Why users don't trust AI, and what your company can do about it

Anthropic published its first Public Record on June 12, 2026, a survey of how the public feels about AI. Only 15% of Americans said they trust AI companies to make decisions about how AI is developed and used.

Companies aren’t waiting for that number to improve. McKinsey's 2026 State of AI survey found that 44% of organizations now have AI scaling across the enterprise, up from 38% a year earlier. Stanford's 2026 AI Index measured the gap from the other side: just 10% of Americans say they are more excited than concerned about AI. Among AI experts, that figure is 56%.

When users don't trust the AI behind a product, they disengage, hold back their data, and take their business somewhere that feels safer. Understanding why the gap exists, and what actually closes it, is now one of the more important jobs a company has.

Why users don't trust AI

Answers that sound right and aren't

The most familiar complaint about AI is that it makes things up. Models produce fluent, confident text, and some of that text is wrong. Unlike a search engine, which hands you a list of sources to judge for yourself, a generative model usually presents its output as a finished statement of fact.

Unfortunately, that reliability is uneven in ways that are hard to predict. Stanford's 2026 AI Index calls this the "jagged frontier": the same top-tier model can score near 100% on a professional coding benchmark and still read an analog clock correctly only about half the time. Workers already say this shapes their behavior. Salesforce found that 71% of workers say consistently inaccurate outputs would break their trust in AI entirely.

The fix isn’t to hope for a perfect model, but to constrain where answers come from and show your sources. For example, when FullStack built Luxer for Lux Research, the assistant was designed to answer only from Lux's own library of more than 16,000 research documents, with every answer cited back to its source. It handled more than 900 client queries in its first two weeks.

Agents that go further than anyone intended

In July, OpenAI disclosed that one of its unreleased models broke out of a test environment and accessed Hugging Face's systems. Days later, Anthropic reported that three of its own models had reached the open internet during cybersecurity tests and gained unauthorized access to the infrastructure of three organizations. By September, a fourth incident had surfaced, and the CEOs of Anthropic, OpenAI, and xAI publicly agreed that frontier development should slow down.

In Anthropic's case, the models didn’t scheme their way out. They reached the internet through a network path that had been left open by mistake, and then did what they had been asked to do, on systems nobody meant to include.

For companies deploying agents, the lesson is practical: give an agent the narrowest set of tools and credentials that will get the job done, keep it inside a boundary you control, and put a person at the decision points that matter. 

FullStack built an agentic call-auditing system for a regulatory compliance provider on exactly that principle. The agent sorts and routes every sentence of every call on its own, but human supervision was moved to the one place it is needed: the final ranked report. Manual review dropped 99% while accuracy held at 93%, on par with human auditors.

Data privacy

Most people are not sure how their personal data is collected, used, or fed into AI systems, and that uncertainty makes them cautious about engaging at all.

A 2026 Malwarebytes survey found that 90% of respondents worry about AI using their data without consent, and 91% support national laws to regulate how personal data is used in AI. Stanford's 2026 AI Index found that even as more people globally see AI as beneficial, 52% say AI products make them nervous.

Privacy in AI systems is mostly a data-architecture question. If your AI reads from five ungoverned copies of the same content, no policy document will save you. When FullStack rearchitected the data layer for a Fortune 500 hospitality company, the goal was one governed access layer standing behind the CMS, the mobile apps, and the AI systems alike, so every downstream tool inherits the same permissions and the same rules.

Biased outputs

AI systems learn from past data, so they can absorb the same patterns and prejudices that shaped earlier decisions. Amazon's experimental recruiting tool was trained on historical hiring data and began downgrading resumes from women, even though gender was never a field in the model.

When people watch AI make unfair calls about who gets hired or who gets a loan, it reinforces the sense that these systems are neither neutral nor accountable.

The answer is to measure, and to keep a person at the decision. FullStack faced this question in its own hiring. When the company built its AI-enabled applicant tracking system, it didn’t simply switch on AI grading and hope. It ran human and AI graders side by side and compared outcomes: a 78% human pass rate against a 79.8% AI pass rate. Grading time dropped from 65 minutes to under two, and the final hiring decision stayed with a person. Parity isn't proof of fairness, but it is a number you can audit, and auditable numbers are where trust starts.

Lack of transparency

Many AI systems work as black boxes, producing outputs with no user-facing explanation of how they got there. That is uncomfortable in a chatbot and unacceptable in healthcare, lending, or hiring, where a decision can change someone's life.

When people don't know how a decision was made, it's hard for them to believe it was fair, and impossible for them to challenge it.

Environmental concerns

Data centers use a great deal of water and power. A large data center can consume up to five million gallons of water a day, roughly what a town of 10,000 to 50,000 people uses in a year. As AI workloads grow, so does the load on this infrastructure, and the public has noticed. A September 2026 poll from AP-NORC and the University of Chicago found that 53% of Americans are extremely or very concerned about AI's environmental impact, up from 41% a year earlier. Among adults aged 18 to 29, concern jumped from 38% to 60%. Nearly half now think AI will do more to hurt the environment than help it.

For someone already skeptical about how responsibly AI is being built, this is one more reason to wonder whether the people behind it are weighing the full cost.

How companies can build user trust in AI

Decide who is responsible

Many AI trust problems trace back to nobody owning the system. When there is no named owner, no written policy, and no process for handling incidents, problems get addressed only after a user finds them.

Start by deciding who in your organization is responsible for AI, then agree on a short code of ethics that sets basic rules: where AI can be used, what data it can touch, and when extra review is required. For higher-risk uses like hiring or access to services, spell out how decisions are checked, how they are explained, and how mistakes get reported and fixed. Keep a human reviewing the outputs that could materially affect a person's life.

It's also worth noting, however, that governance that is too heavy erodes trust from the other direction. A system that hedges every answer or refuses fair questions teaches users that it cannot be relied on, which is the same lesson a hallucination teaches. That is why it helps to track epistemic yield alongside your controls: how often does the system give people an answer they can actually use? Good governance should raise that number, not lower it. If your users are re-prompting, overriding, or abandoning the tool, the guardrails are part of the problem.

Give agents less room than you think they need

An agent should get its own identity and its own credentials, scoped to the task, never a borrowed copy of an employee's access. It should call only the tools you have explicitly exposed to it. Its network access should be an allowlist, not an open door. Anything that leaves the boundary, whether a payment, an external email, or a change to a production system, should pause for a person. And every action should be logged in a way an auditor can read later.

This is the same least-privilege thinking that has governed employee access for decades, applied to a new kind of worker. The difference is that this worker will try very hard to finish its task, so the fence has to be real and not a note in the prompt. Most of these checks can be plain rules that run in a fraction of a millisecond, without asking the model to police itself, which is also how you keep the controls from dragging down the yield described above. When FullStack rebuilt the cost-request process for a Tier 1 automotive supplier, the platform was built so that 100% of the workflow was tracked and visible to the relevant team members, precisely so that AI could later be put to work inside it. You can’t supervise a process you cannot see.

Explain where AI is, and where it isn't

When companies are clear about where AI is used, what it influences, and where people stay in the loop, users have context for judging what they see. Without it, many people assume more automation and less oversight than actually exists.

Soften the black box by explaining how a system uses inputs, how results are reviewed, and how a user can ask for a person to step in. Small forms of context and recourse make AI feel like something people can question, rather than something they have to accept.

Fold environmental impact into the plan

Treat energy and water as part of AI planning rather than a separate sustainability memo. That can mean choosing infrastructure that reduces consumption, being deliberate about where large models run, and talking openly about the trade-offs. The infrastructure itself is changing: NVIDIA's recent data center designs use high-temperature liquid cooling and dry coolers to eliminate most on-site water use.

Where to start

You don’t have to solve everything at once. A practical first step is an inventory: where is AI already used in your products and processes, who owns each system, what can each one access, and what have you told users about it. From there, formalize governance, tighten agent permissions, improve explanations, and set up regular bias and oversight checks.

For most organizations, the goal isn't perfectly safe or perfectly transparent AI. It's showing, visibly and repeatedly, that someone is responsible for how the system is built and what it's allowed to do. As more companies put AI into core services, that evidence of responsibility will matter as much as the capability itself.

If you’re building AI into your business and want it governed from the start, talk to FullStack. We design and deploy AI and agentic systems that are scoped to your workflows, wired into the permissions you already have, and accountable to the people who use them.