NEW RESEARCH REPORT

How humans teach AI to perceive what isn’t said

How humans teach AI to perceive what isn’t said

Laptop displaying whitepaper cover, titled “The neuroscience of subtext: How humans teach AI to perceive what isn’t said”

Download the free white paper on how to build trust and operationalize perception

Your guide to perception for generative AI

LLMs write fluent sentences, but they often miss what humans mean. Tone, intent, sarcasm, hesitation, and cultural signals don’t live in words alone. This whitepaper translates neuroscience and psychology into practical, human-in-the-loop (HITL) workflows that help models read the room, align to brand voice, and respond appropriately in real interactions. You’ll learn how Theory of Mind, mirror neurons, and human pattern processing map to concrete annotation tasks that make perception measurable — and improvable.

Animation showing a text bubble playing an audio file, with pop-ups depicting markers for tone and intent of the message. For tone, the annotator marks the audio as “happy”, and for intent, the annotator marks the audio as “change” and “question”.

50% of companies buy/lease gen AI models from external vendors

Inside the report...

Essential insights and step-by-step playbooks to make
subtext visible in your data:

  • Neuroscience to practice: Theory of Mind, mirror
    neurons, and pattern processing—and how each
    informs perception-first annotation.
  • Where models fail: Real-world misses (sarcasm,
    idioms, prosody, culture) and how they erode trust,
    CSAT, and safety.
  • HITL workflows that work: Tone ranking boards,
    empathy side-by-sides, intent adjudication, cultural
    calibration, multimodal comparisons.
  • Designing for subjectivity: Preference data,
    rubric-based rationales, and live guideline updates for
    nuanced judgments.

  • Annotator diversity: Why native competence,
    cultural range, and experiential depth reduce blind
    spots.

  • Metrics that matter: Tone-alignment and empathy
    scores, IAA where agreement is expected, incidence
    tracking for cultural/brand violations.

  • Action plan: Sourcing playbooks, calibration
    cadence, RLHF/SFT integration, and deployment
    checks.

Machines don’t acquire social intelligence by accident.
We have to encode subtext

McKINSEY

Companies that invest in AI data quality see a 20% improvement in model performance, directly impacting their bottom line.

FORRESTER

Related resources

Article

Beyond words: 10 subtle layers of human context AI still struggles to understand

Article

Disfluency, slang, idioms, and subtext: teaching AI what words don’t say

Article

Inter-annotator agreement for subjective tasks in gen AI

ES