Platform Choice, Trust, and Privacy in the Consumer AI Assistant Market

Beyond the Generalist: Task Niches and the Knowledge Gap in AI

A study of nearly 2,000 AI users shows that while ChatGPT and Gemini lead the market, different tools serve specific jobs like coding. People value privacy, but they mostly take action when they understand how their data is used. Interestingly, users are willing to pay much more to keep humans—not models—out of their conversations than to stop their data from being used for training.

As conversational AI becomes integrated into daily life, users route everything from factual queries to sensitive personal deliberations through a few platforms. While companies often rely on proprietary telemetry (internal data logs) to track usage, these logs are blind to why a user chooses one tool over another. They also do not observe privacy attitudes or the reasons behind task-level choices. This leaves a critical gap in our understanding of the consumer market. Is the dominance of big players due to sheer reach, or is there a deeper structural lock-in happening at the task level?

The blind spots of platform telemetry

Current market analysis often treats AI adoption as a monolith. This approach measures whether a person uses "AI" or not. However, this fails to capture how users actually distribute their workload. A user might rely on ChatGPT for general information but switch to a different model for complex programming.

Existing datasets typically focus on broad adoption trends. They miss "task signatures"—the specific patterns of how different tools are utilized for different purposes. Without this granularity, it is impossible to tell if a platform wins because it is a great generalist or because it has captured a specialized niche. Furthermore, traditional surveys often face the "privacy paradox." This is the gap between stated privacy concern and actual protective behavior.

Mapping the AI-user landscape

To resolve these unknowns, the authors surveyed 1,999 US adults. They specifically targeted the AI-using population. To ensure the results reflected reality, they used post-stratification weighting. This is a statistical technique that adjusts the sample to match known population characteristics like age and education. This process is analogous to adjusting a political poll so the sample's demographic mix matches the actual voting population.

The researchers employed a multi-layered approach to decode user behavior: 1. Task-level allocation: Instead of asking for a single "favorite" platform, the authors asked respondents to identify a primary platform for six distinct task categories. 2. Head-to-head trust rankings: To bypass vague brand perceptions, the study required respondents to rank platforms against each other. This provided a clearer picture of relative reliability. 3. Discrete-choice experiment: To put a price tag on privacy, the authors presented users with hypothetical subscription plans. These plans varied in price and data-handling attributes, such as whether humans review conversations.

Specialized niches and experiential trust

The study finds a market that is highly concentrated but internally segmented. While ChatGPT holds a 58.2% primary share and Gemini holds 25.4%, smaller players have carved out significant territory. The authors report that Claude captures 33% of coding tasks, despite having only a 7% overall primary share .

Similarly, Copilot shows a clear "work tilt." Its share doubles in professional tasks compared to personal ones .

The data also reveals a distinction in how trust is built. For incumbents like ChatGPT and Gemini, trust is largely reputational. Even people who have never used them rank them highly .

However, for challengers like Claude, trust is experiential (earned through direct use). The authors find a 35-point "experiential lift" in trust among Claude users compared to non-users . In head-to-head matchups, Claude outperformed both ChatGPT and Gemini in perceived trustworthiness [Table 3].

Finally, the study quantifies the "privacy-action gap." While over 80% of users express concern about data use, protective behavior is driven more by policy literacy. Literacy refers to knowing whether a model trains on your data. This knowledge is a stronger predictor of action than the level of concern itself [Table 4].

Limits of the current model

The study has notable boundaries. The findings describe the AI-using public specifically. Therefore, the results cannot be generalized to the entire US adult population. This is because AI adoption is heavily skewed toward certain ages and education levels.

Additionally, the privacy valuations come from a stated-preference experiment. This means the dollar amounts represent what people say they would pay in a controlled setting. This may differ from real-world behavior in a market with fluctuating costs. For example, users value avoiding human review at $11.20/month [Table 5]. This represents a theoretical willingness to pay. Lastly, because smaller platforms have relatively thin user bases, profiling those specific segments with high precision remains difficult.

The verdict: Watch the tasks, not the totals

Looking at aggregate market share is insufficient. The evidence suggests the market organizes around "task signatures" rather than universal dominance. A platform can be a minor player in total users but a leader in a high-value niche like coding.

For developers and providers, the takeaway is twofold. Competition will be won in specific functional domains. Also, privacy marketing should focus on human-access guarantees. Users value these significantly more than training policies .

The market is not just growing; it is specializing.

Figures from the paper