-minded people deliberating while insulated from the outside do not converge on an average, but drift further in the direction they were already leaning.

Remote work satisfies these conditions almost perfectly. The people I talk to skew pro-AI (homogeneous), I am physically closed off (insulated), and I bathe daily in same-direction input on Slack and my timeline (repetition). In an office, a skeptic’s offhand remark in the hallway or a different opinion overheard from the next desk used to act as natural counter-pressure. Remotely, that accidental foreign matter drops to nearly zero, and the world closes around only the information I chose. Left alone, my opinion drifts quietly from a moderate starting point (“AI is useful”) toward an extreme (“developing without AI is simply irrational”).

What makes group polarization frightening is that it happens automatically as a dynamic of the environment, independent of any individual’s malice or intelligence. Even believing myself sincere and careful, merely sitting in a homogeneous, insulated place slides my opinion toward the edge. The countermeasure is to break the insulation and homogeneity yourself: keep following AI-skeptical voices instead of muting them and go back to read them regularly; treasure rather than resent the colleagues who offer caution; keep heterogeneous voices permanently in your information diet. It is implementing, as the design of your own information environment, what Sunstein calls heterogeneity as a creative force.

Conclusion

The combination of full-on AI and fully remote work is an environment that readily grows an echo chamber, where three things overlap:

1. A success-story-biased information diet (bubble)
2. The reflex to discredit critics (chamber)
3. Group polarization from insulation

Conviction itself is not the problem; the problem is whether its source is evidence or insulation.

One last note. The remark “that’s an echo chamber” can itself, depending on use, be a discrediting move that lowers the other side’s credibility. But firing back “no, you’re the echo chamber” is the most foolish move available, proving only that you have boarded the echo chamber’s logic yourself. The honest response is to receive the criticism as a falsifiable hypothesis and test it on yourself. This post is the record of that test, and the conclusion concedes more than half the point to my critic: the bubble component in me is real, the risk of chamber formation is high, and so I have to keep inspecting myself with outside voices at regular intervals.

That’s a field report on self-diagnosing the echo chamber, written as an individual all-in on AI-driven development.

References

• C. Thi Nguyen, “Echo Chambers and Epistemic Bubbles” (Episteme, 2020)
• C. Thi Nguyen, “Why it’s as hard to escape an echo chamber as it is to flee a cult” (Aeon)
• Cass Sunstein, “The Law of Group Polarization”
• Eli Pariser, “Beware online filter bubbles” (TED)
• METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”