A new study from Harvard Kennedy School has identified a significant divide in how women and men experience artificial intelligence. According to research published October 14 by the Shorenstein Center on Media, Politics and Public Policy, 57% of women report rarely or never using AI, compared with 43% of men. Women also express greater concern about the technology, report less access to training and workplace encouragement, and are substantially underrepresented among the people developing AI and shaping its public narrative.
USA Today reports that women rated their concern about AI at an average of 6.4 out of 10, compared with 5.9 among men. Men expressed greater curiosity about the technology, while women were more likely to acknowledge uncertainty when answering questions about it. The researchers caution against interpreting those differences as evidence that women lack the ability to understand or use AI. In fact, some of the findings suggest that confidence and actual knowledge are different things.
This distinction deserves considerably more attention because the technology industry has a long history of equating adoption with acceptance and confidence with competence. When people are reluctant to use a new technology, the instinct is often to educate them about its benefits, make it easier to use or develop more persuasive marketing. Those interventions may address some barriers, but they do little to resolve concerns about how the technology actually behaves.
For an industry building systems that will increasingly make decisions, conduct transactions and interact directly with people, skepticism should be treated as useful information about what users expect from the technology.
Trust Is Not the Same as Familiarity
The Harvard researchers found that women reported lower levels of AI familiarity, confidence and trust across multiple measures. They were also more likely to indicate uncertainty rather than guess when answering factual questions about AI. Men answered somewhat more questions correctly on average, but they were also more inclined to express confidence, including in situations where that confidence was misplaced.
Those differences matter because self-reported confidence can influence how people approach technology without necessarily indicating how well they understand its risks. Someone who readily trusts an AI-generated answer may be more likely to use the system, but that behavior does not establish that the answer is accurate or the decision appropriate.
Conversely, a person who questions an AI recommendation, verifies its claims or declines to delegate a consequential decision may be exercising precisely the judgment that responsible AI adoption requires.
The researchers recognize this distinction in their conclusions. They do not argue that every demographic group should use AI more frequently or trust it more enthusiastically. Their concern is whether people have equitable opportunities to understand the technology, influence its development and use it when doing so genuinely benefits them.
That is a considerably more useful objective for organizations developing human-facing AI. Greater trust should follow better performance, stronger accountability and meaningful human authority over the system’s behavior.
Who Designs AI Influences Whose Needs Are Addressed
The study also identifies substantial disparities in who participates in AI development and whose perspectives receive attention.
Women remain underrepresented in technical and leadership positions, and the researchers found that men were named in AI news headlines 14.6 times as often as women. Men were cited as experts 4.3 times as often, revealing a considerable imbalance in the public discussion surrounding a technology expected to influence nearly every aspect of society.
Representation matters in this context because development decisions involve assumptions about how people communicate, what kinds of assistance they need and which risks deserve priority. Those assumptions influence everything from training and evaluation to product design and deployment.
An AI system optimized primarily for task completion may satisfy the requirements its developers established while failing to account for the experience of the person using it. A model may produce a technically correct answer that is poorly suited to the situation, misunderstand the intent behind a request or continue an interaction after its behavior has become inappropriate.
These problems can affect anyone. The research does not establish that women experience every AI failure more frequently than men, and it would be irresponsible to assume that all women share the same concerns. Nevertheless, the findings reinforce the importance of including a broader range of human perspectives in AI design, evaluation and governance.
Organizations should be able to define what appropriate AI behavior means for the people they serve, rather than inheriting every behavioral assumption from the foundation model provider.
The Human Side of AI Is an Engineering Problem
Most AI governance systems concentrate on the machine. They monitor data access, restrict tools, enforce permissions and determine which actions an agent can execute. These controls are increasingly important as artificial intelligence becomes more autonomous, but they provide limited insight into how the system behaves toward a person.
Consider an AI customer-service assistant interacting with someone who has received an unexpected charge. The assistant may have access to the correct account information and remain entirely within its technical permissions. Yet the interaction can still deteriorate if it repeatedly misunderstands the customer’s concern, offers inappropriate reassurance or continues following a script as frustration increases.
The operational workflow may be functioning while the human experience is failing.
That distinction becomes important in healthcare, financial services, education and other environments where trust is influenced by how a person is treated throughout an interaction. It also becomes important as AI systems gain personalities, voices and other humanlike characteristics that encourage people to communicate with them socially.
An AI can sound attentive without correctly interpreting what is happening. It can appear confident while misunderstanding the user’s intention. And because large language models are probabilistic, an organization’s standards for handling these situations cannot depend entirely on the model generating the appropriate interpretation every time.
The system needs a way to govern the interaction independently of the intelligence generating the conversation.
Why VERN Built an Independent Emotional Model
This is the problem that led us to develop VERN’s Emotion Recognition System.
Rather than asking a large language model to infer emotional context and then determine how it should behave based on that inference, VERN uses an independent, deterministic neurolinguistic model to identify emotional signals expressed in language. The same analytical framework is applied consistently, providing signals that VERN OS can incorporate into deterministic behavioral governance.
There is a significant architectural distinction here. Psychology does not have one universally accepted theoretical definition or taxonomy of emotion, and language cannot provide direct access to someone’s internal psychological condition. VERN therefore evaluates expressed emotional signals within a defined neurolinguistic framework rather than treating the model’s interpretation as a diagnosis of what someone feels.
Those signals can inform organizational requirements governing escalation, role containment, emotional calibration and other aspects of AI behavior. An enterprise can define which conditions require intervention, how the AI should respond when relevant signals change and when the interaction should be transferred to a human.
The value of determinism is repeatability. Given the same input and configuration, the emotional model applies the same analytical rules, while the associated governance layer can enforce the behavioral requirements established by the organization.
Deterministic processing does not eliminate classification error, nor does emotional recognition establish whether someone should trust an AI system. Both the recognition model and the resulting controls require independent evaluation. But separating these functions from the underlying LLM provides organizations with a more stable, testable basis for governing human interactions.
That is an important improvement over leaving every interpretation and behavioral decision inside the same probabilistic conversation.
One Governance Architecture Across Agents and People
As AI systems become more autonomous, the human interaction will increasingly be one stage in a larger workflow. A customer-facing agent may consult financial systems, communicate with other agents, prepare an action and return to the customer with a recommendation.
Each stage introduces different governance requirements. Agent-to-agent interactions require controls over delegation, authority, role containment and consequential actions. Human-facing interactions require those same controls, along with consideration of how the person is experiencing the exchange.
VERN OS is designed to govern both through a common architecture.
The organization’s requirements can remain consistent as a task moves between agents and ultimately reaches a human. VERN’s deterministic emotional model provides additional signals when the interaction involves a person, while the governance layer maintains authority over the permitted behaviors and actions throughout the workflow.
This becomes particularly valuable in enterprises using multiple foundation models. Different models may interpret the same conversation or instruction differently, but the organization’s behavioral standards do not need to change every time the intelligence underneath them changes.
That architecture offers a way to combine the flexibility of probabilistic AI with independently enforceable requirements governing the experience.
Earning Trust Requires More Than Emotional Intelligence
The Harvard study should also be understood as a reminder that trust cannot be engineered through conversational behavior alone.
Women reported disparities in training, workplace support, representation and involvement in AI development. Those issues require organizational and institutional changes that no emotional recognition system can resolve by itself. Companies need to involve more diverse perspectives in designing their products, provide equitable access to education and evaluate whether their systems work appropriately for different users and contexts.
Behavioral governance addresses another part of the problem: Whether the resulting AI systems operate
