AI usage
Japan treats AI as a dictionary, the US treats it as a colleague
Asked what generative AI is most like, the second most common Japanese answer was "dictionary or encyclopaedia". In the US and China it was "friend". That difference produces a difference in results.
Occasionally a survey question tells you more than the statistics around it. In Japan's 2026 White Paper on Information and Communications, that question asks what generative AI is most comparable to.
The leading Japanese answers included "dictionary or encyclopaedia". In the United States and China the top answer was "friend". In Germany it was "counsellor".
It reads as a question about metaphors. It is actually a question about process design.
Used as a dictionary, the work does not change
A dictionary answers what you ask it. It assumes you already have the question, and every judgement after the answer stays with you.
That mode has a firm ceiling. Individual tasks get slightly faster, but the number of steps stays the same. Looking things up takes less time, while drafting, checking and getting approval continue exactly as before.
This also explains why Japanese usage concentrates so heavily in text. The white paper puts image and video generation at 16.3% in Japan against 47.6% in Germany and 41.0% in the US, and code generation at 7.0% against 35.4% in the US.
As long as the mental model is a reference book, the use cases stay textual.
Used as a colleague, steps disappear
Answers like "friend" and "counsellor" imply a back and forth. You show your draft, you get criticism, you revise. You get told what assumption is missing. You get offered an alternative you had not considered.
In that mode, AI takes over an entire step rather than assisting inside one. First drafts. Review checklists. Finding what was omitted. A step a person used to perform from scratch stops existing.
Because a step disappears, the shape of the process changes. That is the connection to another white paper finding: only 6.0% to 11.9% of Japanese firms say they have redesigned a business process around AI. Most workplaces have not yet removed a step.
Why the dictionary model took hold in Japan
Part of it is where adoption started. "Make research faster" is safe, easy to explain and hard to argue against. As a first move it was the correct one.
The other part is that nobody owns the output. Without an owner, an AI suggestion cannot be built into the process. If it cannot be built in, it can only be treated as reference material. A tool you treat as reference material is a dictionary.
So the dictionary pattern is not a training failure. It is what an unassigned responsibility produces.
If you want to change how AI is used, do not start with prompt training. Start by deciding who is accountable for the output.
Three decisions that move you across
- Pick one step for AI to own. Own, not assist. Decide that first drafts are produced by AI, and mean it.
- Name the person accountable for that output. Write down the standard they apply. Without a standard, everyone reviews everything, which is slower than before.
- Delete one human step. If nothing was removed, you added a step rather than replacing one, and there will be no measurable effect.
The third is where the resistance appears. It is also the only one that changes the process.
The metaphor question captures how Japanese companies currently position AI. Repositioning it comes down to deciding who is responsible for what, and no purchase order will do that for you.