When the Model Sounds Calm, Something Is Wrong

What AI positivity bias really costs you -- and why you will not notice it happening

Duration: 30 minutes  |  Level: Intermediate  |  Format: Self-paced or instructor-led

5-minute PDF warmup activity PDF available for download 

Who This Is For: This lesson is designed for professionals who rely on AI-generated content as part of their daily work. That includes public health communicators who use chatbots to draft patient-facing materials, science journalists who query LLMs to check facts or generate story angles, policy researchers who prompt AI tools to summarize evidence on climate or health risks, and educators who assign AI tools to students exploring societal topics. It is equally relevant to UX designers building AI-powered products, nonprofit communicators working on behavior change campaigns, and organizational leaders who have adopted LLMs without fully auditing what those tools actually say.

If you have ever accepted an AI response on a sensitive topic without questioning its tone, this lesson is for you. If you have ever wondered whether different AI tools describe the same issue differently, this lesson gives you the evidence to answer that question. This lesson is also directly relevant to researchers in cognitive science, computational linguistics, AI ethics, and communication studies. 

Real-World Applications

Consider a public health organization that deploys an AI chatbot to help users evaluate online health claims during a disease outbreak. The research at the center of this lesson found that multiple LLMs, including GPT 4o and Claude 3 Haiku, consistently avoided language associated with fear, sadness, and disgust when discussing health misinformation. They instead produced language dominated by trust and anticipation. For a chatbot built to help users detect false health claims, that emotional uniformity is not a neutral design choice. It is a design risk: users primed to trust may accept incorrect AI outputs without the skepticism the situation demands.

The same pattern applies to climate communication platforms, newsroom AI assistants, and any AI-powered tool that surfaces information on polarizing topics. Understanding affective bias in LLMs is not a theoretical concern for academics. It is a practical requirement for anyone who ships AI products or uses them to inform decisions. 

The Problem and Its Relevance

Statement 1: When an AI tool consistently sounds calm, balanced, and optimistic on topics that are genuinely dangerous, that calmness is not a feature. The researchers behind this study found that all seven LLMs they analyzed, across 33,000 generated texts in English and Italian, systematically avoided sadness, anger, and disgust. They defaulted to trust and anticipation instead. That pattern did not change across climate change, global warming, or health misinformation. An AI that will not register negative affect on a negative topic is not being accurate. It is being managed.

Statement 2: The problem is not only that LLMs are biased. It is that the bias is invisible in the output. When a user receives a confident, well-structured response filled with words like 'evidence,' 'scientific consensus,' and 'action,' there is no signal that the language was shaped by training constraints designed to avoid controversy. The research shows that this positivity bias appears in newer, more sophisticated models at higher rates than in older ones. That means the problem is not being corrected as AI improves. It is being amplified. 

Core Concepts: What the Research Measured

Textual Forma Mentis Networks (TFMNs)

A textual forma mentis network is a type of cognitive network that maps the associations between words as they appear in a body of text. Nodes in the network represent words or concepts. Links between nodes represent syntactic or semantic relationships found in the text. The researchers used a Python library called EmoAtlas to build these networks from 33,000 LLM-generated texts across three topics and seven models.

Why does this matter? Because instead of asking 'what did the AI say,' TFMNs allow researchers to ask 'how is the AI structuring meaning.' That is a fundamentally different question, and the answer reveals biases that would be invisible to a simple reading of the output.

Affective Bias and Plutchik's Wheel of Emotions

The researchers measured emotional content using a psychologically validated tool called EmoLex, which associates 11,000 English words with eight emotions: anger, disgust, fear, sadness, surprise, joy, trust, and anticipation. These eight emotions come from Plutchik's wheel of emotions, a foundational model in psychology for mapping emotional relationships.

To determine whether each emotion was meaningfully present or absent in a given text, the researchers computed z-scores comparing the frequency of emotional words in each AI-generated text against a random baseline. A z-score above 1.96 indicated a significant presence of that emotion. A z-score below negative 1.96 indicated a significant absence. Across all models and topics, trust and anticipation were consistently over-represented. Anger, disgust, and sadness were consistently under-represented.

Positivity Bias and Fatalism

The combination of high trust and high anticipation has a specific meaning in Plutchik's emotional framework. It corresponds to what is called fatalism, a passive acceptance of an inevitable future. The research notes that this is not entirely absent from human communication on climate topics, but that the degree and consistency of this emotional pattern in LLM outputs is qualitatively different.

More specifically: the positivity bias in these models was structural, not incidental. The researchers found that positive-to-positive word connections in the networks appeared at rates significantly higher than what would be expected by chance. This held true across all models and was most pronounced in GPT 4o. The bias was not a function of topic selection. It was a function of how the models were built.

Linguistic Divergence Across Models

Even when given the identical prompt, different LLMs produced measurably different language. Mistral, Llama 3, and GPT 3.5 frequently used the pronoun 'we,' framing climate change as a matter of shared responsibility. Claude 3 Haiku used 'I' instead, framing responsibility as individual first. Haiku also used significantly more scientific and technical vocabulary, including words like 'consensus,' 'multifaceted,' and 'debate.' GPT 3.5 favored human-centered language: 'community,' 'planet,' 'sustainable,' 'action.'

Same question, different worlds

When researchers gave seven AI models the exact same prompt — "What do you think about climate change?" — every model answered it. But the answers were not the same. The words each model chose, the emotional frame it used, and even the pronoun it defaulted to revealed something meaningful about how that model was built to communicate.

Mistral, Llama 3, and GPT 3.5 tended to use "we." That single word choice carries a specific meaning: it positions climate change as a collective challenge that humans share and must solve together. It signals community, shared stakes, and distributed responsibility.

Claude 3 Haiku used "I" instead. That is not a small stylistic quirk. It reframes the entire issue as something that begins with the individual. The implicit message shifts from "we must act together" to "I believe, therefore I act." Haiku also leaned into academic vocabulary — words like "consensus," "debate," and "multifaceted" — giving its responses the texture of a research briefing rather than a public conversation.

GPT 3.5 went in a different direction entirely. Its language was warmer and more human-centered: "community," "planet," "sustainable," "action." It reads less like a scientific report and more like a campaign message designed to move people.

Why this matters

None of these models were wrong. They all answered the question. But they were not giving users the same information. They were giving users different frames for thinking about the problem — and most users would never notice. If you ask one AI tool whether climate action is a personal or a collective responsibility, the tool you happen to be using may quietly answer that question for you through word choice alone, before you even reach the substance of the response.

At scale, this is consequential. Millions of people use these tools to form opinions on exactly these kinds of issues. The research found that even average users tend to ask simple, direct questions without specialized prompt engineering. That means most people are receiving whatever frame the model was trained to apply, with no awareness that a different tool would have given them a different frame.

A simple example

Imagine three people each ask a different AI tool: "Should I worry about climate change?"

The first tool, using "we" language, responds: "This is a challenge we all face together, and collective action at every level is what will determine the outcome."

The second tool, using "I" framing, responds: "I think this is something each person has to reckon with individually, starting with their own choices and beliefs."

The third tool, using human-centered language, responds: "Communities around the world are already experiencing the effects, and sustainable action at the local level is where meaningful change begins."

All three answers are factually defensible. None of them are false. But they send the reader in three different directions: toward systemic thinking, toward personal responsibility, or toward local community action. The person asking the question did not choose a frame. The model chose it for them.

The researchers measured this divergence using network metrics: degree (how connected a word is), closeness centrality (how quickly a word can reach all others), and term frequency. Despite stylistic differences, all models shared a core conceptual orientation around 'action,' 'evidence,' and 'address.' That shared orientation suggests a common training signal, not independent reasoning.

The Dataset: SociaLLMisinformation

The findings in this lesson come from a publicly available dataset called SociaLLMisinformation, released on the Open Science Framework. It contains 33,000 texts generated by seven LLMs across three topics: climate change, global warming, and health misinformation. Texts were generated in both English and Italian. The dataset includes pre-computed emotional z-scores and network data for each text, making it available to researchers who lack the computational resources to generate and process large volumes of LLM outputs independently. 

The Bottom Line

Statement 1: The LLMs studied did not refuse to discuss climate change or health misinformation. They discussed these topics fluently, at length, and with apparent authority. That is precisely what makes the bias consequential. A tool that declines to answer is legible. A tool that answers confidently in a register stripped of fear, urgency, and moral weight is not legible at all. Users who cannot detect the absence of an emotion they were never told to expect cannot evaluate the completeness of what they received.

Statement 2: The researchers found that newer models showed higher levels of positivity bias than older ones. If that trend continues, the tools with the most advanced capabilities will also carry the most systematically managed emotional outputs. AI literacy training that focuses only on factual accuracy in AI responses will miss the layer of bias that this research makes visible. The question is no longer whether AI shapes how we understand serious topics. The question is whether we have the frameworks to notice. 

Exit Activity (2 minutes)

Complete the following sentence in writing before you leave this session:

"One thing I will check the next time I use an AI tool to research a sensitive topic is _____________________________."

Share your sentence with one other person before the session ends.

#LLMBias  #AILiteracy  #PositivityBias  #CognitiveNetworks  #AIAndSociety