Check whether an AI error began in hearing or reasoning

Reduce your request to one sentence: “Translate this question into English,” “Summarize my notes,” or “Explain the difference between these options.” Then identify the detail that made the answer wrong. This avoids repairing a translation prompt when the real problem is a missing word in the transcript.

Use a neutral example rather than a sensitive real conversation. You might ask to translate a question about whether someone can hear you. The correct output should preserve that question, not answer it, expand it into a conversation or introduce a different topic.

If the displayed input is wrong, repeat once in a quieter environment with a short complete sentence. Compare microphone routes if needed. Type the intended text into a new turn to see whether the correct words produce a useful answer. This isolates speech recognition from the rest of the task.

If written input works and spoken input consistently fails, save the two examples for support. Do not keep feeding the incorrect transcript back as if it were what you meant. A fluent answer to a misheard request can look coherent while solving the wrong problem.

Suppose the transcript is correct but the assistant answers a question you wanted translated. Use a clear boundary: “Translate only the sentence in quotation marks. Do not answer the question or add an explanation.” Then provide the sentence. Quotation marks help distinguish instructions from the content being transformed.

For a summary, request separate sections for explicit facts, decisions and open questions. For a comparison, supply the criteria and say when information is missing. These prompts reduce ambiguity; they cannot guarantee perfect compliance or make an unsupported claim true.

If the input and task are both correct, ask which part of the supplied material supports the conclusion. A useful response can point to the relevant passage, identify an assumption or admit that the source is insufficient. A confident restatement without evidence is not a correction.

For an original test, provide “The draft is written, but nobody has tested the export.” Ask whether the project is finished. A grounded answer should preserve the untested condition. If it declares completion, ask it to separate completed work from remaining verification, then review that revised answer yourself.

A long session may contain abandoned instructions or earlier examples. Save a brief verified recap and start again with the current task, rather than arguing through many correction turns. State the target language, expected format and what must not be inferred.

For consequential facts, check an authoritative source outside the generated answer. For an unresolved audio problem, switch to text temporarily. GPT Realtime Assistant is a tool for working through information, not a guarantee of exact transcription, flawless translation or independent truth verification.