45.3% believe AI fully understands them without being told

Innovation Nippon 2026 ran a ten-question literacy test on generative AI. Each question allowed multiple selections and counted as correct only if the respondent selected all the right options and nothing else.

The strongest results were practical: 53.9% fully correct on what to do after asking AI to plan a trip, 52.5% on the characteristics of generative AI, 50.4% on what to watch for when drafting a business email with it. The weakest were 13.3% on how to phrase instructions to reduce bias in image generation, and 14.0% on what generative AI actually does.

So the etiquette of using the tool is reasonably well known, while what kind of machine it is is barely shared at all. The report adds its own caveat, and it matters: completion rates fell on most items compared with the 2024 edition, but exclusive options such as "none of these is correct" were added, which raised the difficulty. These numbers should not be read as a simple decline.

What matters more than the score is which wrong answers people chose. The most-selected incorrect option, at 45.3%, was that generative AI fully understands the intent of your question without specific explanation and provides the optimal answer. Second, at 40.6%, was that it selects a suitable response from answers prepared in advance.

Those are opposite errors. The first treats the model as a mind reader; the second treats it as an index, like a search engine. Neither matches what is happening, which is a probabilistic assembly of likely next words from patterns in training data.

The first error carries a practical consequence. If the model genuinely understood your intent and returned the optimal answer, there would be no reason to structure your question and no reason to open a source. Checking would be wasted effort. The misconception manufactures a reason not to verify.

That gap in understanding lines up with separate government data. Japan's 2025 Information and Communications White Paper reports that 26.7% of people in Japan had used a generative AI service as of the 2024 survey, up from 9.1% the year before, but well behind the same survey's figures for the United States (68.8%), Germany (59.2%), and China (81.2%).

The reasons matter more than the gap. The most common reason for not using it was that it is not needed for one's life or work, and the next most common was not knowing how to use it. That is not disinterest. It is a missing on-ramp, which makes it a teaching problem rather than a motivation problem.

The same white paper asked about perceived risks. The items most often rated as very risky were criminal misuse by bad actors, being deceived by sophisticated fakes, and the possibility that an AI's answer to a question is not factual. One of the things people fear most is precisely the thing verification addresses.

Put those together and the prescription gets specific. Do not teach which buttons to press. Hand over one correct mental model.

AI does not read your intent. It assembles a plausible continuation from the material it was given. Once that lands, everything else follows without being taught separately. That is why you state your context and your sources. That is why you ask for citations. That is why you check the load-bearing claim at its passage. To someone expecting a mind reader, those steps look like unnecessary friction. To someone who knows how the machine works, they are simply how you use it.

The age breakdown supports reading this as an education question. Completion rates peaked among respondents in their thirties, followed by their forties, while people in their twenties scored low, which the report calls an undesirable result for promoting correct use of generative AI. Being a digital native and understanding AI turn out to be different things.

A mental model travels faster through a tool than through a lecture. When every answer arrives with its sources, when clicking a claim opens the exact passage, and when the system says the sources do not contain this rather than improvising, you learn what the machine is doing without anyone explaining it. The behaviour teaches the model.

That is what Tatsulok is built on. Answers come only from your own documents and curated collections, every claim links to the passage behind it, and an honest gap is treated as a first-class answer. We gather, you judge. The longer you use it, the more accurately you see the shape of the thing.

This week's drill, no signup needed. Ask your usual AI a question from your field twice. First the way you normally would, short. Then again, stating your assumptions, the sources it should use, and the shape of output you want. The difference between the two answers is the meaning of the sentence "AI does not fully understand your intent." Moving from the 45.3% to the 14.0% usually takes exactly one such comparison.