These are six terms I didn’t follow during one week in September, and each card asks you to guess before it turns over.

No. Every count was right, and none of them measured a voice. Read aloud, the pairs sounded like one writer.

I wrote a script to score how different my essay agent's pieces sounded. It counted words, headers and dashes. It called nine of ten pairs different where I heard one writer.

What it is

Whether a measure captures the thing it claims to. That is a separate question from whether it gives the same answer twice, or whether the arithmetic is right. A number can pass both and still measure the wrong thing.

What to watch for

My own check had the same flaw. I was the ground truth, and by the fourth round I knew what the test was looking for, so my answers stopped counting as evidence. If the judge can guess the result you want, you have two stand-ins and nothing real underneath.

How I use it. Ask of any AI metric: what is this a stand-in for, and what did you check it against? Then add one case built to fail, and see whether it does.Copy

That list can’t tell you. It only shows disks the system managed to hand over. This drive was plugged in and powered the whole time.

An external drive showed nothing in the disk list. Twice, three days apart, I wrote it down as not connected. It was plugged in and powered the whole time.

What it is

A drive comes up in layers: the USB device, the storage driver, then the disk the system hands you. The tool I used only reads the last layer. The drive answered when asked what it was and timed out when asked if it was ready, so its electronics worked and its mechanism didn’t.

What to watch for

The failure shows up as silence rather than an error. The list never says a drive is missing; it prints the disks it has, and I read the gap as absence.

How I use it. When a system says something isn’t there, ask which layer is doing the reporting. An AI tool should return “nothing matched” and “you don’t have access” as two different answers, not the same empty list.Copy

Can’t tell yet. Read what the first stage retrieved: if the answer isn’t in there, the ranking never saw it.

I was reading how large video feeds decide what to show, and the design turned out to have the same shape as the search I run over my own notes.

What it is

Systems that pick a few things out of millions work in two stages. Retrieval is cheap and rough: it cuts millions down to a few thousand. Ranking is expensive and careful, and it only scores what survived. The split is a budget decision, since nobody can afford to run the careful model on everything.

What to watch for

Something that never gets retrieved never gets scored, so a miss in the first stage leaves no record at all. Teams tune the ranker because that’s the part they can see, and the fix never reaches the stage that dropped the item.

How I use it. When an answer built on search is wrong, I pull what was retrieved and read it before touching the prompt. If the right passage isn’t in there, no prompt change will fix it.Copy

Neither fact tells you. Hash both files and compare the bytes.

I was recovering photos from a failing drive. A recovered photo and a backed-up one had the same filename, and one file was about twice the size of the other. Neither fact told me whether they were the same picture.

What it is

A filename is a label someone typed. A hash is computed from the file’s bytes, and changing one byte changes it completely. Two files with the same hash are the same file, whatever they’re called.

What to watch for

Names collide, and sizes collide at volume: 68,772 files on that drive had only 50,780 distinct sizes. A hash has the opposite blind spot. It proves identical bytes, not the same photo, so a JPEG saved again looks the same to me and hashes differently.

How I use it. Ask what an integrity check actually compares. A backup that only compares filenames will pass every time, including on a file that was quietly corrupted.Copy

Still on the drive. The format rewrote the index and left the photos where they were. What’s lost is their names.

The recovery scan found thousands of photos on a drive that had been quick-formatted, a drive the computer said was empty.

What it is

A drive holds two things: an index (names, folders, dates, where each file sits) and the data itself. A quick format rewrites the index and leaves the data. Carving ignores the index and finds files by how their bytes begin. I get the photo back, but not its name or its folder.

What to watch for

Short patterns match by chance. The three bytes that open every JPEG turn up in random data about once every 16 MB, which is thousands of fake files across a drive. Matching on a longer marker that real photos carry cut the chance hits by roughly 250 times.

How I use it. Deleting usually cuts the pointer, not the data. When a system says a record is deleted, ask what overwrote it.Copy

Not any more. A later regulation moved it: 2 December 2027 for Annex III systems, 2 August 2028 for Annex I.

I was checking the dates in my notes on the EU AI Act against the official text. They didn’t match, and nothing had been mistyped.

What it is

An amending regulation is a separate law whose job is to edit another one. The original act is never reissued, so its old text keeps circulating unchanged. Regulation (EU) 2026/1744 moved the AI Act’s high-risk obligations from 2 August 2026 to 2 December 2027 for Annex III systems, and to 2 August 2028 for Annex I systems.

What to watch for

A search for the AI Act returns the original text, and the article you cite from it is authentic and out of date, because the amendment doesn’t contain the words you searched for and never comes up.

How I use it. I put a source and a checked-on date beside every date in a compliance document. An AI that answers legal questions should work from the consolidated text and show its as-of date in every answer.Copy

Checked 30 Sep 2026 against Regulation (EU) 2026/1744, Official Journal 24 July 2026, Article 1, the amendment to Article 113(c).