Excessive transcription and errors in processing with LLM could change our communication habits

There are two postings recently published on golem.de which highlight the
risks of AI usage in the first place / at the first step and in the further processing
of AI produced artifacts:

1. Excessive transscription of every (business) conversation and/or meeting may lead into a situation, where each and every peace of word, even not important side conversations and small talk could have serious legal and or business consequences. And further, given the fact that LLM are not error free, one single small error such us recording relevant instead of irrelevant in the transcript could lead into legal damages when this transcript is used in court, see more
   https://www.golem.de/news/kuenstliche-intelligenz-anwaelte-warnen-vor-ki-transkriptionstools-in-meetings-2605-208512.html

2. But usually nobody wants to read very long texts, so you let an LLM summarize such transcripts into meeting notes or something similar. May be the mentioned above transcripts are stored in a database for later filtering and/or retrieval in a LLM RAG system. And here comes the second important point. Analysis published by microsoft scientists shows that
   current LLMs may introduce further sparse but severe errors to these transcripts, when they are processed further and that silently corrupt documents, compounding over long interaction, such as long context, thinking and LLM assistants working in a loop, see more in
   https://www.golem.de/news/kuenstliche-intelligenz-ki-modelle-zerstoeren-dokumente-bei-delegation-2605-208559.html and here
   https://arxiv.org/abs/2604.15597

   Large Language Models (LLMs) are poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm (e.g., vibe coding). Delegation requires trust - the expectation that the LLM will faithfully execute the task without introducing errors into documents. We introduce DELEGATE-52 to study the readiness of AI systems in delegated workflows. DELEGATE-52 simulates long delegated workflows that require in-depth document editing across 52 professional domains, such as coding, crystallography, and music notation. Our large-scale experiment with 19 LLMs reveals that current models degrade documents during delegation: even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows, with other models failing more severely. Additional experiments reveal that agentic tool use does not improve performance on DELEGATE-52, and that degradation severity is exacerbated by document size, length of interaction, or presence of distractor files. Our analysis shows that current LLMs are unreliable delegates: they introduce sparse but severe errors that silently corrupt documents, compounding over long interaction.

Why this two points could lead in a change of our communication habits?

1. Obsessive transcription could push all side conversations and all small talk to a place outside a meeding
2. The knowledge of transcription errors could lead to a LLM adjusted communication style and language subset, which is less prone to transcription errors

3. The fact, that further LLM processing introduces errors and that current AI assistants based on Gemini 3.1 Pro, Claude 4.6 Opus or GPT 5.4 loose about 25% of the content when going over 20 further processing steps could lead in kind of a court safe conversation and communication style where you enforce your (legally safe) weighting, valuation or judgement with certain semantic formula, such as

   • neither nor
   • in each and every aspect not relevant, in all completely irrelevant
   • I don't like this way, because this way leads to self censorship
   • this is uncomfortable, because this could move each and every team building process away from the public towards dark back stage places
   • last, but not least

Last, but not least, you should ask yourself:
Is this possible change in communication habbits or conversation styles an avoidable situation?

• Do you want court safe communication in meetings, LLM adjusted language, usage of semantic formulas and all side conversations and all small talk to be pushed into dark backstage places?
  Most probably not.

• Is it energy saving to add extra tokens to the LLM context just to enforce your legally safe semantic formula remains intact after every AI processing step?
  Definitely not.

• Could any AI regulation help, where each and every AI created and/or LLM processed content is to be marked as AI generated?
  Yes of course.
  If you are in the EU, you might be lucky, because the EU AI Act article 50 goes in this direction.