AI fluency practical assessment: what it measures

What is being measured

The assessment observes how someone delegates work to AI, checks its output, decides whether to release it, explains their decisions and hands work to another person. It assesses the submitted actions and supporting evidence, not how frequently someone uses an AI tool. This English edition summarizes the public Korean reference reviewed on September 17, 2026; the full fixed rubrics and policy details remain linked below.

Five dimensions and their evidence

The conceptual reference is the 4D AI Fluency Framework developed by Rick Dakan, Joseph Feller and Anthropic. Preterview's five dimensions and rubrics are its own operational definitions; they are not an official implementation or endorsement of that framework. See the framework.

The assessed material is typed text and AI work recorded inside the assessment. Earlier sessions before September 7, 2026 used additional steps; changing the flow does not silently rescore those earlier reports.

What is not measured

• Facial expressions, voice, personality or attitude; this practical assessment does not collect video or audio.
• General job knowledge as a standalone examination. Flagging an embedded factual issue as needing verification can count.
• Rank or percentile among other candidates.
• The quality of the AI output by itself, rather than the candidate's actions.

How scoring works

Tasks come from a reviewed bank with predefined defects. The server first matches findings against known anchors and uses model classification for ambiguous cases. A candidate's findings ledger is sealed after the review step; later answers do not simply add findings to it.

Scoring and narrative feedback are separate. Fixed server rules can cap dimension scores, and the model cannot change those caps. Verification uses a server calculation; the model supplies its explanation and evidence. Quoted evidence must match a continuous passage in the candidate's actual text. Unsupported quotes are removed; dimensions without required evidence may be marked unscored under the scoring rules.

The overall score is a weighted average on a 0–100 scale. All five dimensions must be scored before an overall score is issued. A report with missing dimensions must not be compared on the same overall scale with a complete report.

Interpreting the behavior scale

The underlying rubric has levels 0–4. Reports display the corresponding values 0, 2.5, 5, 7.5 and 10. The reference mark of 7.5 points to rubric level 3; it is not a hiring cutoff and does not trigger an automatic selection decision.

At the stronger end, delegation specifies the purpose, context, constraints and output, and explains revised instructions. Verification identifies concrete checks and problems in revised AI output. Judgment states release criteria, sensitive-data boundaries and responsibility. Defending decisions explains weaknesses and failure conditions. Collaboration identifies the AI contribution and includes a colleague's review in the handoff.

The complete, fixed level-by-level rubrics are in the Korean reference. Score caps are applied separately, so a capped score can be lower than the uncapped behavior level; the report shows the applicable reason. Honest uncertainty with a concrete next verification step is distinguished from unsupported certainty.

How to use a report

Compare only results for the same task ID. Different task forms have not been statistically equated. There is no validated human score-band distribution in the reviewed reference, and simulated score distributions must not substitute for one.

Read the five dimension scores, behavior descriptions, evidence and factual flags before the overall score. Use the report to prepare follow-up interview questions or discuss training needs. Do not use this document alone to accept or reject a candidate, rank different tasks, or compare experience levels as if the scores were adjusted for seniority.

AI produces the score; an individual human review is not automatically included. Hiring and personnel decisions belong to the institution using the tool. Preterview provides the scoring tool, criteria, quoted evidence, version history and a rescore route. The institution handles its notices, lawful basis, decisions and its own appeal procedure.

Reproducibility and limitations

Repeated model scoring can vary. The reproducibility reference publishes dated measurements and the targets they do or do not meet. A single session is a single task, not a complete judgment of a person's ability. Paste and window-switch counts are client-reported reference signals and do not affect the score.

The Korean reference cites external research about observed AI-use behaviors. Such research is not a norm for Preterview candidates and must not be used to interpret an individual's score. Translation of this document does not expand the availability of institutional assessments to new languages.

Rescoring

Requests use the contact link below the report. The default request window is 30 days from issue, subject to the notice for that assessment and its shorter remaining retention period. Rescoring is impossible after the source material is deleted. The operations team reviews requests; requesting a rescore does not guarantee an increase.

The sealed findings record is not manually edited. If extraction is reviewed, it uses the text from before the ledger was sealed. The same scoring function applies: matching samples are accepted; a one-level disagreement uses the lower value; larger disagreements require a third sample and use the median. Rescoring creates a new report version and fingerprint while retaining the prior version. Scores can go up or down.

Access, retention and versions

Personal practice results are available to the participant; educational assessments also allow the institution's designated viewer. Employer rounds allow candidate and employer access under the assessment's notice, with employer access after the response window closes. The employer profile requires the stated notices and consent steps.

For employer rounds, the reference states a default retention period of 90 days after the response window closes, followed by deletion of source material and detailed feedback while a summary remains. Downloaded copies are the institution's responsibility under the same retention terms. Non-accepting invitees' name and email are removed when the response window ends. Confirm the specific notice for the assessment you are taking.

Reports distinguish scorer, task-bank, prompt and result versions. A rescore updates the result version and fingerprint. When methods change, measurements must be read against the corresponding scorer version rather than borrowed from an older one.

Open the complete Korean reference, rubrics and policy tables. This explanation does not replace the service terms or the notice shown for a particular assessment.