Academy

AI and data · Beginner · About 35 minutes, plus practice

Verify an AI Draft Before You Use It

Check a short AI-assisted planning draft claim by claim. Build an evidence table, correct unsupported statements, and preserve unknowns so a teammate can understand which parts were checked and which still need work.

What you will make

A corrected draft and a claim table showing source support, corrections, and unresolved questions.

A fluent draft can make unsupported dates, numbers, or citations appear settled. Verification gives you a usable editorial record. It also helps you recognize when a task needs better evidence or a person's judgment instead of another generated answer.

  • Separate factual claims, inferences, and suggestions in a draft.
  • Verify citations and inspect calculation inputs and units.
  • Correct or remove unsupported claims while recording unresolved evidence.
Before you start
  • A short draft and the public documents or structured data authorized for the exercise.
  • A spreadsheet or text editor. Access to an approved AI interface is optional; the checks can be completed manually.

Project connections: Bluefield VA comprehensive plan; Haysi downtown; Clintwood downtown reuse and housing; Red Onion industrial website and marketing; Clinchco and Nora workforce and former-school reuse.

Work through the method

  1. Inventory the claims

    Identify checkable dates, counts, names, units, and status claims in a small draft. Assign IDs and define the authorized public source set.

    Check your work: The inventory covers the selected draft.

  2. Locate source evidence

    Open each public source and locate its passage or table. Check title, date, organization, geography, and page. A working link alone does not support a claim.

    Check your work: Supporting evidence has a precise location.

  3. Check the wording

    Compare each statement with source context, qualifications, and exceptions. Mark supported, contradicted, or unresolved. A second model answer does not replace reading the evidence.

    Check your work: Inferences are distinguished from source statements.

  4. Audit any arithmetic

    Check original structured inputs, units, exclusions, and denominators. Recalculate a small example in a spreadsheet. Review generated code and methods where relevant; keep missing distinct from zero.

    Check your work: Checked arithmetic is reproducible from documented inputs.

  5. Make the correction

    Correct supported statements and remove or flag unresolved ones. Keep proposals labeled. Identify questions requiring human review; do not automate grading, eligibility, or professional judgment.

    Check your work: The corrected draft keeps uncertainty visible.

  6. Save the evidence trail

    Save the draft and evidence table with review date, scope, calculation checks, and AI-assistance disclosure. Trace one corrected sentence to its evidence.

    Check your work: Another reader can repeat the selected checks.

Prompts to adapt

Replace the bracketed details with your own question and authorized source material. Check the output rather than treating it as evidence.

A claim-by-claim draft review

Using only [authorized public draft] and [named source documents or data], list checkable claims with IDs. For each, provide the draft wording, exact source page/section or table location, supporting evidence, and a status of supported, contradicted, or unresolved. Keep inference and proposal separate from source statements. Identify missing inputs, units, and denominators for numeric claims. Do not invent citations or certify professional conclusions. Draft corrections only where the provided evidence supports them.

  • Open each evidence location independently.
  • Recalculate numeric examples from the original inputs.
  • Check that unresolved claims remain visible in the corrected text.
  • Do not use this output for automated judgment of people or projects.

Audit a short planning draft

  1. Select a short draft based on authorized public evidence and identify its claims.
  2. Check every selected claim and source location; label unresolved evidence.
  3. Recalculate any small numeric example with explicit units and inputs.
  4. Save the corrected draft with the evidence table and review scope.

Bring the method into a product

  • Corrected draft with unresolved claims visibly marked
  • CSV or text claim-verification table

Suggested formats: PDF, TXT, CSV.

Reflect: Which unresolved claim would change the reader's decision most if it turned out to be wrong?

Avoid these mistakes

  • A valid URL can still be irrelevant to the sentence citing it.
  • Missing data and a measured zero have different meanings.
  • A model's agreement or confidence is not independent source verification.

Check your understanding

These short checks are practice, not course grades. Read the explanation, revise your reasoning and try again when useful.

Which check establishes citation support?
What should you inspect in a generated calculation?
How should an unsupported claim appear in the reviewed output?

Submit your practice product

Use the existing course login. Your file and notes are visible to you and the instructor, not the public or other students. Official graded assignments still go to Canvas.

Submit a PDF, image, DOCX, UTF-8 text/CSV, JSON or GeoJSON file. For several products, combine the evidence, text alternatives and notes into one accessible PDF or DOCX, or submit each file separately with the same project title. Keep native ArcGIS projects and complete shapefile bundles locally. Upload only work you can share with the instructor; leave out confidential client records and sensitive species or infrastructure locations.

My academy work

Sources and further help

Guidance may change. Follow the current source and check the software version, license, local rules and data limitations before applying the method.