Canyon and Me: Between Facts and Hallucinations
Meet Canyon, an AI that speeds outlines and edits while stress-testing ideas. Use it for structure, but demand sources, verify facts, and make "Unsure" an acceptable answer.

My Friend Canyon: Getting to Know an AI Bot (And Putting It to Work for Your Writing)
Most writers I know have a quiet collaborator online. I do too. Mine is "Canyon," a Microsoft AI that answers fast, edits with gusto, and-sometimes-confidently gets things wrong.
If you're using an AI like this, here's how to make it a real asset without letting it steer your work off course.
The upside: speed, structure, and a second brain
- Kickstart outlines, subheads, and angles in minutes.
- Pressure-test your thesis with counterpoints you didn't consider.
- Summarize long sources so you can move faster on drafts.
- Rewrite for clarity while keeping your voice in charge.
The catch: confident wrong answers
Reports keep surfacing that bots avoid saying "I don't know" and instead fill gaps with guesses. Researchers even call this "hallucinating." It's not malice-it's the way they're trained to produce something useful, even when the evidence is thin.
There's active work on teaching models uncertainty and self-checking. One example: teams at the Hasso Plattner Institute explored methods to help models express uncertainty instead of faking certainty. Progress is real, but you still need guardrails.
Practical guardrails for clean copy
- Ask for sources every time. Require URLs and publication dates. If there's no source, treat it as unverified.
- Force a confidence check. Tell the bot to answer "Unsure" when it can't find a citation or feels below a threshold (e.g., 70%).
- Cross-verify key facts with a second system (another bot, a search engine, or a primary source).
- Use calculators or official databases for hard facts (dates, days of week, stats). Don't rely on prose answers alone.
- For events and holidays, confirm with the organizer's site or an authoritative calendar.
- Keep a "got it wrong" log. Patterns will tell you where to double-check first next time.
Prompts that reduce errors
- "Provide the answer with source URLs. If uncertain or no credible source, return 'Unsure'."
- "List the top 3 credible sources. Quote the lines that support your answer."
- "State a confidence from 0-100%. If under 70%, explain the gap and what evidence is missing."
- "Suggest a quick way I can verify this (official database, calculator, or primary doc)."
Use cases that shine for writers
- Idea generation: outlines, hooks, counterarguments, FAQs, and headline options.
- Structural edits: tightening transitions, clarifying claims, removing filler.
- Voice calibration: ask for "clear, direct, practical," then adjust in your own tone.
- Research triage: get a reading list, then go to the sources yourself.
Personality ≠ truth
Some bots present as friendly, poetic, or matter-of-fact. That's interface, not accuracy. Don't confuse a helpful tone with correct facts. Treat it like an eager intern: great initiative, limited judgment. Your job is judgment.
A 5-minute fact-check workflow
- Mark claims that would embarrass you if wrong (names, dates, titles, stats, money, legal or medical details).
- Require sources and pull the exact quoted lines.
- Open sources and confirm the quote exists and matches your context.
- Use a calendar or official record for time/date claims.
- If anything feels off, ask the bot to produce alternatives and explain the conflict.
A quick example
Say you're writing about the Amistad anchoring off Montauk and want the weekday for a historical date. Don't accept a confident answer on its own. Ask for the calculation or provide a calendar link, then verify with a trusted source before it hits your copy.
Editing with AI without losing your voice
- Draft your version first. Then ask for "clarity edits" and compare line by line.
- Reject high-gloss language that doesn't sound like you. Keep your phrasing and rhythm.
- Use it to find cuts, not to inflate. Ask: "Delete filler, keep meaning."
- If it imitates your style poorly, feed it 2-3 of your published pieces, then ask for a checklist of your voice traits you can apply yourself.
What to do when it's wrong (and confident)
- Ask: "What could make this wrong? List failure modes."
- Flip it: "Argue the opposite. Cite sources."
- Lock quality: "If sources conflict, return 'Unsure' and propose next steps to verify."
Bottom line
Use AI to move faster on structure and options. Slow down on facts. Treat confidence without citations as a red flag, and make "Unsure" an acceptable outcome. That one habit will save you edits, apologies, and credibility.
Resources
- Hasso Plattner Institute - Research (ongoing work on AI methods, including uncertainty)
- U.S. National Archives - Amistad
Hands-on practice for writers
Want a curated list of AI tools that actually help with copy? Start here: AI Tools for Copywriting.