Privacy and product design.
Cases where the design of a product tells you what its makers actually optimized for. Includes the two-post account recovery saga that took 60,000 views to resolve.
What does a product’s design reveal about its makers?
The design of a product shows what its makers actually optimized for. Enbridge, worth roughly $110 billion, shipped a new billing site that capped passwords at 13 characters and would not take the fourth digit of an American Express CVV. HumanSlop argues that is a failure of organizational quality, not resources, and that software built by humans gets no quality pass. I quit Duolingo after a 1,000-day streak because the streak had become evidence that I was opening the app, not learning a language: the metric satisfied, the mission missed.
Can AI read your email without costing you your privacy?
Email is an open pipe anyone can write into, and it is also where we send the keys to everything else. AI can finally sort mail by what a message means rather than where it came from, but the naive design uploads your inbox to someone else’s servers and someone else’s model. Email is awful, and AI can fix it only if we protect the inbox explains why I built Premail to be bring-your-own-key, with local models through Ollama, and why its goal is less time spent on email rather than more engagement.
What should software do when it cannot do its job?
Comoji deliberately stands down while macOS Secure Input is active, because an emoji utility has no business watching you type a password. After a reboot left Secure Input stuck, held first by Chrome and then by Apple’s login process, Comoji looked broken, even to me. The fix was not to work around the protection. It was a small lock on the menu bar icon and a help article reachable from the app. Building defensive software is about that distinction: “not my bug” is not a complete answer, and software that stops should be able to tell you why.
What happens when the automated path fails?
When my son’s Microsoft account was hijacked, the attacker changed its recovery address, and Microsoft treated that address as more authoritative than the iCloud sign-in we still controlled or the Family Safety group where I am his organizer. Microsoft account recovery is security theater documents that maze. The account came back only after the post passed 60,000 views and a Microsoft employee connected me with an escalation team, which is why it should not have taken 60,000 views. Support lines fail the same way: if a customer asks for a human, give them one, and use AI to equip that person.
Building Defensive Software
Comoji stopped working after a reboot because macOS Secure Input got stuck. The fix was not a workaround: a lock icon and a help article that explain the pause.
Read→HumanSlop: How Does a $110 Billion Company Ship Software This Bad?
AI slop gets the criticism, but a $110 billion utility shipped a billing site that caps passwords at 13 characters and rejects American Express CVVs.
Read→Email Is Awful. AI Can Fix It, but Only If We Protect the Inbox.
Email desperately needs AI classification, but the inbox is too sensitive to casually hand to another cloud service. The case for privacy-first filtering.
Read→The One Place I Wouldn’t Replace Humans With AI Is Customer Care
AI can automate customer service, but it should never stand between a frustrated customer and a human being.
Read→Why I Quit Duolingo After a 1,000-Day Streak
Is Duolingo worth it? After a 1,000-day streak, my answer was no. How gems, leagues, and streak pressure turned language learning into a game of points.
Read→Microsoft Gave My Son’s Account Back. It Shouldn’t Have Taken 60,000 Views.
Microsoft restored my son’s hijacked account after an escalation, but the ordinary recovery process should not require a viral blog post to work.
Read→Microsoft’s Account Recovery Is Security Theater
Microsoft has the evidence to restore my son’s hijacked account, yet its recovery process protects the attacker. Here is why that is security theater.
Read→