Talking to an AI: Guide to Voice-Based Information Processing

Artificial intelligence (AI), including large language models (LLMs), is sometimes used in The Sunil Abraham Project (TSAP) for recording and processing information through voice. This is currently an experimental part of the project’s information-processing workflow. Tools such as Google Gemini’s meeting note-taking and ChatGPT’s voice mode are used as assistive tools to capture, organise, process, summarise, and examine information provided through speech. They are not used as direct content-creation tools.

Information recorded or processed with the assistance of AI does not automatically become final project content. The resulting material goes through human review, checking, and verification before it is relied upon. Only relevant parts of the processed transcript or other AI-assisted output may subsequently be incorporated into the project’s records or publications; the complete AI output is not treated as the final record.

Voice communication with an AI is also different from ordinary conversation between people. People routinely rely on shared memory, context, pronunciation, previous conversations, and an assumption that the other person knows what they are referring to. An AI may not have access to the same context. This becomes particularly important when the speaker is communicating in a language that is not their first language, or when names, Indian languages, regional expressions, technical terms, and other less common words are involved.

Tools used or considered

The following tools are used or being considered in this experimental workflow. They are selected because they can combine voice input with AI-based transcription, processing, or interaction. Ordinary voice recorders are not included because they would require a separate transcription and processing stage.

• Google Meet with Gemini AI notes — Used for meetings, where Gemini takes notes and transcribes the discussion.
• Google Gemini for in-person meetings — Used for in-person conversations, with Gemini providing transcription and AI-assisted notes.
• ChatGPT Voice — Used to provide information directly through voice and have ChatGPT process it conversationally.
• ChatGPT Live / Gemini Live — Real-time AI conversations. These are known options but have not been used extensively in TSAP so far.

The following principles are intended to make voice-based communication with an AI clearer and more reliable.

Use common sense

Remember that you are talking to an AI system such as ChatGPT or Google Gemini. Give it the context it needs, but do not assume that it shares your memory, assumptions, or access to previous conversations and documents.

There is also no need to over-explain everything. An AI may already understand technical terms, specialised vocabulary, and concepts that would require more explanation when speaking to another person. Use your judgement about what needs to be explained and what does not.

Speak clearly and do not rush

There is generally no need to speak unnaturally slowly, but speaking clearly and allowing small pauses between separate thoughts can make voice recognition and subsequent processing easier.

Natural speech is appropriate. The important consideration is that words, sentences, and changes of thought should be sufficiently clear for the AI to identify them correctly.

Use precise words, and slow down for uncommon words

There is usually no need to replace a technical or unfamiliar word with a long explanation simply because it might be difficult for a human listener. An AI may already understand the word.

Names, Indian words, regional expressions, uncommon terms, acronyms, and technical vocabulary may nevertheless be more susceptible to voice-recognition errors. Such words should be spoken clearly and, where necessary, slowly or more than once.

For particularly important or unusual terms, spelling the word can also help.

Stay with one topic at a time

It is useful to complete one subject before moving to another. Rapidly moving between unrelated subjects can make it harder for an AI to determine which information belongs to which discussion.

When several subjects need to be recorded, deliberately marking the transition can reduce confusion.

For example:

“That is the background. Now I want to discuss the website.”

Give the AI structure

Voice communication can include simple verbal markers that help the AI understand the organisation of the information.

Examples include:

• “New topic.”
• “Background.”
• “Sunil said…”
• “My response is…”
• “This is an action item.”
• “This needs verification.”
• “End of this section.”

These markers are particularly useful when a conversation contains multiple speakers, several subjects, or a mixture of facts, opinions, questions, and proposed actions.

Be especially careful with names, dates, numbers, and other precise information

Small transcription errors can significantly change the meaning of information. Names, dates, amounts, page numbers, version numbers, identifiers, URLs, acronyms, and similar details therefore deserve particular attention.

If an AI appears to have misunderstood an important word or number, correct it immediately rather than assuming that the error will be corrected later.

For example:

“The date is 17 September, not 7 September.”

Distinguish facts from ideas, opinions, and uncertainty

When providing information by voice, it can be useful to tell the AI what kind of information is being provided.

A speaker may be stating:

• something they know to be a fact;
• something they remember but have not checked;
• an opinion;
• a hypothesis;
• a proposal;
• something said by another person;
• or something that still needs verification.

This distinction should be preserved during AI processing. Information should not become more certain merely because the AI has rewritten it in fluent language.

Do not rely on conversations or documents that the AI cannot access

Humans frequently use references such as “I told you yesterday”, “we discussed this last week”, or “look at the document I sent you earlier”. These references work when both people have access to the same memory or material.

An AI may not have access to the particular conversation, meeting, message, or document being referenced. When the information is important, repeat the relevant information or provide the document or notes again.

A reference to an inaccessible source should not be treated as evidence that the AI has seen it.

Treat AI processing as assistance, not verification

An AI can help record, transcribe, organise, summarise, compare, and process information. It can also help identify questions, inconsistencies, missing information, or areas that may require further investigation.

These functions do not, by themselves, establish that the underlying information is correct.

A fluent AI output can still contain a transcription error, misunderstanding, omission, or unsupported assumption. Important information should therefore pass through appropriate human checking and verification before being treated as final.

The objective is to use the AI as an additional layer of assistance between human communication and the subsequent human-reviewed record, rather than treating the AI as the author or verifier of the information.

📄 This page was created on 19 September 2026. You can view its history on GitHub, preview the fileTip: Press Alt+Shift+G, or inspect the underlying source.