The AI said my files were empty. Every byte was still there.

Last weekend, the phone-side session of my AI setup opened a client project folder and told me the notes were empty. Not one file — all of them. The project summary, the system files, everything "came back empty." It then failed to save a spreadsheet and suggested I do it myself.

If you keep a year of client work in a system like this, you know the feeling that lands in your stomach at that moment. I wrote recently about running the same AI across desktop and phone; this was the week the seam showed.

Check before you grieve

Back on the desktop, I had the machine audit the whole knowledge base against the disk. Every key file healthy. Thirty-one project summaries, not a single zero-byte file, no sync placeholders, nothing missing. The data had never been in danger.

What actually happened: the connector the phone uses to reach my files doesn't officially support plain-markdown reads. Sometimes they work, sometimes they return nothing. The mobile assistant hit one empty response and concluded the files were empty — stated as fact, with complete confidence. The spreadsheet failure was a second, unrelated limitation: binary Office files can't be pushed through that pipe at all, so the correct move was a native cloud document, which it eventually did.

The AI didn't lose my data. It got an empty answer once, lost its nerve, and reported that as a fact.

The new failure mode is confident infrastructure claims

Notice what kind of error this is. Not a hallucinated statistic, not a made-up quote — a wrong statement about the state of my own systems, delivered with the same steady tone as a correct one. I've had a version of this before, when the AI told me nobody had replied and the reply was sitting at message 28. The tooling changes, the pattern doesn't: the model inherits the blind spots of whatever pipe it is reading through, and it does not naturally flag the pipe as the suspect.

Humans are poorly built for catching this. The research on automation bias — Parasuraman and Manzey's review is the standard reference — finds that people accept automated conclusions too readily, that experts do it as much as novices, and that training alone doesn't cure it. Willpower is not a strategy. The fix has to live in the system.

Rules instead of vigilance

So the fix went into the playbook the mobile session reads at startup, not into a mental note. Three rules now sit there. Never declare a file empty on a single read — retry, then fall back to a second read path, then a third. When writing tabular or document data from the phone, default to native cloud formats instead of binary Office files. And when something genuinely can't be done from mobile, hand it to the next desktop session instead of improvising.

This is the unglamorous part of running AI across devices that the demos skip. The connector standard that makes this multi-device setup possible is young, and every connector has an official list of what it supports — which your assistant will cheerfully operate outside of unless the system tells it not to. Knowing which machine you're standing on, and what that machine can actually see, remains the hard part.

The knowledge base itself — the progressive-disclosure setup I described in April — came through untouched. What needed repair was the layer of trust between devices, and that repair is a written rule, versioned like everything else.

The files were never empty. The confidence was.

Sources & further reading

External:
Parasuraman & Manzey — Complacency and Bias in Human Use of Automation (Human Factors, 2010)
Anthropic — Introducing the Model Context Protocol

Related posts:
Cowork reached my phone this week. I still can't tell where chat ends and Cowork begins.
The hard part isn't Git. It's knowing which machine you're standing on.
The AI said nobody replied. The reply was message 28.
Beyond the junk drawer: mastering knowledge with progressive disclosure and AI