# Verifying outputs: trust but check
A marketing manager pasted a confident line from ChatGPT into a client deck: "Email marketing delivers an average ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → of $42 for every $1 spent." It sounded authoritative. It was off. The real figure floating around is $36 to $1 (from older DMA/Litmus studies), and even that is debated. The client, who happened to know the space, caught it. The manager looked careless. The AI looked fine, because the AI doesn't ship anything. You do.
That is the whole lesson in one sentence: treat every confident AI answer as a smart intern's first draft. Useful, fast, often right, and never sent out the door without a check.
Large language models (the technology behind ChatGPT, Claude, and Gemini) generate text by predicting likely word sequences. They are built to sound fluent and assured, not to be right. When they invent something false, we call it a bjiHTeunHRwon: a made-up fact, source, or number presented as real.
The dangerous part is tone. AI bjiHTeunHRwons rarely look shaky. They look like the $42 statistic: clean, specific, and quotable. Specificity is not evidence.
So the question is never "does this sound right?" It is "can I confirm this?"
Run any AI output that leaves your hands (a report, an email, a post, a slide) through four quick checks. Memorize them as SNND: Sources, Numbers, Names, Dates.
If the AI cites a study, link, or quote, find the original. Do not trust a citation just because it has a title and a year.
A common failure: the AI invents a real-sounding paper or attributes a real quote to the wrong person. Ask directly:
> "List the exact sources for each claim above, with links. If you are not certain a source exists, say so."
Then click the links. Fabricated URLs often 404 or lead somewhere unrelated.
Statistics, percentages, prices, and calculations are the highest-risk items. AI is famously shaky at arithmetic and loves a tidy-sounding stat.
Re-check every number against a primary source, or have the AI redo the math step by step. Even better, ask it to use a tool: in 2026, ChatGPT, Claude, and Gemini can all run code to compute exact answers instead of guessing.
People, companies, products, job titles. AI mixes these up constantly, especially for less-famous individuals. It might attribute a real quote to the wrong CEO or invent a "Director of X" who never existed. Confirm spelling and role before publishing.
"As of 2024..." may be wrong. Models have a training cutoff (the date after which they learned nothing new) and may not know recent events unless they search the web. Verify any date-sensitive claim, and check whether your tool actually looked it up or answered from memory.
Google's own guide on responsible AI use makes the same point: AI can make mistakes, so verify important information.
Here is the workflow in action. Suppose you asked an AI:
> "What percentage of the human body is water?"
It replies confidently: "Approximately 90% of the human body is water."
Plausible? Sort of. Water is everywhere in the body. But the real figure for an adult is roughly 50% to 65% (closer to 75% for newborns). The 90% number is wrong, and it would embarrass you in a health newsletter.
How you catch it:
1. Numbers flag: it is a stat, so it gets checked automatically.
2. Quick primary check: a glance at any medical source (Mayo Clinic, USGS Water Science School) gives the correct range.
3. Confirm before shipping. Done.
Thirty seconds of checking saves a public mistake.
You do not have to do all the checking by hand. Good promptingpromptingPrompt engineering is the practice of designing and refining text inputs to guide large language models toward accurate, relevant, and reliable outputs.View full definition → turns the AI into your first line of review.
Ask for confidence and caveats:
> "Rewrite the summary above. Mark any claim you are less than 90% sure about with [CHECK]. List what evidence would confirm each one."
Force it to separate fact from guess:
> "Which statements here are established facts versus your inference? Label each."
Use tools for math. Instead of trusting mental arithmetic, ask the model to run code. Here is a tiny example you can paste into ChatGPT, Claude, or Gemini's code feature, or run yourself in Python:
# Verify a claimed ROI figure
revenue = 36000 # revenue attributed to email campaign
spend = 1000 # amount spent
roi_per_dollar = revenue / spend
print(f"ROI: ${roi_per_dollar:.2f} for every $1 spent")
# Output: ROI: $36.00 for every $1 spentWhen the number comes from real code instead of a prediction, you can trust the arithmetic. You still verify the inputs, but the math itself is solid.
Not everything needs forensic review. Calibrate effort to consequences.
Brainstorming, draft outlines, casual rewrites, private notes. If a small error costs nothing, a quick read is enough.
Internal emails, team docs, first drafts of content. Verify the numbers and names. Skim the rest.
Anything public, legal, medical, financial, or sent to a client or executive. Check every source, number, name, and date against primary references. In 2026 there are already real cases of lawyers sanctioned for citing AI-invented court cases. Do not become one.
A simple rule: the closer it gets to a real person making a real decision, the harder you check.
Knowledge check
1. In the lesson's opening story, what was the actual problem with the marketing manager pasting the '$42 ROI' line into a client deck?
2. According to the lesson, why do AI hallucinations tend to be especially dangerous?
3. The lesson notes that large language models generate text primarily by doing what?
4. Select ALL correct answers about the SNND verification checklist described in the lesson.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about how the lesson recommends verifying AI-cited sources.
Sélectionnez toutes les réponses correctes.
Checking should be a habit, not a heroic act of willpower. Bake it into how you work.
Pass 1: Generate. Let the AI draft freely. Speed over caution.
Pass 2: Verify. Switch hats. Run SNND. Mark anything unconfirmed. Only then ship.
Keeping these passes separate matters. When you generate and verify at the same time, you tend to do neither well.
If a claim is important and you cannot find a source quickly, ask a second AI. Pose the same factual question to Claude and Gemini, for example. When they disagree, you have found a spot that needs a human and a real source. Agreement is not proof (both can be wrong), but disagreement is a useful red flag.
Note the errors your AI makes repeatedly. Maybe it always rounds stats up, or invents plausible LinkedIn-style job titles. Knowing your tool's habits makes you faster at catching them.
In 2026, ChatGPT, Claude, and Gemini all offer modes that search the web and show citations. Use them for anything fact-heavy. A linked source you can open beats an unlinked claim every time. But remember: a citation existing does not mean it says what the AI claims. You still open the link.
The shift that makes all of this stick is psychological. Stop reading AI output as a reader receiving facts. Start reading it as an editor reviewing an intern's draft.
An editor assumes there are mistakes and goes looking for them. An editor is not impressed by confident prose. An editor's name goes on the final product, so an editor checks.
That is your job now. The AI writes the draft. You decide what is true enough to ship.