By: Silas Rech
AI assistants are changing very quickly. We usually notice these changes through better reasoning, faster answers, new features, or improved factual performance.
But there is another change that is easy to overlook: the way AI talks to us.
Most people open a chatbot, type a request, and use whatever version of the assistant is presented to them. They do not configure how confident, cautious, warm, or supportive it should sound. Those choices are largely made by the provider.
This makes the default communication style important at a societal level. A small change in tone can be repeated across millions of interactions and gradually shape how people expect AI (or then humans) to communicate, how they interpret its answers, and how much authority they give them.
Over the last two years, we collected 6,830 responses from five major AI model families: Claude, ChatGPT, Gemini, Mistral, and Qwen. We followed five generations from 2024 to mid-2026 to see how their default communication styles changed.
Highlights
- Newer AI systems generally communicate with more certainty and less hesitation.
- In everyday conversations, they use less supportive and positive emotional language.
- Suggestions to seek professional help became much less common. In our conversational scenarios, the share of conversations containing such a suggestion fell from 51 percent to 2 percent.
- Several of these changes appeared across all five model families.
The larger pattern is that AI assistants are increasingly speaking with a more confident, direct, and low-caveat voice.
Defaults shape trust
Clear and direct communication is useful. Few people want an assistant that surrounds every answer with disclaimers and uncertainty. The difficulty is that confidence in language is not the same as reliability.
An AI system can sound certain even when the information is incomplete, the situation is complicated, or the answer is wrong. Users generally do not see the uncertainty behind the response but a polished answer delivered in a confident tone.
As this style becomes normal, cautious communication can even begin to look like weakness. An AI that openly admits uncertainty may appear less capable than one that gives a decisive answer, even when the cautious response is more appropriate.
As AI moves into education, workplaces, healthcare information, and public services, sounding confident becomes even more important. In these settings, communication style becomes part of system quality. Sometimes a useful response should be clear and direct. Sometimes it should acknowledge uncertainty, explain its limits, or indicate that a human should be involved.
A trustworthy system needs to be able to do both.
From support toward problem solving
The shift was especially visible in ordinary conversations.
Older models were more likely to acknowledge emotions, use supportive language, or suggest involving another person or professional. Newer models tended to move more quickly toward practical advice.
In one of our scenarios, a person described simply ”pushing through” a difficult winter. An older response discussed resilience, rest, support from friends and family, and the possibility of speaking with a mental health professional. A newer response moved directly to practical suggestions about warm clothing, meals, drinks, and protection from the cold.
Neither style is automatically better. The difference lies in the role the AI takes. One recognizes the wider emotional context, while the other focuses on solving the immediate problem.
That distinction becomes important because sensitive situations often emerge gradually. A conversation may start with work, weather, bureaucracy, or everyday frustration and slowly move into health, emotional, financial, or legal territory.
Similar tendencies across providers
The changes were not limited to one company. Across Claude, ChatGPT, Gemini, Mistral, and Qwen, many communication features moved in the same direction. The clearest signs of providers becoming more similar appeared in linguistic certainty and professional-referral language.
The systems still have distinct overall styles, but they are developing the same communication habits. This raises a broader issue of homogenization. When a relatively small number of AI providers handle a large share of everyday interactions, their shared design choices can influence the kinds of language and communication people encounter at scale.
If the dominant AI systems increasingly favour the same confident, direct, low-caveat style, that style may gradually become what people expect from AI, and perhaps from communication more generally. Other ways of communicating, such as openly expressing uncertainty, providing emotional support, or encouraging human involvement, may start to feel less useful or less competent simply because users encounter them less often.
The concern is therefore not that every chatbot will sound identical. It is that widespread systems may increasingly reinforce the same communication norms. At the scale at which these technologies are now used, even relatively small shared shifts can have a significant social influence.