About 60% of US adults are not confident that AI developers will ensure the technology is used responsibly. That is the headline finding of a Pew Research Center survey of more than 5,000 adults, conducted in February 2026 and published on June 17. It sits beside an estimated 1.5 billion people using generative AI and large language models every week, for everything from trivial daily tasks to sensitive personal questions. The two figures are the interesting part: distrust of the companies is not showing up as non-use of the products.
The argument built around those numbers separates two kinds of trust that usually get collapsed into one. The first comes from use — you work with a system, you watch what it does, you decide how far it can be relied on. The second is inherited from the company behind it. If a developer looks careless about how it builds or controls its systems, users arrive already wary, and a trust gap opens before the first interaction. If the developer looks responsible, the halo works in reverse: people start using the system more readily and assume it can be relied on.
That halo is thin. A company with a spotless reputation loses it fast once its AI performs badly or produces dangerous output. Reputation buys the first session, not the tenth.
Cross the two kinds of trust and you get four combinations. Two are unremarkable: trust the developer and trust the AI, or distrust both. The other two are where the real behavior lives — trusting the company while doubting the specific system, and distrusting the company while finding its AI useful and reliable.
The essay reaches for ordinary institutional analogies, and they hold up. You can distrust a bank and still think well of the teller you deal with, because you judge the person separately from the institution. Patients trust their doctor while distrusting the health system that employs them. The reverse happens too: an impeccable bank can employ someone who mishandles your money, at which point your assessment of that individual outweighs the institution's reputation. People do the same with AI, separating institutional trust in the developer from trust in what the model can actually do, and deciding largely on what the system says and does. The company's reputation stays in the calculation without necessarily dominating it.
Trust also moves. The essay runs a scenario: Jane uses an AI from a company called XYZ, arrives with high trust because XYZ is well known and seen as building responsibly, and gets useful and accurate answers. Her trust in the system rises, and her trust in XYZ rises with it — the product's competence is read as evidence the reputation was earned. Then she asks for personal advice and gets an alarming answer, one proposing actions that could cause serious harm. Trust in the system collapses, and most likely takes some of her trust in XYZ with it, because people treat product and maker as connected. The transfer runs both ways: news that a company's system went off the rails can change how users feel about the model even if they have personally never seen it misbehave.
There is a second argument running underneath, and it is the one that will age better. The standard debate asks whether a machine will become conscious. The less asked question is what sustained interaction with these systems does to human minds. Interaction is two-way. People who talk to AI constantly change how they think about the world and how they think conversation with other people should go, and the scale of that shift has not been properly studied. No one needs to wait for conscious AI to research it; minds are already adapting to systems that have no consciousness at all.
My assessment is that the four-box framework is coarse, and the essay says so itself — trust moves gradually along a wide spectrum, and high-versus-low is a blunt instrument. But the framework earns its place by locating where the market actually is. If 1.5 billion weekly users coexist with 60% institutional distrust, then the dominant quadrant is low trust in the developer and high enough trust in the product to keep using it. That has a consequence the essay states without pressing: reputation is a weak lever on this industry. The usual corrective — customers leaving because they do not trust the maker — is not operating. What moves users is the behavior of the system in front of them, which means the only discipline that reaches an AI company is a product failing visibly, individually, to someone who then tells people.
Which makes the piece's one concrete harm the load-bearing case, and the place where its own logic runs out. Overtrust, it warns, is especially dangerous around mental health, where a user can take a system's output for therapeutic advice and end up in a worse position than they started. But the entire model of trust here depends on users calibrating from experience — trying the system, watching it, adjusting. Someone in distress taking harmful advice as counsel is precisely the user who cannot run that loop, and the essay does not say who runs it instead. Developers are told they need both kinds of trust, institutional trust to operate as a business and capability trust so people want to use the product, and that misunderstanding where trust comes from will cost them reputation. That is advice about the company's exposure, not the user's.
The author closes by citing the entrepreneur R. M. Williams, who held that trust is the easiest thing to lose and the hardest to rebuild — someone disappointed by a company or a system will not quickly decide to trust either again. It is a good line with a condition attached: it only bites if people actually withdraw. On the evidence of the Pew number against the usage number, this industry has already learned it can grow without being trusted, and is under no visible pressure to find out what withdrawal would feel like.