What is Hallucination in AI Models?

Large language model (LLM) hallucination—frequently termed confabulation in cognitive science—is the tendency of generative artificial intelligence models to produce statements that appear authoritative, articulate, and plausible, but are completely ungrounded in factual reality or the provided context.

In creative writing or casual chat, hallucination can be harmless or entertaining. However, in professional knowledge management, corporate strategy reviews, medical research, and legal proceedings, hallucination is a catastrophic failure mode. A meeting note that falsely claims an executive approved a $500,000 budget or committed to an unverified deadline can derail organizations.

The mechanical root causes of hallucination

LLM hallucination is an emergent property of modern autoregressive Transformer architectures:

1. Probability over truth: Transformers are trained via next-token prediction: given a string of tokens, calculate the probability distribution $P(w_t \mid w_1, \dots, w_{t-1})$. The model optimizes for linguistic fluency and statistical plausibility, with no inherent concept of objective empirical fact.
2. Context window compression: When long meeting transcripts exceed model attention bounds, attention heads lose resolution, causing models to fill narrative gaps with generalized training patterns.
3. Sycophancy and overconfidence: Foundation models are fine-tuned via RLHF (Reinforcement Learning from Human Feedback) to generate helpful, cooperative responses. When faced with ambiguous questions, models often prefer inventing a plausible answer rather than admitting ignorance.

Strategies to eliminate hallucination in meeting records

Strategy	Mechanism	Efficacy in Note Apps
Strict Grounding (RAG)	Model can only extract facts present in the text	Very High
Verbatim Citations	Require model to quote exact sentence timestamps	Critical for trust
Deterministic Sampling	Set temperature to 0.0 or 0.1	High (reduces randomness)
Negative Constraints	Explicitly instruct: “If not mentioned, write None”	Very High

Anti-hallucination architecture in Oyma

Oyma rejects speculative, ungrounded AI summaries. Our meeting intelligence architecture enforces strict evidentiary grounding:

• Source transcript verification: Every summarized bullet point, decision, and action item is linked to the exact timestamp in the verbatim Whisper transcript.
• Conservative sampling: Summarization prompts use low temperature parameters to eliminate creative improvisation.
• Full human auditability: You can click any generated summary bullet to jump immediately to the underlying audio recording and verify the exact words spoken.

With Oyma, your meeting notes remain a faithful, trustworthy record of reality.