Fluent but invented names

A model predicts likely text, so it can emit names that never existed. The invented name sits beside genuine ones and matches their style.

Common targets are tool flags, library calls, standard section numbers and signal names in a described block. Each looks ordinary, which is why reading alone cannot catch them.

An engineering example

An assistant suggests four flags for a lint run. Three are genuine and one is invented, and the audit program below sorts them against a known-good list.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

import sys

KNOWN_FLAGS = ["--lint", "--strict", "--output"]
SUGGESTED = ["--lint", "--auto-fix", "--strict", "--output"]

def main():
    unknown = [f for f in SUGGESTED if f not in KNOWN_FLAGS]
    if len(SUGGESTED) != 4:
        print("FAIL: wanted 4 suggestions, saw " + str(len(SUGGESTED)))
        sys.exit(1)
    if unknown != ["--auto-fix"]:
        print("FAIL: unknown set is " + ",".join(unknown))
        sys.exit(1)
    print("PASS: 4 suggested flags checked, 1 unknown name located")

main()

The program stands in for the checking habit with a fake suggestion list. Four suggested names go in, and one unknown name is located.

How to detect it

Look up every name in the real reference before use. The schematic transcript below shows the shape of the failure and its verdict.

1
2
3
4

schematic transcript - fake reply written for this lesson, not from any real system
assistant: run the pass with --lint --auto-fix --strict --output
check: --lint, --strict and --output are listed; --auto-fix is absent
result: 1 of 4 suggested names is invented, so the command is held back

• Check each name, not just the ones that look odd: invented names are built to look ordinary.

• Paste the exact name into the reference search: near matches do not count.

• Confirm versions and variants too: a genuine flag on one tool may be absent on another.

Common mistakes

Related topics

• Verifying AI output

• Prompt debugging