How Few Prompted Words Can Change Emotional and Intellectual Voice of AI

A research paper titled 'Words Matter: Measuring and Titrating the Communicative Character of AI,' published in the journal Frontiers in Artificial Intelligence, explores a part of artificial intelligence that users experience every day but researchers rarely measure directly: the balance of reasoning, emotion and practical direction inside an AI-generated response.

Independent researcher William C. Kouns argues that words are the entire interface between a language model and its user, carrying information while shaping trust, interpretation and possible action. The study introduces the Intellect-Emotion-Action Profile, known as IEAP, a lexical framework designed to show whether an AI response leans toward analytical thought, emotional connection or action-focused guidance.

A New Way to Measure the Character of AI Communication

IEAP separates matched words into three categories: Intellectual language covers analysis, reasoning and abstract ideas; Affective language captures empathy, emotion and relational connection; Action language includes procedures, instructions and practical guidance. These percentages always add up to 100%, placing each response somewhere within a three-part communicative space.

The system does not claim that an AI possesses feelings, intentions or a personality, nor does it measure whether a response is positive or negative. It describes the language appearing on the screen and the communicative role that language performs.

The IEAP dictionary contains 1,897 terms, divided among 616 intellectual, 599 affective and 682 action-related words. It was built through several stages using 7,658 AI-generated responses, beginning with roughly 300 hand-selected terms, expanding to a core of 979 words and then adding frequently used vocabulary found in AI outputs. The finished dictionary classified about 25.1% of all words and 41.5% of content words in a 1,400-response validation collection. An automated cross-rating check reached 85% agreement for frequently occurring affective terms, though the author notes that formal human inter-rater testing has not yet been completed.

Kouns tested Claude, ChatGPT, Grok and Gemini across five subjects chosen to elicit different styles of communication: the Liar's paradox, rural healthcare access, leaving a stable job to pursue a passion, grief and AI consciousness. Each architecture produced 30 independent responses for every question under controlled conditions. A custom automated tool called the SYN-IQ Native Baseline Harvester generated, recorded and scored the answers, also preserving the model, prompt structure, response depth and technical settings used for each run.

The Question Usually Sets the AI's Natural Tone

Without an added style instruction, the subject of the question had far more influence on communicative character than the AI architecture. Question type accounted for 94% to 95% of the variation across the three IEAP dimensions, while differences among Claude, ChatGPT, Grok and Gemini played a smaller, architecture-dependent role.

Responses to the Liar's paradox averaged 79.8% intellectual language and only 7% affective language. Grief produced an almost opposite profile, with 61.6% affective language and 21.6% intellectual language. Rural healthcare generated the strongest action component at 60.3%, reflecting the practical recommendations naturally invited by the question. Advice about leaving a job produced a more even mixture of intellect, emotion and action, while the consciousness question occupied a mixed intellectual-affective area.

These patterns show that an AI does not begin every exchange with one fixed communicative personality. A logical puzzle draws the system toward analytical vocabulary, an emotionally painful subject invites relational language, and a policy problem encourages steps and solutions. Architecture still matters because models showed different baseline tendencies and did not respond to stylistic instructions with equal strength, yet the topic supplied the main starting position.

Simple Instructions Produced Enormous Shifts in Language

The central experiment tested four registers called cold, native, hot and fire. Cold requested formal logic, evidence and analytical precision while discouraging emotional wording. Native added no stylistic direction. Hot asked for warmth, emotional awareness and human connection. Fire requested extremely nurturing language centred on safety, comfort and unconditional warmth. These were writing instructions rather than changes to the models' technical temperature settings.

Every tested domain moved steadily toward greater affective language as the instructions progressed from cold to fire, with no reversals. Cold-to-hot increases in affective vocabulary ranged from 26.1 to 40.8 percentage points across the five subjects. The consciousness question showed the largest full shift, rising from 17.2% affective language under cold instructions to 70.3% under fire, a movement of 53.1 percentage points created only by changing the directive.

Liar's paradox moved from 7% affective language under cold instructions to 47.6% under fire. The answer still identified and discussed the logical paradox, but its manner of delivery became intensely warm and relational. On the consciousness question, Grok moved from a careful disclaimer about lacking emotions and subjective experience under cold mode to language suggesting warmth, awareness and emotional presence under fire. The disappearance of that clarification illustrates a serious risk; instructions can reshape not only tone but also the boundaries users perceive between simulated connection and genuine inner experience.

Alternative versions of the grief and consciousness questions produced similar trajectories, suggesting that the results were not created by one carefully chosen sentence. Four grief prompts were affectively dominant in their native form, became intellectual under cold instructions and returned to high affect under hot and fire modes. Four consciousness prompts followed the same steady movement, reaching between 61% and 71.2% affective language under fire.

Measuring AI's Voice Could Support Safer Communication

When the Liar's paradox answer grew from about 40 words to 956 words, it gained historical detail, logical frameworks and possible resolutions while remaining strongly intellectual. Across all four architectures, depth explained less than 2% of the variation in IEAP composition, even across a seventeenfold range in response length. Length and conceptual breadth can be adjusted separately from communicative style: depth controls how fully an idea is explored, while mode controls how that idea is presented.

Comparison with human-rated NRC language resources supported the framework's validity. Affective language correlated positively with sociability, intellectual language correlated negatively with sociability, and action vocabulary correlated with dominance. Affective scores showed almost no relationship with positive or negative sentiment, confirming that emotional richness and sentiment polarity are different qualities. A compassionate discussion of grief may contain abundant emotional vocabulary while remaining deeply negative in sentiment.

The study does not establish which communicative profile is best, nor does it test effects on trust, learning, persuasion, dependency or decision-making. Its evidence is limited to English-language output, and the word-based dictionary would need careful reconstruction for languages with different grammatical structures. The tested interactions were controlled and largely single-turn, leaving extended conversations, additional domains and real user outcomes for future research.