DuckDuckGo published survey results on August 25, 2026 showing that 56% of self-described AI enthusiasts and 32% of all AI users in a 1,944-person US sample have told a chatbot something they withheld from people close to them, while a majority remain unaware that most services train on those conversations by default.

What the survey measured

The Pennsylvania-based search company commissioned and ran the study itself. Responses came from 1,944 United States adults aged 18 and over, collected between June 17 and June 27, 2026 through an online sample sourced from the PureSpectrum panel, according to DuckDuckGo. Quotas were applied to balance the sample against US Census demographics for age, gender and region. Results were reported unweighted.

The stated margin of error for the full sample is approximately plus or minus 2.2 percentage points at the 95% confidence level, widening for subgroups. DuckDuckGo puts the figure at roughly plus or minus 6 points for adults aged 18 to 24. Figures drawn from open-ended questions, such as the single biggest AI privacy concern named by each respondent, rest only on those who supplied an answer. Percentages may not total 100 because of rounding and multiple-response questions.

That methodology note is unusually detailed for a vendor-commissioned consumer study, and it matters here because several of the headline numbers describe subgroups rather than the full sample.

Disclosure behaviour

The central finding concerns what people tell chatbots that they do not tell anyone else.

Among AI users in the sample, 32% said they had told a chatbot something they had withheld from a close friend, parent, colleague, doctor or therapist. Among respondents who described themselves as AI enthusiasts and use chatbots, that figure reached 56%.

The parental split is sharper still. Of parents or guardians with children in the household who use AI, 43% reported telling an AI system something they had never told a doctor, therapist, close friend, parent or child. For AI users with no children at home, the comparable figure is 25%.

One inconsistency in the published material is worth flagging. The summary section of the report describes the 43% parent figure as covering information never shared "with another person," while the body section defines it against a specific list of confidants: doctor, therapist, close friend, parent or child. Those are not the same claim. The narrower body formulation is the one the questionnaire appears to have measured, and it is the safer reading of the number. A similar looseness appears around the 32% figure, which the summary section frames as information kept from "close friends, parents, or medical professionals" while the body section adds colleagues to the list.

DuckDuckGo's own framing for why the gap exists is behavioural rather than technical. A typical web search runs to a few words; a chat session invites longer and more personal input. Search queries expose interests. Conversations, according to the company, carry the potential to expose thought processes and communication styles as well.

The knowledge gap

The survey then tested awareness of six specific facts about how conventional AI chat services store, review and disclose conversations. The six were: that chats can be subpoenaed by courts or law enforcement; that chats can be made public in a lawsuit; that employers can access work or enterprise AI accounts; that settings governing deletion and storage vary by provider; that free accounts often carry weaker privacy protections than paid accounts; and that human employees can review conversations. DuckDuckGo cited published reporting from outlets including CNN, TechCrunch, Computer World, CNBC, the Washington Post, Axios and Fortune as examples for each.

Awareness was low across the board. According to the survey, 53% of respondents did not know, or were not sure, that their conversations are used for AI training. Only 25% knew that AI chats can be subpoenaed by the government. Across AI users and non-users alike, 43% did not know any of the six facts. Among self-described active AI users, roughly a third fell into that same category.

The company frames this as an industry condition rather than individual failure, drawing a parallel with other forms of online tracking where the mechanics are invisible at the point of use.

What happens once people are told

The more consequential half of the finding is what the numbers do after disclosure.

Presented with how most AI services handle conversations, 58% of respondents said they were uncomfortable with their conversations being used to train AI. Of the full sample, 30% specified feeling "very uncomfortable."

Trust levels sit lower still. Only 14% of respondents said they "mostly" or "completely" trust large AI companies to protect their data. A further 39% reported no trust at all. The United States government scored worse on the same question, with 48% reporting "no trust at all."

Read together, the two sets of numbers describe a market where the product is used widely, understood poorly, and disliked once understood. That combination has appeared repeatedly in survey work over the past eighteen months. Usercentrics research published on July 1, 2025 found 59% of 10,000 consumers across Europe and the United States uncomfortable with data used to train AI. A Verve survey released in December 2025 put the figure at 65%, up from 45% two years earlier, with 97% of respondents wanting greater transparency from platforms. A Cloaked survey published on April 2, 2026 found only 18% of 1,009 Americans trusting AI to keep personal data secure. A Shift Browser survey released on March 3, 2026 recorded 81% concerned about AI accessing personal data while daily use kept climbing.

DuckDuckGo's 14% trust figure is the lowest of that group, though question wording differs enough across the studies that direct comparison is unsafe. What is consistent is direction.

The commercial context the survey does not name

The report does not discuss advertising directly, but the disclosure behaviour it documents sits on top of an ad market that has spent two years learning how to use conversational data.

Meta confirmed on October 25, 2025 that interactions with Meta AI, including voice and text exchanges, would feed content and advertising personalisation from December 16, 2025, across Facebook, Instagram and other apps, with no opt-out. WhatsApp conversations were excluded unless users linked those accounts in Accounts Center.

OpenAI took a different route. The company confirmed advertising tests on January 16, 2026 for logged-in adults in the United States on free and Go tiers, publishing five principles that included a commitment that conversations stay private from advertisers and that data is never sold to them. The pilot went live on February 9, 2026 at a $60 CPM. An April 30, 2026 privacy policy update formalised advertiser data-sharing in binding language while keeping both commitments intact in letter, with advertisers receiving aggregate view and click counts rather than conversation content. StackAdapt's entry into the pilot framed the format as closer to contextual advertising than to audience-based targeting.

Whether users can distinguish between those two models is precisely what the DuckDuckGo numbers put in doubt. A person who does not know that conversations are used for training by default is unlikely to be tracking which provider promises what.

Third-party measurement supports the concern. A University of California, Davis study covered on May 16, 2026examined network traffic across 20 chatbot services and found conversation text transmitted in plaintext to Microsoft Clarity by three services through session replay instrumentation, alongside advertising connections to as many as 13 distinct networks in a single session on one platform.

Litigation and the disclosure surface

The subpoena awareness figure, at 25%, is the one with the most direct precedent behind it.

A federal magistrate judge ordered OpenAI to produce 20 million de-identified consumer ChatGPT conversation logs to news plaintiffs in November 2025, and a stay request was denied on November 13 of that year. The logs covered a random sample of complete conversations from December 2022 to November 2024, excluding Enterprise, Edu, Business and API customers, and each log contained a full exchange of multiple prompt and output pairs.

Regulatory attention has followed a parallel track. The Federal Trade Commission issued orders to seven companies on September 10, 2025 under Section 6(b) of the FTC Act, requiring detailed reports on safety practices, data handling and monetisation at consumer-facing AI companion products. That inquiry sought engagement metrics segmented by age group and documentation of age-gating practices.

Product liability actions have reached the same material from another direction. A wrongful death complaint filed on June 11, 2026 in San Francisco County Superior Court drew on chat logs as evidence, alleging that memory features and anthropomorphic presentation constituted product defects.

Each of those routes converts stored conversations into discoverable records. None of them is visible to a user typing into a chat box.

No opt-out signal exists for chat

Web tracking has spent fifteen years building user-side controls. Conversational AI has none of comparable standing.

The failed attempt was Do Not Track, a browser header whose working group never agreed a definition of tracking and whose compliance was voluntary. Its successor, GPC, carries statutory backing in a growing list of US jurisdictions and travels as a one-bit HTTP header instructing sites not to sell or share personal information. Firefox, Brave and DuckDuckGo send it. Chrome, Safari and Edge do not support it natively.

Neither mechanism reaches chat content. GPC governs sale and sharing for cross-context advertising; it says nothing about whether a conversation enters a training corpus. On the publisher side, controls do exist: a site can block GPTBot in robots.txt to keep pages out of training data, though the block is prospective only and content still reaches training sets through intermediaries. No equivalent exists for the person typing the prompt. The available control is a per-provider setting, buried at different depths in different products, which is itself one of the six facts 43% of the sample did not know.

The vendor behind the numbers

DuckDuckGo is not a neutral party to this research, and the report says so plainly by ending with a product pitch.

The company launched Duck.ai on June 9, 2024, an anonymous chat interface that acts as an intermediary, stripping the user IP address and substituting DuckDuckGo's own before requests reach underlying model providers. The UC Davis study noted that Duck.ai goes further than most services on this point. The report closes by directing readers to that product and its privacy policy.

The company has been active on adjacent fronts through 2026. It began blocking YouTube ads by default on iPhone, Windows and Mac in July 2026. On August 5, 2026, according to material published alongside the survey, it filed an amicus brief in the appeal of the federal decision that Google unlawfully maintained a monopoly in general search, arguing the company would repeat the default-distribution pattern in AI.

None of that invalidates the data. It does mean the framing was chosen by a competitor to the companies the survey asks about, and the questions were not written by a disinterested party.

Why this matters for marketing

Three implications follow for practitioners.

The first concerns signal quality. Conversational input is being sold on the premise that it carries richer intent than search queries, and the DuckDuckGo data supports that premise more strongly than any vendor deck. People who disclose to a chatbot what they hide from a doctor are producing exactly the high-value, high-sensitivity signal that behavioural targeting has always wanted. The 43% parent figure identifies the household segment producing the most of it.

The second concerns fragility. Signal built on a disclosure the person did not understand they were making is not durable. Once informed, 58% of the sample objected. The commercial cost of that gap has been measured elsewhere: Usercentrics research published in June 2026 across seven markets found one in four consumers had cancelled a purchase over AI data practices, and 52% willing to pay a 7% premium for brands that address the concern.

The third concerns positioning. Trust in AI companies sits at 14% in this sample, below the 18% Cloaked recorded four months earlier and below the 16% who told Shift Browser they trust AI answer engines "a great deal." Brands buying placement inside conversational surfaces are buying adjacency to a category the public does not trust, and the aggregate-reporting model most AI ad platforms have adopted means they cannot easily demonstrate otherwise. The argument that AI firms benefit commercially when users treat chatbots as confidants has been made directly, most recently by technology writer Kate O'Neill in May 2026.

The survey's own closing observation is the one with the sharpest commercial edge: AI chat remains early enough that most people have not formed habits with it. Whether the disclosure gap closes before those habits set is the open question, and the answer determines whether conversational data becomes a durable advertising asset or a regulatory liability.

Timeline

Summary

Who: DuckDuckGo, the Pennsylvania-based privacy search and browser company founded in 2008, which commissioned and conducted the research itself. Respondents were 1,944 United States adults aged 18 and over drawn from the PureSpectrum online panel.

What: A survey titled The AI Privacy Problem, finding that 32% of AI users and 56% of self-described AI enthusiasts have told a chatbot something withheld from a close friend, parent, colleague, doctor or therapist; that 43% of parents who use AI report the same against 25% of AI users without children at home; that 53% did not know or were unsure that conversations are used for AI training; that only 25% knew chats can be subpoenaed; that 43% knew none of six basic facts about chat storage, review and disclosure; that 58% are uncomfortable with training use once informed, 30% "very uncomfortable"; and that 14% "mostly" or "completely" trust large AI companies with their data while 39% report no trust at all, against 48% reporting no trust at all in the US government.

When: Fieldwork ran from June 17 to June 27, 2026. The results were published on August 25, 2026.

Where: The sample covers United States adults, balanced by quota against Census demographics for age, gender and region, with results reported unweighted. The margin of error is approximately plus or minus 2.2 percentage points at 95% confidence for the full sample and around plus or minus 6 points for the 18 to 24 subgroup.

Why: The findings quantify a gap between how personally people use conversational AI and how little they understand about the retention, review, training and legal disclosure of those conversations. That gap sits directly beneath an advertising market now buying against conversational intent, following Meta's December 2025 activation of AI chat data for ad targeting and OpenAI's advertising pilot launched in February 2026. DuckDuckGo is a commercial participant in the same market through Duck.ai, which the report promotes, and the framing of the questions carries that commercial interest.