The Trust Thermometer

Trust in AI is usually measured the way temperature is measured. A scale. A number. A gradient that rises or falls in small, predictable increments. Edelman releases the survey. The dashboard turns green or red. The policy team writes a memo. The number is treated like a thermostat reading — something to nudge upward through campaigns, training, and reassurance.

The 2026 Edelman Trust Barometer broke that frame. Trust in AI is not a thermometer. It is a cliff.

Across nearly 34,000 respondents in 28 countries, surveyed between October 23 and November 18, 2025, Edelman found something that the gradient model could not explain. The single largest predictor of whether a person trusts AI is not their age, not their income, not their education, not their political orientation. It is whether they have personally had a useful experience with the tool. And the gap between those who have and those who have not runs from 26 to 46 percentage points across the markets surveyed.

Twenty-six to forty-six points is not a gradient. It is a discontinuity. The data does not show people warming slowly toward AI as they learn more about it. The data shows people on one side of a wall or the other. The wall is one meaningful use. Once a person has had it, trust rises sharply. Until they have, trust does not move — regardless of how many decks the change-management team produces.

The Cliff in the Numbers

The country breakdowns make the cliff visible at scale. In China, 87 percent of respondents say they trust AI. In Brazil, 67 percent. In Germany, 39 percent. In the United Kingdom, 36 percent. In the United States, 32 percent. The gap between China and the United States is fifty-five percentage points — a gap so wide that the two populations are effectively living in different relationships with the same technology.

The temptation is to read the country differences as cultural. Eastern markets trust technology more than Western markets. Younger populations trust more than older populations. The familiar explanations queue up.

The Edelman Trust Institute looked at the underlying mechanism and found something different. The fall 2025 Edelman Trust Barometer Flash Poll — a separate study of 5,000 respondents across Brazil, China, Germany, the United Kingdom, and the United States, fielded between October 17 and October 27, 2025 — isolated the experience variable. The question was simple: have you personally benefited from AI at work? In four of the five markets, respondents who said yes were 26 to 46 percentage points more likely to trust AI than respondents who said no.

The country differences in headline trust do not disappear once you control for experience. But they get smaller. Much smaller. The reason China leads the trust rankings is not principally that Chinese culture is more trusting of technology. It is that more Chinese workers have had a useful AI experience. The trust follows the experience. The experience does not follow the trust.

This is what makes the cliff metaphor more accurate than the thermometer one. A thermometer assumes continuity — small movements in the input produce small movements in the output. A cliff assumes discontinuity — the person on one side has a categorically different orientation than the person on the other side, and the only thing that moves them across is the crossing itself.

What the First Interaction Decides

Amy Edmondson published her foundational work on psychological safety in 1996 and 1999, studying nursing teams at a hospital. She expected to find that better-performing teams made fewer medication errors. She found the opposite: better teams reported more errors, not fewer. The teams that performed well were not error-free. They were error-honest. Psychological safety — which Edmondson defined as the belief that the team is safe for interpersonal risk-taking, the belief that you will not be punished for asking a question or admitting you do not know something — was the variable that distinguished the teams that could surface errors from the teams that buried them.

Apply the framework to AI adoption and the cliff comes into focus.

The first interaction with an AI tool is an interpersonal risk-taking moment. The person at the screen does not know if their question is too basic, too vague, or too revealing of what they do not know. They do not know if the system will produce something useful or something embarrassing. They do not know whether their query history will be reviewed by someone with authority over their performance evaluation. They are exposing a piece of cognitive uncertainty to a system whose social meaning has not yet been established.

If that first interaction produces a small, useful result — an answer that clarifies something, a draft that saves twenty minutes, a summary that surfaces a pattern they had missed — the uncertainty resolves. The tool moves from “potential threat to my professional standing” to “resource I can use again.” The trust does not arrive through persuasion. It arrives through the experience of a non-punishing return on a small act of vulnerability. This is Edmondson’s framework applied to tools instead of teams: the safety to take the risk produces the experience that justifies the trust.

If that first interaction produces nothing — or worse, produces something the person has to clean up, defend, or hide — the uncertainty does not resolve. It hardens. The risk was taken. The return was negative. The system has now been classified as unsafe. No amount of subsequent training will reset that classification quickly, because the human nervous system does not weight evidence symmetrically. A negative first experience produces a defensive posture that subsequent positive experiences must overcome, not simply add to.

The 26-to-46-point gap is the visible shadow of this asymmetry. The people on the trust side of the cliff are not more enlightened. They had a useful first interaction. The people on the resistance side are not more fearful. They did not.

The European Position

For European SMEs — the 50- to 500-person companies that make up the audience for almost everything Bluewaves writes about — the cliff has a specific geometry that the global averages obscure.

Eurostat released its 2025 ICT survey on December 11, 2025. The headline number: 20 percent of EU enterprises with ten or more employees used AI technologies in 2025. Twenty percent. Which means eighty percent did not. The growth from 2024 was 6.5 percentage points — a solid acceleration — but the starting point was 13.5 percent, and the destination is still a minority of European businesses.

The breakdown by enterprise size is more revealing. Large enterprises (250+ employees) reached 55 percent adoption. Medium enterprises (50–249 employees) reached 30 percent. Small enterprises (10–49 employees) reached 17 percent. The country range is wider still: Denmark at 42 percent, Finland at 37.8 percent, Sweden at 35 percent at the top; Romania at 5.2 percent, Poland at 8.4 percent, Bulgaria at 8.5 percent at the bottom.

The European SME landscape, in other words, is mostly pre-cliff. Most companies have not yet introduced AI tools to most of their employees. Most employees have not yet had a first interaction. Which means most of the trust formation that will determine whether AI takes root in European SMEs has not yet happened.

This is unusual. We are accustomed to thinking about adoption as something that happens after the fact — measured in retrospect, optimised through iteration. The European SME context inverts the timing. The first interactions are still in the future. The cliff has not yet been crossed by most of the population. The conditions of the crossing are still designable.

That is a rare position. It is also a fragile one. Because the design choices that determine whether the first interaction succeeds are being made right now, in procurement decisions, deployment templates, and onboarding emails — and most of those choices are being made by people who do not realise they are designing a cliff.

What Most Deployments Get Wrong

I have watched dozens of AI deployments in European companies over the past two years. The deployments that fail tend to fail in similar ways. The deployments that succeed tend to succeed for similar reasons. The pattern is not subtle once you know where to look.

The failing deployments treat the first interaction as a training problem. The thinking goes: if we train people well enough on the tool, they will use it well. So the deployment lead schedules a one-hour training session. A vendor representative walks through the features. People take notes. Some questions are asked. The session ends. People return to their desks, where the tool is now installed alongside seventeen other applications, and they are expected to integrate it into their work.

The first interaction, in this configuration, is typically a person sitting alone at their desk, opening the tool for the first time after the training, trying to remember what they were supposed to do. They type something tentative. The tool produces something generic. The person does not know if the output is good or bad — they cannot evaluate it, because they have not used the tool enough to develop calibration. They close the window. They go back to the workflow they already know. The first interaction is not a failure. It is a non-event. And a non-event is, in Edmondson’s framework, a confirmation that the risk was not worth taking. The trust does not form because nothing happened that would form it.

The succeeding deployments treat the first interaction as a relational moment. They engineer it. Not the tool — the moment. They identify a specific, small, real task the person already does. They sit with the person — or pair them with a colleague who has used the tool — for the first interaction. They watch the output together. They name what is useful and what is not. They make sure the person leaves the first interaction with a single concrete instance of “this tool helped me with something I had to do anyway.” That instance becomes the trust anchor. Subsequent interactions are no longer first interactions — they are continuations of an experience the person already classifies as positive.

The succeeding deployments understand something that the failing ones do not: the first interaction is not a step in a training curriculum. It is the moment the cliff is crossed or not. Everything else is downstream of that moment.

The Mass-Class Divide

The Edelman data surfaced another finding that European SMEs should read carefully. In the markets surveyed, nearly two-thirds of managers reported using AI tools regularly. Among non-managers, the figure was approximately one in four.

This is not a competence gap. It is a first-interaction gap. Managers have had more opportunities to engage with AI tools — they are present in vendor demos, included in procurement conversations, given early access during pilots, and sometimes assigned a personal license before broad rollout. They have had multiple first interactions, some of which produced useful results. They are on the trust side of the cliff.

Non-managers, in most of the deployments I have observed, get a different first interaction. They get an email announcing the tool. They get a link to a recorded training. They get a deadline by which they are supposed to be using it. They do not get a colleague sitting beside them for the first query. They do not get an explicit, small, real task to use the tool on. They get the configuration that produces non-events.

The trust gap between managers and non-managers in the Edelman data is not principally a gap in skill, intelligence, or openness to technology. It is a gap in the quality of first interactions. The managers had the relational moments. The non-managers got the email.

For an SME deploying AI, this means the question “is our team adopting the tool?” is the wrong question. The right question is: “what was each person’s first interaction with this tool, and was it good enough to put them on the trust side of the cliff?” If you cannot answer that question for each person, you do not have an adoption strategy. You have an adoption hope.

The Asymmetry of Negative First Impressions

There is a piece of stress physiology worth naming here. Robert Sapolsky, in Why Zebras Don’t Get Ulcers, documents the asymmetry of threat detection in mammalian nervous systems. A predator missed is a meal lost. A predator unnoticed is death. The nervous system is calibrated to weight negative signals more heavily than positive ones, because the cost of a false negative — missing an actual threat — is higher than the cost of a false positive — treating something safe as a threat.

This asymmetry travels into the workplace. A person who has had one negative experience with a tool will require multiple positive experiences before the negative classification is overwritten. The exact ratio varies — research on negativity bias has produced estimates ranging from three-to-one to five-to-one — but the direction is consistent. Negative first impressions are sticky. Positive first impressions are reinforceable. The substrate is not symmetrical.

This is why the Edelman gap does not close quickly through subsequent exposure. A person who had a negative first interaction with an AI tool is not simply someone who has not yet had a positive one. They are someone whose nervous system has already encoded the tool as a source of risk. The next exposure must work harder than the first one would have, because it is fighting an existing classification, not building one from scratch.

For deployment teams, the operational implication is severe. You cannot run a bad first interaction and then fix it with a better second interaction. The first interaction is not a draft. It is the foundation. The deployments I have seen succeed treat it with the seriousness that finding deserves.

What the Data Says to Do

I am cautious about prescriptive lists, because they tend to convert nuanced findings into checklists that produce the same compliance-driven non-adoption the findings were trying to prevent. But the Edelman data, combined with Edmondson’s framework and Sapolsky’s negativity asymmetry, points to a small set of structural choices that European SMEs should make before deploying their next AI tool.

Design the first interaction explicitly. Do not assume it will happen. Do not assume training will produce it. For each person on the team, identify a real, small task that the tool can demonstrably help with, and design the moment when the person uses the tool on that task. The design is not a checklist. It is a relational moment — a colleague present, a question expected, an outcome named.

Sequence by readiness, not by org chart. The first people to interact with the tool should be the people most likely to have a useful first interaction. This often means starting with curious users, not with senior leaders. The curious users will produce trust signals that the rest of the organisation can observe — which lowers the perceived risk of the first interaction for everyone who comes later.

Separate the rollout from the evaluation. During the first interaction window, no usage dashboards should be visible to managers. No metrics should be reported up the chain. The first interaction is a learning moment. Treating it as an evaluation moment closes the safety that the learning requires. Once trust has formed — once the team is on the right side of the cliff — usage data can become operationally useful. Before that, it is a threat that prevents the very behaviour it is trying to measure.

Build for the failed first interaction. Some first interactions will go badly. The tool will produce something useless or misleading. The deployment design should include an explicit, low-cost path for the person to flag the failure and try again with help. The failure is not the problem. The failure unsurfaced is the problem — because an unsurfaced failure hardens into the negative classification that Sapolsky’s research describes, and subsequent positive experiences will have to fight it for months.

Measure the experience, not the activity. The relevant question is not how many queries the team made this week. It is how many people, asked directly, would say the tool helped them do something they actually had to do. That measurement is harder. That is why most organisations skip it. The skipping is why most deployments produce activity without adoption.

The Integration

Here is the tension I want to hold, because collapsing it would falsify the data.

AI tools are useful. The Edelman respondents who reported regular AI use also reported sharp gains in confidence, productivity, and trust. When the tool helps people understand complex ideas, trust rises by roughly forty points across markets, and by nearly fifty points in the United Kingdom. The benefits are real. The data is not ambiguous about this.

AI deployments are mostly failing to produce those benefits. Eighty percent of European enterprises with ten or more employees are not using AI. Among those that are, the gap between managers and non-managers is structural, not coincidental. The benefits are unevenly distributed because the first interactions are unevenly designed. The tool works. The deployment, mostly, does not.

Both things are true. The technology has crossed a capability threshold. The organisations using it have not crossed a design threshold. The capability is uniform. The trust is bimodal.

The European SME context — where most companies are still pre-adoption, where most employees have not yet had their first interaction — is the rare moment when the design threshold is still in front of the organisation, not behind it. The decisions being made right now, in procurement and onboarding and the first emails about the new tool, are decisions about which side of the cliff each person will land on. The decisions are reversible only at high cost. They are designable only at low cost — if the design happens before the deployment, not after.

Edelman gave us the data. Edmondson gave us the framework. Sapolsky gave us the mechanism. The integration of the three is the operational picture: trust is built or broken in the first interaction, the first interaction is a moment of psychological risk-taking, and the nervous system that processes that moment weights the outcomes asymmetrically. None of this is a thermometer. All of it is a cliff.

Most organisations are designing for the hundredth efficient interaction. The adoption problem is on the first safe one.

The thermometer reading does not move because the thermometer is the wrong instrument. What the data is measuring is not a temperature. It is a position. And position, unlike temperature, is not adjusted by warming the room. It is determined by where the person was standing when the moment that mattered arrived.

The first interaction is the only one that matters. Most organisations are not designing it. That is the adoption problem nobody is solving — and the rare opportunity, for the SMEs that are still pre-cliff, to design it before the cliff is fixed.