What happens to a decision when an AI model confidently asserts something that simply is not true? A pre-registered study titled When Hallucinations Travel seeks a concrete answer to that question. In two experiments, participants read an AI response to a medication choice, with the strength of the hallucination varying. The outcome is sober but telling: as the fabricated explanation looked more convincing, it changed how participants interpreted risk, the causal narratives they formed, and how well they later recalled where information came from.
The researchers call this travelling: hallucinated mechanisms seep into the user's own reasoning. Participants adopted fabricated explanations in their decision language, relied less on the original facts and reconstructed source provenance more poorly. For anyone working with high-impact information, this is an important observation: a hallucination is not only an error in the output, but an input that can subtly yet systematically distort human judgement.
From error rate to decision dynamics
It is tempting to reduce hallucinations to a figure: what percentage of answers are wrong. Recent literature suggests that this is too narrow. The Survey of AI Hallucinations and Mitigation (AMCIS 2026, Thompson) organises the definitions, types and causes of hallucinations and stresses that they are socio-technical phenomena. They arise from a combination of data quality, model architecture, prompts and evaluation mismatches, but their effect only truly unfolds in use. In healthcare, law and finance, these risks are especially consequential, because a plausible-sounding but incorrect claim can lead directly to a different choice.
An arXiv review on hallucinations in organisation-bound AI advisers makes this even more concrete. The study distinguishes scepticism, factual verification, the success of that verification and the eventual reliance. The conclusion is sobering: warnings and 'hallucination risk' labels often have little effect, and people frequently continue to rely on incorrect information despite explicit awareness. Simply telling someone that an answer might be unreliable does not, by itself, change behaviour.
Trust as a risk bearer
Alongside the content of decisions, the pattern of trust also shifts. The article The trust crisis in artificial intelligence in Technology in Society (Cheng et al.) reports, on the basis of a large-scale survey, that hallucinations undermine both cognitive and emotional trust in AI, and thereby weaken the effectiveness of human-AI collaboration. The degree to which a tool fits the task determines how strongly that breach of trust carries through.
The consequence is a double movement. After a series of good answers, overconfidence easily arises: users switch to autopilot and check less. A single visible error can then permanently reduce the willingness to use AI at all. For governance, this means the question is no longer only how good the model is, but where AI may advise, where it may only summarise, and where a human always takes the primary decision.
From model tuning to verifiable decision architecture
If better models alone are not enough, where is the gain to be found? Partly in detection. A NIST publication on hallucination detection with diversion decoding describes a method for deriving an uncertainty measure during generation and thereby signalling hallucination risk, more efficiently than a number of existing approaches. That makes hallucination risk usable as a signal: an indication to check an answer more thoroughly.
But a detection score is only useful once it is attached to a process. Taken together, the studies point in the same direction: hallucinations belong in a decision architecture, not merely in model statistics. In concrete terms, that means coupling uncertainty signals to mandatory follow-up steps — an extra source check, a comparison between multiple models, or explicit human review — and making those steps visible and traceable.
What this means in practice
For professionals working with confidential or high-impact information, the common thread is clear: treat hallucinations as a decision risk. Record which AI answers were weighed in which decisions, which signals occurred in the process and which verification was carried out. That way it remains traceable and accountable when a questionable answer was noticed and what was done with it.
Here lies the role of a verification console such as Vera. Vera is not a language model and not a chatbot, but a verification layer: the design is to have AI answers checked through multiple models and to make those checking steps visible, so that users gain more insight into where answers diverge or are uncertain. This fits the research finding that standalone warnings do little — it is about workflow-bound checkpoints. Through evidence-logging, it can be recorded which answers were weighed and which verification was performed, which makes later review possible.
The privacy side follows on from this. In Vera's design, pre-processing and anonymisation take place on EU infrastructure, and the workflow is set up so that only anonymised content is sent onward to the selected AI models; if the privacy check fails, nothing is forwarded. Documents can be viewed and edited within the same environment in Vera Office, so content does not need to leave the secured context for that purpose.
The sober undertone of the research remains important: no tool provides certainty about correctness, and Vera does not remove hallucinations. What a verifiable architecture can do is reduce the chance that a fabricated claim travels unnoticed into a decision — by making uncertainty visible, enforcing verification and recording the trail of choices. The professional final judgement remains, as it should, with the human.