How do I verify what AI produces?
The text came out beautifully — but is the figure in it correct? AI’s most dangerous trait is not that it gets things wrong, but that it says the wrong thing in a confident voice. 🔍
A well-written sentence is not a true one. That is why every serious use needs a verification step.
Short answer: four things are always checked — figures, names, dates, claims. The rest of the text is style. ✅
Where does the error come from?
BU BÖLÜMÜN ÖZETİ
- Built on prediction
- There may be no source
- It may be out of date
- It does not know your data
Not from a fault, but from how it works. ⚙️
Built on prediction
Where it does not know, the tool does not stop; it produces the most likely continuation. Gaps fill with fluent but invented sentences. 🧩
There may be no source
The information given may rest on nothing; asked for a source, it can even produce a title that does not exist. 📚
It may be out of date
On changing matters — regulation, prices, rates — it falls behind. These items are always verified externally. 📅
It does not know your data
If you did not supply your own product and pricing information, it puts a general guess there; the brief logic sits in the brief guide. 📎
The four checks
BU BÖLÜMÜN ÖZETİ
- Figures
- Names
- Dates
- Claims
On every output, automatically. 🧪
Figures
Percentages, amounts, ratios, durations. If a figure enters the text it is not published without a source. 🔢
Names
People, institutions, products and regulations. Similar-but-wrong names are the most frequent error. 🏷️
Dates
Effective dates, versions, periods. One outdated item makes the whole text look wrong. 📆
Claims
Absolute statements like “best”, “guaranteed”, “definitely”. These carry both accuracy and regulatory risk. ⚖️
How do you verify?
BU BÖLÜMÜN ÖZETİ
- With your own data
- With a primary source
- With a second tool
Three methods, in rising confidence. 🔎
With your own data
Price, scope and duration come from your source. Your list governs, not what the tool said. 📋
With a primary source
Regulation, official guidance, the manufacturer’s own page. Blogs and second-hand sources do not count as verification. 🏛️
With a second tool
Asking the same question elsewhere and comparing makes doubtful spots visible. Not proof, but a warning. 🔁
How much checking, for which work?
BU BÖLÜMÜN ÖZETİ
- Low risk
- Medium risk
- High risk
Three tiers by risk level. 🎚️
Low risk
Internal notes, idea lists, drafts. A quick read is enough; errors are costless. 📝
Medium risk
Customer emails, product descriptions, social content. The four checks apply in full. 📤
High risk
Contracts, proposals, regulation, health and finance content. Here the tool produces only a draft; an expert decides. ⚠️
How does verification become habit?
BU BÖLÜMÜN ÖZETİ
- The check line
- Sources beside figures
- A second pair of eyes
- An error log
By embedding it in the flow, not the rulebook. 🔄
The check line
One line at the end of every template: “figures, names, dates, claims checked?” That line reminds. ✅
Sources beside figures
Every figure entering the text carries its source. A figure without one comes out of the text. 📌
A second pair of eyes
Outgoing content is read by a second person. This was already true before AI. 👀
An error log
Caught errors go on a short list; recurring ones become rules in the brief. 📓
What should I do today?
BU BÖLÜMÜN ÖZETİ
- Step 1: set the risk tiers
- Step 2: add the check line
- Step 3: ban unsourced figures
- If you want help
Three steps, one day. 🪜
Step 1: set the risk tiers
Which work is low, medium, high? The checking load is distributed accordingly. 🎚️
Step 2: add the check line
Four words at the end of every template. A minute’s work. ✅
Step 3: ban unsourced figures
One rule, the biggest protection. A figure without a source is not published. 🔢
If you want help
Let us build the verification flow: use the consult your expert form. For your current usage see the business AI usage audit; the whole sits on the AI consultancy page. 🎯
Related reading from the archive: why outputs come out generic · what AI cannot do.
📝 Notes From the Field
Two AI-generated figures made it into a proposal; both were plausible and neither was correct. The client asked about them in the meeting. From that day one rule applied: a figure without a written source comes out of the text. It never happened again.
📖 Short Glossary
Invented output: the tool filling a gap it does not know with a plausible sentence. Primary source: the official or manufacturer origin of the information. Check line: the four-word reminder at the end of a template. Risk tier: classification by how far the output travels and what harm it could do.
⚡ Quick Summary
AI says wrong things in a confident voice. 🔍 Four things are checked on every output: figures, names, dates, claims. Verification uses your own data, a primary source and, when needed, a second tool. Checking load follows the risk tier.
🎯 Next Step
Let us build the verification flow: use the consult your expert form. The brief side sits in the brief guide; for your current usage see the usage audit.
Frequently Asked Questions
Sık Sorulan Sorular
Someone who knows the subject. A check by someone who does not is only a spellcheck. 👤
Final legal and medical text. The boundary sits in the delegation limits guide. 🚫
No: source lists can be generated too. Verification is not complete until the links are opened one by one and the content is seen to actually say what was claimed.
They are more checkable because they show sources, but the interpretation of a source can still be wrong. Trusting them without opening the link carries the same risk.
Some of it, and that is normal. A net gain remains; besides, the cost of one published wrong figure exceeds months of checking time.