A 461-publication review mapped tourism AI gains and risks—but did not pool outcomes
The synthesis linked faster service and information search with bias, hallucination, privacy and deskilling. Only 12 included publications received its highest quality rating.
The short version
- What happened
- A PRISMA review searched Web of Science and Scopus through 9 September 2026, identified 3,825 records and included 461 publications in a thematic synthesis.
- Who it affects
- A traveller may get faster help but still needs accurate advice and clarity about whether a person or system is responding. Hospitality workers need oversight rules that protect judgement, privacy and opportunities to build skills.
- What remains unproven
- 461 publications
Keep in mind
461 publications
Two databases
English search terms
12 high-quality publications
No pooled effect
Many self-reported studies
Search ended 9 September 2026
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Efficiency and authenticity pull in opposite directions
Generative AI can shorten tourists' information search and help hospitality businesses produce content or automate routine service, but those efficiencies sit beside bias, hallucination, privacy trade-offs and employee deskilling, according to a peer-reviewed systematic review published on 29 October. The authors argue that machine-generated empathy can also violate guests' expectations when an interaction feels automated rather than genuinely attentive.
That tension matters for travellers and workers because hospitality depends on judgement, trust and interpersonal service as well as speed. A chatbot may answer a standard question at any hour, while a fabricated recommendation or opaque use of personal data can cause harm. For staff, AI may remove repetitive work or narrow opportunities to develop expertise. The review does not establish how often each benefit or harm occurs in real businesses.[1]
A broad map of 461 publications
The team followed PRISMA 2020 and searched Web of Science and Scopus in two rounds, covering work from 2021 to 9 September 2026. The supplementary search retrieved 3,825 records; after 1,283 duplicates, screening and full-text review, 461 publications were included. Of these, 398 were empirical—270 quantitative, 77 mixed-method and 51 qualitative—and 63 were theoretical or review publications.
Two reviewers performed design-sensitive quality appraisal. Only 12 publications, or 2.6%, were rated high quality; 395 were rated moderate and 54 low. The synthesis counted article-level evidence and developed a framework for balancing efficacy and risk, but it did not pool comparable outcome sizes or estimate a causal productivity, revenue, satisfaction or employment effect. A large corpus is not the same as strong intervention evidence.[1]
The distinction that matters
What the review can and cannot establish
It provides a broad evidence map rather than one pooled effect estimate.
Search
3,825 records from two databases
461 publications included
Quality
Design-sensitive appraisal
12 high; 395 moderate; 54 low
Inference
Thematic synthesis
No pooled causal effect
What managers can use—and what remains unknown
The authors propose tiered human oversight: automate low-risk tasks, keep people supervising consequential or ambiguous interactions, and retain human control where authenticity, privacy or safety matters most. That is a governance proposal derived from the literature, not a tested operating model. On-trip studies using embodied systems or sensors were comparatively scarce, and many underlying studies relied on self-reported intentions rather than observed behaviour.
The search used English terms in two databases, omitted backward and forward citation searches and largely returned English-language work. Its September cutoff also excludes the newest deployments. The authors reported no funding and no commercial conflicts. Stronger evidence would compare human, AI-assisted and automated service in real operations, measuring accuracy, guest outcomes, worker workload, skill development, complaints and privacy incidents over time.[1]