AI-generated SQL — 6 mistakes that run without an error (and how to catch them)

The dangerous AI query is not the one that fails — it is the one that runs and answers a slightly different question. Six of them, each one to solve right here:

1. The join that inflates a total

Asked for revenue, an assistant joins orders to their items and sums the order amount. Every order is repeated once per item.

SELECT SUM(o.amount)
FROM orders o
JOIN items i ON i.order_id = o.id;

On Atterna's practice marketplace this exact query turns €552,645 of paid orders into €1,239,938. Nothing errors; the number is just more than twice too big. Ask back: "Is any amount summed after a join to a table with several rows per order?"

2. Counting rows when you asked for people

Marketing asks "How many customers ordered?", and the assistant writes this:

SELECT COUNT(buyer_id)
FROM orders
WHERE status = 'paid';

Ask back: "Does this count customers or orders? Use COUNT(DISTINCT …) for people."

3. A LEFT JOIN undone by WHERE

Assistants often add a filter on the joined table in WHERE. The rows the LEFT JOIN was meant to keep disappear.

SELECT u.id, COUNT(o.id) AS paid_orders
FROM users u
LEFT JOIN orders o ON o.buyer_id = u.id
WHERE o.status = 'paid'
GROUP BY u.id;

Ask back: "Should this condition be in ON, so rows without a match stay with zero?"

4. NULL quietly changes the average

Empty values don't raise errors. They change what an aggregate covers.

SELECT AVG(rating) FROM catalog;

Ask back: "How many rows have NULL in this column, and should they count as zero, be excluded or be reported separately?"

5. Grouping by a timestamp

You ask for orders per day, and the assistant groups by the column that holds the order time:

SELECT created_at, COUNT(*)
FROM orders
GROUP BY created_at;

6. A correct query that answers the wrong question

The last mistake isn't in the SQL at all. The query is right; the conclusion it is used for is not.

-- the query is correct
AUTUMN20 orders   €91.81 average
without the code  €36.60 average

This is where AI help stops being enough: the assistant can compute both averages, but it may not ask who could use the code in the first place unless you do.

A two-minute routine before you use AI-written SQL

1. Say what one row of each table is, and what one row of the result should be.
2. Look at every join: can it repeat a row you then sum or count?
3. Look at every filter: is a condition on a LEFT JOINed table in WHERE instead of ON?
4. Check the counts: does COUNT count the thing you were asked about, and does DISTINCT belong there?
5. Check NULLs and dates: what happens to missing values, and is the latest period complete?
6. Read the conclusion separately from the query: does the number support the claim, or only sit next to it?

Paste the answer or the SQL into the AI answer checker and it flags joins that repeat rows, LEFT JOINs undone by WHERE, = NULL and incomplete periods, then writes the follow-up question for you. To build the habit on real mess — cancelled orders, refunds, baskets from several sellers — Start teaches SQL from zero (days 1–2 with no account, all ten free with an account) and case 1 puts a convincing, wrong number in front of you.

FAQ

Can I trust SQL written by ChatGPT or another AI assistant?

Trust it the way you would trust a colleague's first draft: it usually runs and is often right, but it can double-count after joins, drop rows in a LEFT JOIN or count the wrong thing. Check the joins, filters, counts and NULL handling before the number is used.

What is the most common mistake in AI-generated SQL?

There is no reliable ranking, but a frequent and costly one is summing or counting after a join to a table with several rows per record, which inflates totals without any error. The fix is to aggregate before joining or to join only what you need.

How do I check an AI's SQL if I'm still learning SQL?

Ask the assistant to state what one row of the result means, to show the row count before and after each join, and to explain every filter. Then compare a total with a simpler query on one table. Our checker flags common patterns, such as joins that repeat rows, in your browser; it can't run the query on your data.

Will AI replace the need to learn SQL?

AI makes writing SQL faster, which makes reading it more important. The work that remains is deciding what to count, catching a query that answers the wrong question and explaining what the number can't tell you.