Users Pay $11 a Month to Keep Humans Out of Their AI Conversations

A new working paper from Profound researcher Jennifer Zou offers the most granular picture yet of how US adults actually use AI assistants — which platform they choose, what they use it for, how much they trust it, and what they would pay to control their data. The study surveyed 1,999 US adults in June 2026, weighted to the AI-using population rather than the general adult population, and its findings challenge several assumptions that underpin how brands and marketers think about the AI assistant landscape.

<h2>The market is concentrated but not uniform</h2>

ChatGPT is the primary assistant for 58% of users and Gemini for 25%, making the two platforms dominant in aggregate. But the aggregate view obscures a more interesting reality at the task level. Claude holds 33% primary share within coding tasks — against just 7% overall — nearly matching ChatGPT (39%) and far exceeding Gemini (15%). Copilot roughly doubles its share in work tasks compared to personal ones. ChatGPT and Gemini are generalists tilted toward informational and everyday use; the smaller platforms have carved out defensible niches that their overall share figures do not reveal.

The implication for brands optimising for AI visibility is significant. A brand that tracks only ChatGPT and Gemini is measuring the majority of the market but missing the task-specific contexts where other platforms dominate. A developer tool brand, for instance, may find that Claude is the more consequential platform for its category — even though Claude's headline share is a fraction of ChatGPT's.

<h2>Trust is earned, not inherited</h2>

The study's most striking finding on trust concerns the gap between reputational and experiential trust. ChatGPT and Gemini are already trusted by people who have never used them — 54% and 60% of aware non-users rank them among their three most trusted platforms, riding the brand recognition of OpenAI and Google. Claude is different. Only 41% of those aware of it but not using it rank it top-three, but that figure rises to 76% among its actual users: a 35-point experiential lift, nearly 1.5 times ChatGPT's 24-point lift and the largest of any platform measured.

In head-to-head comparisons restricted to users who have experience with both platforms, Claude is ranked above ChatGPT (59–41) and above Gemini (66–34). ChatGPT is ranked above Gemini (60–40). The ordering is consistent and statistically clean. The competitive implication is that Claude's reputation understates how its actual users regard it, while the incumbents' reputations roughly match or exceed their users' experience. For brands, this matters because the platform a user trusts most is likely the one whose recommendations carry the most weight.

<h2>The privacy paradox is an information problem</h2>

More than 80% of users report concern about how their conversation data is used, yet roughly 60% do not know whether their assistant trains on their conversations, and only 18% have ever paid for a plan with better privacy protection. The dominant predictor of protective behaviour is not concern but policy literacy: knowing whether one's assistant trains on conversations is associated with 0.34 additional protective tools adopted — the largest effect in the regression model and larger than a full point of concern on the five-point scale. Concern matters, but its coefficient is roughly a third the size of literacy's.

The study reframes the privacy paradox as an information problem rather than a preference problem. Users are not indifferent to privacy; they are uninformed about the specific practices that would prompt action. Disclosure and defaults, the paper argues, are what would move protective behaviour — not campaigns designed to raise concern.

<h2>What users actually pay to protect</h2>

The discrete-choice experiment asked users to choose between hypothetical AI plans that varied on monthly price and three data-handling attributes: human review of conversations, use of conversations for model training, and sponsored answers. The results reveal a steep hierarchy. Users pay most to avoid human review of their conversations — $11.20 per month — nearly four times what they pay to avoid training data use ($2.97), with avoiding sponsored answers in between ($6.46). Valuations rise with task sensitivity: for highly sensitive tasks, the willingness to pay to avoid human review reaches $15.24 per month.

The finding inverts the public debate. Training data use dominates regulatory and media attention, but it is the feature users value least of the three. The prospect of a person — an employer, a contractor, a government requester — reading their words is what users are actually willing to pay to prevent. For AI platforms competing on privacy, the paper suggests that human-access guarantees may be a more commercially potent differentiator than training-policy commitments.

<h2>AEO Updates Takeaway</h2>

For brands and marketers thinking about AI visibility, this study offers three practical reframings. First, platform share at the aggregate level is a poor guide to platform importance at the task level — category-specific measurement matters. Second, the platform a user trusts most is not necessarily the one with the highest headline share, and trust is built through use rather than reputation, which means challenger platforms can earn disproportionate influence in specific categories. Third, the privacy conversation that AI platforms are having with regulators is not the same conversation their users are having with themselves — and brands that understand that gap will be better positioned to navigate the trust dynamics of AI-mediated recommendation.

<h3>References</h3>

[1] Zou, J. (July 2026). Platform Choice, Privacy, and Task Allocation in the Consumer AI Assistant Market. Profound.