After years as a Gmail power user, moddedbear documents how explicitly created filters and whitelisted senders no longer guarantee inbox delivery — Gmail's ML classifier treats user rules as one signal among many rather than a hard override. The slow accumulation of silent failures (missed order confirmations, password resets, and human replies) made the service untrustworthy for someone who actually configures their mail.
Submitted the post-mortem to HN where it hit #1 with 830 points, signaling broad agreement among engineers that Gmail's deterministic filter behavior has degraded. The strong upvote response indicates this is a widely-felt frustration in the developer community, not an isolated complaint.
The editorial frames this as a structural ML problem: a model tuned for the population mean (users who never touch settings or run their own domain) inevitably misclassifies long-tail behavior. Properly authenticated self-hosted mail with green SPF/DKIM/DMARC still gets flagged because Gmail now weights volume and engagement patterns over authentication standards.
Moddedbear migrated to Fastmail specifically because filter rules are deterministic and the operator's incentives align with paying customers rather than ad inventory. The argument is that when you're the customer rather than the product, the provider has a direct obligation to honor your explicit configuration.
The HN thread surfaces a consistent failure mode where indie SaaS receipts, GitHub notifications from low-volume repos, and calendar invites from new contacts get silently dropped into spam or unchecked categories. This creates a chilling effect on small senders who cannot accumulate the reputation signals Gmail now requires, even with perfect SPF/DKIM/DMARC configuration.
A developer writing as moddedbear published a 3,000-word post-mortem on why they migrated off Gmail after years as a power user. The trigger wasn't pricing, privacy, or a flashy competitor launch. It was a slow accumulation of silent failures: legitimate mail — order confirmations, password resets, replies from human correspondents — quietly routed to spam or, worse, dropped into categories the user never checks. The post hit #1 on Hacker News with 830 points and the comment thread filled within hours with engineers reporting the same pattern.
The specifics matter. The author describes filters they explicitly created being overridden by Gmail's ML classifier. Whitelisting a sender in Gmail's UI no longer guarantees that sender's mail will land in the inbox — the model treats user rules as one signal among many, not as a hard override. Self-hosted mail from a properly configured domain (SPF, DKIM, DMARC all green) gets flagged anyway, because reputation in Gmail's universe is increasingly about *volume and engagement patterns*, not authentication.
The author's solution: a paid provider (Fastmail in this case, though Proton and self-hosted MX setups appear repeatedly in the HN thread) where filter rules are deterministic and the operator answers to paying customers rather than ad inventory.
This is not a story about one developer's preferences. It's a clean illustration of what happens when an ML system is tuned for a population mean and deployed against a long-tail distribution. Gmail's spam filter is optimized for the user who never touches settings, never runs their own domain, and treats the inbox as a passive feed. That user is the majority. Everyone else is collateral.
The HN thread surfaces a consistent failure mode: transactional mail from small senders — indie SaaS receipts, GitHub notifications from low-volume repos, calendar invites from new contacts — gets probabilistically silenced. Google's own Postmaster Tools dashboard shows senders their spam rate but offers no path to appeal individual classifications. There is no human review queue. There is no SLA on deliverability. For a sender, "delivered to Gmail" and "visible to the recipient" have quietly become different things.
Compare this to how the same problem is handled in adjacent systems. Stripe, for instance, exposes detailed webhooks and a dashboard showing exactly which events fired and which were retried. AWS SES gives you bounce and complaint feedback in near-real-time. Gmail is the only critical piece of infrastructure most developers depend on that ships with neither observability nor an appeals process. The asymmetry is the story.
The community reaction also reveals a generational shift. A decade ago, the consensus advice for "my mail goes to spam" was *fix your SPF/DKIM/DMARC*. That advice is now necessary but insufficient. Multiple commenters with perfectly authenticated domains, IP reputation tools, and warm-up routines report the same outcomes as the author. The variable has moved from *technical hygiene* to *behavioral reputation*, and behavioral reputation is a black box.
A secondary thread worth noting: the post lands at a moment when Google is rolling deeper LLM features into Gmail (summarization, smart reply, agent-style triage). Each of these features takes another behavioral signal off the user's plate and feeds it back into the same opaque classifier. The feedback loop tightens. The long-tail user gets quieter.
If you ship transactional email — receipts, password resets, magic links, notifications — the operational answer is to stop treating SMTP delivery as a success signal. The relevant metric is *Gmail tab placement*, which you cannot directly observe. Proxies that work in practice:
- Seed accounts across Gmail, Outlook, Yahoo, and at least one independent provider, and monitor placement daily. Tools like GlockApps and MailReach automate this; rolling your own with a few free Gmail accounts and a cron job costs an afternoon. - Instrument click-through on transactional links as a deliverability canary — a sudden drop in click rate on password resets almost always means a Gmail classification shift, not a product bug. - Authenticate every domain — SPF, DKIM, DMARC with `p=reject` — but treat this as table stakes, not a fix. Add BIMI if you have a registered trademark; it correlates with placement even though Google denies it does. - For critical flows (account recovery, payment confirmation), provide a non-email fallback. SMS, in-app notification, or a "resend via different channel" button isn't paranoia anymore; it's a SLA.
For your own inbox, the calculus is different but related. If you depend on receiving mail — for sales, recruiting, support — a Gmail address is now a liability proportional to how unusual your correspondent graph looks. Founders, indie hackers, and anyone with a contact list weighted toward strangers should assume Gmail is silently filtering 5–15% of their legitimate inbound and price the migration accordingly. Fastmail at $5/month or a self-hosted Stalwart instance with rspamd are both viable; the cost of running them is now lower than the cost of a single missed sales reply.
The broader pattern here — opaque ML systems optimizing for the median while degrading the tail — will repeat across every consumer-grade service that adds an AI layer. Search results, social feeds, app store rankings, and now mail filtering all share the same failure mode: there is no longer a deterministic contract between the user's stated preferences and the system's output. Power users either accept this, route around it with paid alternatives, or build their own. The moddedbear post is notable not because it's novel but because it's *legible* — a careful write-up of a problem most engineers can feel but few have documented. Expect more of these as LLM-driven personalization eats more infrastructure, and expect the paid-email category to keep growing at exactly the pace Gmail keeps getting smarter.
I can appreciate LLMs for some use cases, but writing emails for the user is the one that really baffles me.It's one thing if you don't speak English well and could use some help making yourself understood, but the amount of native speakers using this is so strange to me. How does this hel
A big question about this push for AI/LLM features in products, and I think very related to public sentiment on AI, is that if these features were so desirable and useful why do they need to be so forcefully promoted?If these features were so useful, the internet would be full of articles and v
It's really bizarre at this point. I'm okay with things like having one-click options for simple replies like "That time works for me" (Google Messages on Android is hilariously bad at these but it's at least useful occasionally). I'm not okay with it suggesting a whole
While I haven't had this issue with Gmail, I recently got a new computer and the first two weeks for full of moments like this. It's shocking to me how much we've let popups go rampant on everything. Perhaps the worst offender is Windows update, as it won't even let you use your
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Looking for your alternative?Let me give you some (non financially motivated) praise for Fastmail.It has everything Gmail has - even app passwords, hide my email, and ios integration. The only criticism is the calendar doesn’t autocomplete addresses so that’s a bit more typing than I would like. But