Reading And Hearing Are Predictions Too

The sentence nobody said

A meeting transcript comes back clean and readable. Somewhere in it is a sentence that reads perfectly and was never spoken.

Not a mishearing. Not "their" for "there". A whole fabricated sentence, grammatical and plausible, sitting between two real ones.

This has been studied in clinical settings, where transcription tools are used at scale for patient notes, and researchers found invented content appearing in a meaningful fraction of segments. Some of it was invented medical detail.

If you assumed transcription was a mechanical process, this is the moment to update.

Why it happens

The same reason everything else in this library happens. A speech model is predicting the most likely text given an audio signal. It is not matching sounds to a dictionary.

So when the signal is poor, it does what a predictor does: it produces the most plausible continuation.

The worst conditions are the predictable ones:

• Silence and long pauses, where there is no signal to constrain the prediction
• Crosstalk, where two people overlap
• Accents and speech patterns under-represented in training
• Background noise, music, or a bad connection
• Technical terms and proper nouns, where a common word is more probable

That last one matters most. Names, drug names, product names, and figures are exactly what gets replaced by something more ordinary.

Where to check a transcript

You cannot re-listen to a whole meeting. You do not need to.

Check the points where something was decided or committed to:

• Who agreed to do what, by when
• Any number, date, or amount
• Names of people, companies, or products
• Anything you will forward or act on

Everything else is context. If the summary of the discussion is loosely worded, little is lost. If the action item names the wrong person or the wrong date, that propagates.

Images have a different failure

Vision models fail in an almost opposite way. They are good at the gist and bad at the detail.

Ask what a chart shows and you will usually get a correct description of the trend. Ask for the value at a specific point and reliability drops sharply.

The weak spots:

• Tables, where a value gets attributed to the wrong row or column
• Axis values read approximately rather than exactly
• Dense forms, where nearby fields blur together
• Handwriting, which is genuinely hard
• Counting anything in an image

The dangerous part is the confidence

Both failures share the property that makes everything in this library dangerous: the output looks the same when it is wrong.

A transcript with an invented sentence reads like a transcript. A table read from a screenshot arrives as a tidy set of numbers with no indication that two came from the wrong row.

The practical rule: if a number came out of an image, check it against the source before it goes anywhere. If you have the underlying file, use the file rather than a screenshot of it. That single habit removes most of this risk.