If a user shares “It’s fine” and the system announces, “Your friend is definitely angry,” it has turned an uncertain interpretation into an apparent fact. A more careful response would explain that the friend might genuinely be untroubled or might be closing the conversation, and that the preceding exchange could help distinguish between those possibilities.
Even people who know each other well get this wrong. Sensitivity to context should include recognising when the available context is insufficient.
This is where linguistics and AI research meet. Pragmatics identifies the aspects of communication worth examining: implied meaning, reference, communicative purpose and social relationships. Researchers can turn these into tasks that test whether a system responds appropriately when the context changes.
The same perspective is useful when writing prompts. “Help me tell my friend I’m not coming tomorrow” leaves much unspecified. Are you declining a casual invitation, cancelling an important commitment or trying to create some distance?
“My friend invited me to her birthday, but I need to look after my mother. Help me decline warmly and suggest another day to meet” supplies a situation, a relationship and a communicative goal. The system has more to work with when choosing its wording.
We can also ask AI to explore an ambiguity rather than resolve it prematurely: “My friend replied ‘okay’ after I cancelled. What are some possible interpretations, and how could I check in without putting pressure on them?” That approach leaves room for evidence we do not yet have.
Translation makes the issue particularly clear. Depending on the situation, mai pen rai might become “It’s okay,” “No, thank you. I can manage,” or “Never mind.”
Each translation performs a different conversational function: accepting an apology, declining help or withdrawing a request. Choosing between them requires understanding what the expression is doing in that exchange. These are illustrative examples, not results from a test of any particular AI model.
Across languages and cultures, the challenge becomes more complex. Directness can sound considerate in one setting and abrupt in another. Indirect language can protect a relationship while also creating room for misunderstanding. Success in one language therefore gives us only a partial view of a system’s ability to interpret communication across communities.
A deeper question remains: does responding appropriately amount to understanding language as humans do?
Emily Bender and Alexander Koller argue that a system trained solely on linguistic form cannot acquire meaning grounded in the world and in communicative intent, as they define meaning. On that account, fluency alone does not establish understanding. Their argument concerns learning from form alone; applying it to a particular AI system requires considering what that system learns from and how it connects language to other information. Bender and Koller (2020)
For everyday users, there are more immediate things to look for. Does the assistant take the preceding conversation into account? Can it explain which parts of its interpretation are supported by the message and which are guesses? When several readings are plausible, does it acknowledge them or ask for more information?
A short phrase such as mai pen rai opens up these much larger questions. Some meaning is carried by the words. Some comes from shared history, tone, relationships and what a speaker leaves unsaid.
AI may become better at reading between the lines. It also needs to recognise when there is too little evidence to do so.
Because sometimes, when someone says “It’s fine,” it really is.
