USC Researchers Test Adaptive AI Support to Reduce Dangerous Overreliance

USC researchers tested five adaptive interventions against a stubborn AI problem: people often trust advice too much, or reject it when they should listen. The study, highlighted through Google News coverage, found that changing an assistant’s behavior around a user’s trust level improved several decision outcomes.

The central result is not that people need more explanations. Persistent explanations sometimes made decisions worse. Instead, the researchers found that support worked better when delivered at specific moments of unusually high or low trust.

That distinction challenges a familiar product strategy. Many AI systems provide the same confidence cues, explanations, and interaction speed to every user. The new research argues that safer decision support requires a moving target: the interface must respond to how trust develops during repeated interactions.

The findings come from controlled studies involving science questions, simulated clinical diagnoses, and a large language model assistant. They offer a practical design direction, but they do not establish that real products can reliably infer trust or reproduce the results in live workplaces.

What the USC study actually changed

The researchers treated user trust as a changing input to the interface, not as a fixed attitude measured after the session.

Tejas Srinivasan and Jesse Thomason examined how trust affects whether people accept AI recommendations. Their paper, Adjust for Trust, appeared in the proceedings of the 2026 ACM Conference on Intelligent User Interfaces.

AI-assisted decision-making means that a person makes the final choice after receiving a recommendation from a model. Appropriate reliance occurs when the person accepts good advice and rejects bad advice.

Inappropriate reliance includes two opposite failures. Overreliance happens when someone accepts an incorrect recommendation. Under-reliance happens when someone rejects useful advice that would have corrected an initial mistake.

The researchers found strong relationships between self-reported trust and subsequent behavior. Across tested conditions, correlations between trust and switching to the AI answer ranged from 0.796 to 0.949.

The pattern became especially important at the extremes. Medical participants accepted 26% of incorrect AI diagnoses when trust was high, compared with 8% when trust was lower.

Low trust produced the reverse problem. Doctors rejected correct AI diagnoses 68% of the time during low-trust interactions, compared with 40% in other interactions.

Those results matter because a single instruction to “verify AI output” cannot fix both failures. A suspicious user needs evidence that makes a good recommendation easier to assess. An overly trusting user needs friction that makes an incorrect recommendation harder to accept automatically.

The study therefore tested trust-adaptive interventions. These are interface changes activated when a user’s reported trust crosses a defined threshold.

When trust was below 5 on a 10-point scale, the assistant could provide a supporting explanation. When trust was above 8, it could provide a counter-explanation describing why its recommendation might be wrong.

The researchers also tested slower interactions. One intervention required highly trusting users to wait 10 seconds before submitting a final decision. That pause aimed to interrupt an immediate switch to the AI answer.

These interventions changed the assistant without changing its underlying predictive capability. The system did not become more accurate because it received a better model. It changed how and when it presented information.

That is the study’s key event. Adaptive decision support moved from a general design proposal to a controlled test across several reliance conditions, including a large language model experiment.

Google News gave the finding a broader audience through coverage from Tech Xplore. However, the relevant evidence comes from the peer-reviewed paper and its reported experiments, not from the news aggregation layer.

Why Google News readers should care about calibrated AI trust

The study shifts responsibility from users alone toward the teams designing how AI advice reaches them.

Most safety guidance asks users to remain skeptical, check sources, and keep a human in the loop. Those practices still matter. However, they assume that vigilance remains stable throughout a long sequence of decisions.

The paper presents trust as dynamic. A correct answer can raise confidence in the assistant, while a visible mistake can sharply lower it. Recent outcomes can then bias how the user handles the next recommendation.

That creates pressure for companies building copilots, clinical systems, financial tools, and enterprise agents. A static interface can amplify yesterday’s experience even when today’s recommendation deserves a different response.

A user who sees several correct answers may begin accepting later suggestions without enough scrutiny. Another user may encounter one obvious failure and then ignore good recommendations throughout the rest of the task.

Neither response reflects the model’s actual reliability on the current case. The interface needs to help the user evaluate each recommendation without pretending that trust has disappeared from the process.

Supporting explanations helped when trust was low. Across relevant conditions, the study reported reductions of 13% to 31% in under-reliance. Total inappropriate reliance fell between 9% and 38%.

Final decision accuracy improved between 10% and 19% in those low-trust conditions. The largest gains appeared when the simulated AI was overconfident, a setting where users had more reason to develop low trust.

Counter-explanations addressed the other extreme. When trust was high, they reduced overreliance by 10% to 23%. Total inappropriate reliance fell by 19% to 36%, while final accuracy improved by 8% to 20%.

The most visible headline figure is a reduction of up to 38% in inappropriate reliance. Another experiment produced an accuracy improvement reaching 20%.

Those are relative improvements across particular experimental conditions. They should not be read as a universal performance guarantee for adaptive interfaces.

The pressure also extends beyond model developers. Enterprise buyers must decide whether a product’s safety controls are meaningful during actual use. A generic warning displayed during onboarding says little about how the system responds after trust changes.

Product managers need a different measurement plan. Adoption, time saved, and answer acceptance are incomplete if a system rewards indiscriminate reliance.

A high acceptance rate can look successful while hiding automation bias, the tendency to follow automated advice even when contrary evidence exists. A low acceptance rate can hide wasted capability and unnecessary manual work.

Appropriate reliance offers a more useful objective. Teams should measure whether people distinguish good recommendations from bad ones, not whether they agree with the system more often.

This matters for readers who found the story through Google News because the result applies to everyday AI use, not only specialized medical software. Search assistants, writing tools, and workplace copilots all shape trust over repeated exchanges.

Knowledge workers can already preserve supporting material and compare AI answers with original documents. A searchable personal knowledge base provides one way to keep evidence close to the decision.

Still, personal discipline cannot compensate for every interface choice. Systems that always speak fluently, answer immediately, and display uniform confidence can make careful evaluation harder than it needs to be.

Adaptive support beats explanations that never switch off

The important mechanism is selective intervention, because explanations can create new errors when they appear at the wrong time.

Explainable AI often assumes that more reasoning will produce better oversight. The study complicates that assumption.

The researchers compared adaptive supporting explanations with explanations shown persistently. Always displaying an explanation did not deliver the same gains.

In one diagnosis condition, persistent explanations worsened final accuracy. In another task, they increased under-reliance during low-trust interactions.

The authors suggest that repeated exposure to misleading explanations can reduce confidence in explanations themselves. A plausible-looking rationale attached to an incorrect answer can also encourage acceptance rather than scrutiny.

This is the primary conflict behind the research: static assistance versus trust-adaptive assistance. Static systems provide the same intervention regardless of the user’s state. Adaptive systems decide whether extra support, contradiction, or friction is appropriate for the moment.

The counter-explanation is especially notable. Instead of adding more support for its preferred answer, the system presents a reason that answer might be wrong.

That design breaks with the conversational style of many assistants. Current products usually aim for a clear response backed by a coherent rationale. Coherence can make uncertainty less visible, particularly when the model’s language sounds more confident than its evidence warrants.

A counter-explanation introduces structured doubt. It does not tell the user to reject the recommendation. It creates a competing argument that makes automatic agreement less attractive.

The forced-pause experiment pursued the same goal without generating additional explanatory text. Highly trusting users had to wait 10 seconds before making their final choice.

That delay reduced inappropriate reliance and improved accuracy in the tested setting. The result supports cognitive forcing, an interface technique that requires deliberate action before users can follow automated advice.

Friction normally has a negative meaning in product design. Teams remove clicks, reduce waiting, and shorten workflows because faster interactions often improve adoption.

Decision support changes that calculation. Speed becomes a risk when the user can approve an uncertain recommendation before considering contradictory evidence.

A well-placed pause can protect judgment. A pause placed before every low-risk action can produce alert fatigue and encourage users to work around the system.

The paper’s approach tries to distinguish those cases through trust. When reported trust is high, a short delay or counterargument can slow acceptance. When trust is low, a supporting explanation can prevent reflexive rejection.

The large language model experiment adds another layer. Participants used a Llama 3.1 assistant to answer questions from the AI2 Reasoning Challenge.

The researchers generated multiple answer-and-rationale pairs. The majority answer served as the model’s prediction, while its frequency across generations acted as a confidence estimate.

When trust was low, the system surfaced a rationale supporting the model prediction. When trust was high, it selected a rationale supporting the alternative answer, if one existed among the generated samples.

Decision accuracy reached 67.8% with the adaptive LLM, compared with 58.4% without intervention. Total inappropriate reliance fell from 83.5% to 68.2%.

Under-reliance during low-trust moments fell from 65.1% to 44.4%. Overreliance during high-trust moments fell from 56.4% to 42.4%.

These results strengthen the mechanism, but they also reveal its dependence on generated reasoning. An LLM can produce a persuasive explanation that contains invented or unsupported details.

The authors explicitly identify hallucinated explanations as a concern in high-stakes settings. Adaptive delivery cannot make unreliable reasoning safe merely by choosing a better moment to display it.

Independent research also warns against treating explanations as a universal remedy. A 2026 study on AI error difficulty found that explanations were most useful for moderately difficult errors that users could still verify.

That boundary fits the USC findings. Adaptive support works by promoting evaluation. It cannot help much when the person lacks the information, expertise, or time needed to evaluate the recommendation.

The practical lesson is narrower than “personalize every AI interface.” Teams should first identify where trust affects acceptance, then test a small intervention against measurable decision errors.

Trust sensing creates its own safety problem

An assistant cannot adapt safely to trust unless it can estimate trust without manipulating users, invading privacy, or confusing confidence with competence.

The experiments relied on explicit self-reported trust. Participants reported how much they trusted the assistant, and predefined thresholds determined when the system intervened.

Most commercial tools do not ask for a trust score after every interaction. Repeated surveys would slow workflows and might change the behavior being measured.

Implicit estimation offers an alternative. A product might infer trust from answer acceptance, editing patterns, response time, requests for sources, or repeated overrides.

Each signal is ambiguous. A fast acceptance might indicate trust, deadline pressure, distraction, or careful prior knowledge. Rejection might reflect justified skepticism, personal preference, or misunderstanding.

The system could therefore apply the wrong intervention. It might confront a careful expert with unnecessary counterarguments while reassuring an uncertain novice with a weak explanation.

That failure would be more than a usability issue. A trust-adaptive assistant can influence the psychological state it claims to measure.

For example, displaying “the AI is thinking” may increase perceived effort and competence. A counter-explanation may lower confidence even when the original recommendation is well supported.

The paper tested an “AI is thinking” delay during low trust and did not find the same benefit as a supporting explanation. Waiting alone did not supply the evidence needed to reassess a useful recommendation.

The finding highlights a distinction between productive friction and theater. A delay can encourage reflection when someone is rushing toward acceptance. It should not imitate deeper reasoning when no additional reasoning occurs.

Privacy presents another problem. A detailed trust model could become a behavioral profile covering hesitation, compliance, expertise, and susceptibility to persuasion.

Organizations would need clear limits on collection, retention, and secondary use. Trust signals gathered for safety should not quietly become performance scores or targeting inputs.

Clinical deployment raises the stakes further. The study used simulated diagnosis tasks, not real patient care. Controlled experiments offer internal consistency, but they omit institutional pressures, incomplete records, liability concerns, and multidisciplinary review.

Doctors also differ in specialty, experience, and familiarity with a case. A fixed threshold on a 10-point trust scale cannot capture every relevant source of uncertainty.

The paper acknowledges another limitation: real environments often lack immediate feedback about whether a decision was correct. Without outcome feedback, neither the user nor the system can update trust against a reliable ground truth.

Many business decisions produce delayed or disputed outcomes. A hiring recommendation, fraud alert, or strategic forecast may not receive a clean correctness label for months.

Trust calibration becomes harder in those settings. The assistant might adapt to confidence signals while remaining unable to determine whether the confidence is justified.

The study also counts under-reliance and overreliance within a combined inappropriate-reliance measure. Yet the costs of those errors vary by domain.

A false negative in medicine may carry different consequences from unnecessary testing. A financial institution may value catching fraud differently from reducing false alarms.

Product teams therefore need utility models that reflect actual harms. Improving a combined score can conceal a dangerous increase in the more costly error type.

Research published in PLOS One previously proposed adaptive trust calibration based on reliance behavior and cognitive cues. The USC work adds controlled evidence, but real deployment still requires validation beyond a laboratory interface.

A separate Communications Psychology study found that people adjusted their reliance across artificial advisers with different profiles. Their integration remained significantly below a modeled optimum.

Together, these studies suggest that adaptation is possible but imperfect. People do respond to perceived adviser quality. They do not consistently combine advice in the accuracy-maximizing way.

The risk is that companies treat trust adaptation as another personalization feature. Safety claims require a higher standard, including preregistered evaluations, domain-specific error costs, and tests against manipulation.

Google News readers should therefore interpret the headline result as evidence for a design hypothesis, not a finished safety layer. The experiments show that context-sensitive interventions deserve testing. They do not show that any current assistant understands when a specific user trusts it too much.

What comes next for adaptive AI decision support

The next phase depends on three signals: field trials, trustworthy trust estimates, and evidence that adaptive interfaces protect skills over time.

The first signal is replication in real work. Researchers need field studies where professionals face genuine constraints, delayed feedback, and consequences beyond experimental scores.

Healthcare provides an obvious test case, but deployment should begin with bounded decisions and independent review. Researchers should report separate rates for accepting incorrect advice and rejecting correct advice.

If adaptive interventions improve both measures without increasing workload or alert fatigue, the paper’s central claim becomes stronger. If gains disappear outside controlled tasks, the mechanism will need narrower use.

The second signal is a defensible trust-estimation method. Explicit ratings are transparent but burdensome. Behavioral inference is less intrusive during work but easier to misinterpret.

A credible system should show which signals trigger an intervention. Users should be able to correct the estimate and disable personalization without losing access to basic safety information.

Evaluations should also test demographic and expertise differences. A trust detector that works for one group can misclassify another group’s hesitation, language patterns, or interaction speed.

The third signal is evidence about long-term skill retention. Short experiments measure immediate choices, while daily AI use can reshape how people search, verify, and remember.

A 2025 clinical study reported that physicians’ unassisted adenoma detection declined after regular exposure to computer-aided detection. The clinical findings00139-7) prompted debate about possible deskilling and alternative explanations, including workload.

Adaptive support should therefore be judged across two timelines. It must improve the current decision without weakening the person’s ability to decide when the system is unavailable.

That requirement favors interfaces that keep evidence visible and encourage independent judgment. It argues against tools that merely add persuasive text around a recommendation.

For everyday knowledge work, users can create their own verification layer. Saving source material, documenting why a recommendation was accepted, and using knowledge blending can make later review easier.

Product teams can begin with similarly modest controls. They can test counterarguments after rapid acceptance, request evidence before high-impact actions, and track whether users catch seeded model errors.

They should avoid inferring that hesitation equals distrust or that agreement equals success. The target is decision quality, not compliance with the assistant.

The Google News attention around this research reflects a wider concern about AI dependence. The paper adds something more useful than another warning: a testable mechanism for changing interfaces when trust becomes extreme.

Its strongest result is also its most cautionary one. More explanation was not always better. Support improved decisions when its form and timing matched the user’s state.

The next useful question is not whether people should trust AI. It is whether a specific system helps them recognize when trust is warranted, when doubt is necessary, and when neither the user nor the model has enough evidence.

Watch for products that publish those measurements instead of offering generic assurances. Ask whether their safety prompts change behavior, whether users can inspect the trigger, and whether performance survives outside a controlled demo.

Adaptive decision support earns credibility only when it improves judgment without turning trust itself into another opaque prediction.