When the Numbers Are Right and the Summary Still Lies

Sweden’s PISA 2025 scores landed today. They are ugly. Reading down 21 points, maths down 18, science down 8. A third of 15-year-olds miss the basic level in reading and maths. That part is not in dispute.

I asked Gemini for the breakdown: students with a Swedish background, second-generation immigrants, and students born abroad.

It gave me the numbers. Then it wrote the moral of the story for me: the decline is “driven to a large extent” by students with a Swedish background crashing, while first-generation immigrants stay “stable (but low) or rise slightly.

Technically, Swedish-background students did drop more from their own 2022 starting point: 2.89 percent versus 2.51 percent for the second generation. The gap between those two percentage drops is 0.38 percentage points.

That is one jelly bean in a bag of 300. Not a landslide. Not a scandal inside the scandal. Just a rounding error wearing a press-release coat.

I am not here to rank children by passport. The point is smaller and more annoying. The model had the table. It still chose the headline that sounded safest. Correct digits, tilted sentence. That is what a safety layer looks like when it starts editing reality instead of blocking it.

If an AI can do the arithmetic and still sell you the wrong story, the filter is no longer protecting you. It is protecting the narrative.

Check the source yourself, then decide which sentence you trust: the table or the summary.

Skolverket: PISA 2025

Leave a Reply