AI Conversations: Share the Context Behind the Words

Imagine a creator looking at a fifth poster proof and typing to an AI assistant: “Wonderful, another round of changes.” The creator knows that the same misspelt name has returned three times. Their frustration feels obvious. Yet the sentence itself could also express enthusiasm about another opportunity to improve the design.

The missing information is the history that makes the remark mean something. For creators and business owners, a useful question is therefore: which parts of my intended meaning are actually available to the assistant? Research on communication offers a reason to examine that gap before treating a fluent reply as confirmation that everything was understood.

Notice the information you are adding in your head

Keysar’s April 1994 study examined what is called the illusory transparency of intention. Across four experiments, people read scenarios containing information that clarified whether a speaker meant a comment sarcastically. They then judged how an addressee without that information would interpret it.

Knowing the speaker’s intention made readers more likely to think the addressee would perceive it too. This happened even in written communication, where vocal intonation could not supply the missing clue. The finding concerns judgements in those experimental scenarios; it does not mean that people never recognise sarcasm.

In the poster example, mentally supplying the repeated mistake changes how “wonderful” sounds. The practical issue is whether that background was communicated. Here, egocentric bias refers to using one’s own viewpoint as the reference for an interpretation.

Separate confidence from successful communication

Kruger, Epley, Parker and Ng’s December 2005 research on e-mail investigated this problem across five experiments. People tended to overestimate how effectively they could convey tone electronically. The researchers linked that overconfidence partly to difficulty setting aside the tone the sender experienced while composing a message.

These were studies of communication between people, not tests of present-day AI systems. Their relevance here is the distinction between an intended meaning and evidence that a recipient identified it. Feeling that a sentence is unmistakable does not settle how it reads from another information position.

For an AI-assisted task, make that distinction visible in the work. If your intended meaning is dissatisfaction with a recurring error, the next draft should address that error. A friendly acknowledgement on its own gives you little to inspect.

Read the 2026 AI study as evidence about expectations

On 7 August 2026, Lea Marie Petrasch and Regina Jucks published an experiment involving 327 university or college students. Participants viewed German-language scenarios of student–chatbot exchanges and judged whether the chatbot would recognise sarcasm. They also selected its expected next response.

Background information suggesting a sincere or sarcastic intention shifted both kinds of judgement. The final ambiguous statements were spoken neutrally. Participants’ knowledge therefore influenced what they expected the chatbot to understand, beyond what that neutral utterance alone conveyed.

The distinction matters: this measured human expectations about illustrated exchanges, not a chatbot’s actual sarcasm-detection accuracy. The sample and scenarios limit generalisation, and there was no human-recipient comparison condition. The order of questions may also have influenced expected responses. It would overstate the evidence to conclude that all AI users make the same mistake, or that every current model fails to recognise irony.

Give the remark a purpose in a fictional task

Return to the imagined poster. Suppose the creator wants a firm correction request to send to a supplier. The useful brief would explain the repeated spelling problem, identify the approved spelling and specify the requested action. It would also distinguish a private expression of frustration from the message intended for the supplier.

For example: “I am frustrated because the guest’s surname is still wrong after three correction requests. Draft a courteous, firm message asking for the approved spelling to be used and checked before the next proof. Keep the event details unchanged.” This is an invented brief, not a transcript or a claim about an actual supplier.

The creator would still need to provide the correct spelling and relevant proof. Adding emotional context does not supply missing factual details. It clarifies why the correction matters and what kind of help is wanted.

Notice how the brief separates three things: the writer’s feeling, the evidence behind it and the action to request. You can keep a humorous opening while making those decisions explicit. There is no need to make a business task depend on whether the assistant guesses the joke.

Make the interpretation something you can check

The following routine is an applied editorial suggestion, rather than a procedure validated by the studies above. Use it when an ambiguous remark could send a draft or decision in the wrong direction:

• State the intended meaning: say whether you are expressing approval, disappointment, uncertainty or a request for revision.
• Supply the relevant background: include the fact that changes the interpretation, such as an earlier correction that remains unresolved.
• Name the requested action: specify what the assistant should produce or examine next.
• Inspect the interpretation: ask for a short paraphrase of the task, then compare it yourself with the intended meaning and the resulting work.

Treat that paraphrase as an output to assess. If it introduces a motive you did not provide, correct it. If two interpretations remain plausible, resolve the difference before relying on a draft. Keep the paraphrase short enough to compare with the brief and the resulting work.

Keep the context relevant and the conversation natural

Relevant context need not become a complete personal history. For the poster, the recurring spelling error and approved version matter. Unrelated experiences add little to that correction request. Choose the background needed for this task and preserve the distinction between what happened and what you infer from it.

When work moves into another conversation, the continuity problem becomes practical too. MaryChuks’ guide to keeping project decisions clear across chats explains how to carry approved details forward. Use such a note to preserve the facts that give a later instruction its meaning.

Clarity leaves room for personality. You can joke, express annoyance or change your mind while stating the next action plainly. Before approving the result, check that it responds to the intention you made explicit. Shared understanding becomes more useful when you can see its consequences in the work.

Sources

Primary research checked on 28 October 2026. The practical brief and review routine are editorial applications, not tested interventions from these papers.

• Keysar, B. (April 1994). The illusory transparency of intention: Linguistic perspective taking in text. Cognitive Psychology, 26(2), 165–208.
• Kruger, J., Epley, N., Parker, J. and Ng, Z.-W. (December 2005). Egocentrism over e-mail: Can we communicate as well as we think?. Journal of Personality and Social Psychology, 89(6), 925–936.
• Petrasch, L. M. and Jucks, R. (7 August 2026). Linguistic perspective-taking in verbal communication with LLMs: Evidence of an egocentric bias. Frontiers in Psychology, 17, 1894763.