All professions

Organisation type

Register Operations

The problem

When the tool reads a value off a scanned document, the trail often does not show what it proposed and what the person changed, and two years later a wrong field copied into other documents can no longer be traced back. Once the tool has been right many times in a row, staff stop checking its flags on identity documents. A reply to a public request goes out in an official tone but with a wrong date, name or reference to an act, and with fields that were never meant to be published.

26 courses · 1 specific to this profession

A documented failure pattern

“2?7/14” read as “237/14”, with high confidence

  1. What the tool doesAn extract of the fields of an old scanned act, prepared by an AI tool for the register officer who enters it in the electronic register.
  2. Where the tool goes wrongThe tool replaces the stained digit with a plausible one, without marking it illegible.
  3. The verification stepPut the image crop beside the extracted value and compare them digit by digit.
Source: arXiv, 2025, and 1 more.
More: why it is easy to miss and the full check

Where the tool goes wrong, and why it is easy to miss

Once entered, the act is linked to a number that may belong to another property, and the error is copied into every document later issued from the record. The tool did not mark the digit under the stain as illegible: it replaced it with a plausible digit and gave a confidence rating for the whole field, not for that digit. Nothing in the extract shows the stain, so an officer working from the extract alone has no reason to open the image; the visible digits are read correctly and the number has a regular format, so the result passes every format check.

The check that catches it

  1. For every key field (property number, cadastral zone, dates, personal numbers), put the image crop and the extracted value side by side and compare them digit by digit.
  2. Instruct the tool to write "illegible" for any character that is not clearly visible, and send back any extract in which a damaged field comes out complete.
  3. Fill in a missing character only from another document in the same file (a map, a certificate, an earlier entry) and note which one.
  4. In the audit trail, record the value the tool proposed, the final value, its source and who confirmed it.

Where the failure pattern is documented: A study accepted at NeurIPS 2025 on hallucinations when reading degraded documents (arXiv, June 2025); Anthropic's documentation on the limits of image reading (living page, read in September 2026).Sources: arXiv, 2025; Anthropic, 2026.The case above is ours, written for work in Albania; the laws it names were checked against the official text.

What it looks like

The scanned actProperty no.: 2?7/14. The second digit is under an ink stain and cannot be read.

The AI tool's extractProperty number: 237/14. Confidence: high.

Translated from the Albanian original.

Courses for this type of organisation

A short course for register staff: data integrity, document verification, public requests and the audit trail when an AI tool enters the workflow.

What you leave with

The trail specification sets what is recorded for every value read off a scan: who, when, which document, which tool, what it proposed, what the person decided and why. State changes and decisions are recorded, clicks are not, those who produce the trail cannot change it, and the sampling rule travels with its own log. The specification passes the course test when a colleague, with the trail as the only source, rebuilds the history of an entry within five minutes. You learn to use an AI tool in register work while keeping the audit trail, and to answer public requests with the publishable fields only.

Courses that apply to every profession

25 more courses are the same for every field: the foundations of AI, checking its answers, agents, and what to do when you get stuck.

Register Operations

The catalogue lists all 26 together, with filters for format and time.