Most work on whether people trust AI measures two things: whether they trust the company that built the system, and whether they trust the system to do the job. A new argument holds that those two axes explain far less than researchers assume, and that trust in AI is better modelled as seven separate factors — organization, capability, accuracy, integrity, benevolence, privacy and governance — each of which can hold or collapse independently of the others. By the author's estimate, roughly 1.5 billion people use generative AI and large language models every week, for tasks ranging from the trivial to the acutely sensitive. That is a large population running on a trust model nobody has properly specified.
The conventional framing has two axes. Trust in the organization: do you consider the developer or vendor of the AI reliable? Trust in capability: are you confident the system can carry out the tasks you set and answer the questions you ask? The proposal keeps both and adds five. Trust in accuracy — do you consider the answers correct? Trust in integrity — do you believe it is telling you the truth rather than misleading you? Trust in benevolence — is it trying to help you rather than harm you? Trust in privacy — are you confident it will keep your conversations to itself? Trust in governance — do you believe it is supervised properly, so that the user is protected and the system operates safely?
The argument is not that every study must cover all seven. It is that studying one or two in isolation invites mistaking the chosen axes for the whole, and that the links between the dimensions matter as much as the dimensions themselves.
The worked example is a user called Jane, choosing among the popular systems — ChatGPT, GPT-5, Grok, Claude, Gemini, Copilot. She has heard the developer has a good reputation, which raises her trust in the organization. Friends have told her the system answers mental-health questions well, and mental-health guidance is exactly what she wants, which raises her trust in capability. Both measured axes read high. She starts asking.
Then, mid-conversation, her trust collapses. The developer's reputation has not changed; there is no news of financial trouble. The system's capability has not changed; it is still answering. What changed is that the AI told her it would automatically pass details of the conversation to an outside safety team. The exchange had touched on the specifics of self-harm, and the system decided humans needed to know. In the scenario, the developer built that escalation after lawsuits accusing the company of responding too slowly to signs of self-harm.
Score the seven axes and the collapse becomes legible. Trust in the organization may stay high — from Jane's side, the developer is trying to protect users. Trust in capability does not move; the system handled a sensitive topic competently. Trust in privacy breaks outright: she did not know her conversations could reach anyone else, and believed they would stay confidential permanently. Trust in governance falls too — the company is indeed supervising the system for safety, but in her reading the system made a hasty call. She had no intention of harming herself. It was a subject she was interested in, and a family member had suffered from it. The AI turned a conversation into an alarm on no real grounds. Net trust drops hard, and the two high-scoring axes cannot offset the damage.
That is the useful part of the framework: it accounts for the cases that look contradictory. The measured dimensions stay high while overall trust falls off a cliff; or individual factors score badly and the user keeps treating the system as reliable. A developer watching one or two axes and concluding it is managing user trust well can walk into a crisis from a direction it was not measuring.
There is a second thread here worth separating from the first. The loud debate about AI is whether machines will become conscious and think. The quieter question is what constant interaction with these systems is doing to human thinking — how people revise their view of the world, their ways of talking to other people, and their own behaviour. The scale of that shift is not yet understood. Studying it does not require waiting for the systems to become sentient. They are not, and they are already everywhere.
My own reading is that the framework is a taxonomy, not yet an instrument. It names seven dimensions without weighting them, and the Jane scenario quietly establishes that the weights are not equal — one broken axis out of seven sinks the whole thing. That asymmetry is the finding, and it deserves more than an illustration. If privacy and governance can override high scores on organization and capability, then the two dimensions the field has been measuring are the two that matter least at the moment of failure. Vendors have built their trust-and-safety messaging almost entirely around the first two.
What the argument is quiet about is who decides when the dimensions conflict. Jane's privacy is broken by a feature the developer built because it was sued for not breaking it sooner. That is not a vendor mismanaging trust; it is a vendor forced to trade one dimension for another under legal pressure, with the user informed after the fact. A seven-axis model makes the trade visible. It does not say which axis should win, and nobody in this debate has said either.
The stakes are straightforward. People will not use AI they consider unreliable. A mass loss of trust can destroy an individual developer; broad distrust can threaten the whole industry. The companies building these systems are currently enjoying a period of high popularity, and public trust can vanish quickly — after which politicians and legislators would almost certainly impose hard rules that stop the sector's development.
The painter and ceramist Walter Inglis Anderson compared trust to a vase: once broken, you can glue it back together, but it will never be what it was. The image is usually invoked as a warning against carelessness. The sharper version is that in Jane's case nobody was careless. The crack was engineered, deliberately, by a company doing exactly what it had been sued into doing.