Aim
AI tools are increasingly used to support knowledge intensive work such as literature review and data curation. While such tools systems promise efficiency gains, they also shift cognitive demands to the verification of AI outputs. [AD1.1][JI1.2]This paper reports on an exploratory, human centred investigation conducted during the development of SciLire, an AI supported data extraction tool, with the aim of understanding how people validate AI generated outputs in practice.

Method
SciLire is developed through iterative engagement with domain experts from a range of scientific backgrounds, including biomedical manufacturing, plant genomics and agricultural science, using the system in their own projects. Observations of tool use, use case discussions, and user feedback informed features to support user verification of AI outputs. These practice based observations were later complemented by a user study that explored how different forms of supporting information influenced users’ checking and judgment of AI generated answers.

Results
We found that users’ verification strategies varied widely depending on factors such as project goals, project maturity (e.g., exploratory vs confirmatory) and tolerance for error. These factors shaped how participants interacted with SciLire and their long-term adoption of the tool. Results of the user study indicated that verification was more efficient when the system provided excerpts from source documents, while AI‑generated explanations supported faster judgments but also encouraged uncritical acceptance of incorrect outputs.

Discussion/Implications
These findings position designing methods for appropriate oversight and verification of AI outputs as a central element for effective AI tool development. From a human factors perspective, the challenge is to design systems that support a range of verification strategies aligned with different project goals and user needs, enabling appropriate trust calibration. In practice, this requires balancing potential AI-assisted productivity gains against increases in cognitive demand and potential reductions in the quality of human judgment.

We introduce a human-centred, ethical, assured, responsible, and secure (HEARS) design framework developed in alignment with the EPSRC National Artificial Intelligence for Collective Intelligence Hub (AI4CI). By addressing systemic vulnerabilities in the design and deployment of AI agents, HEARS establishes the sociotechnical foundation necessary to safely harmonise human-machine intelligence within shared digital ecosystems. By shifting focus from isolated AI capabilities to collective interconnected human-machine ecosystems, HEARS principles can optimise user acceptance, adoption and continued use of AI agents:
(i) Human-centred Design (HCD): Optimisation of e.g. trust, affordances, accessibility, adaptability, and usability. Examples of our research include HCD of agentic interventions – including transparent normative messaging for human-agent collectives (N=400), peer-awareness (N=176), and personalisation (N=176).
(ii) Ethics: Resolving data harvesting, confidentiality, and anonymity challenges under regulatory standards with communication models to deliver transparency.
(iii) Assurance: Designing verifiable protocols to evaluate and communicate AI agent safety and reliability and elevating user preparedness and systemic resilience in situations where AI does not perform as intended or expected.
(iv) Responsibility: Expanding the ethical scope to assess broader societal impacts, ensuring that collective AI systems inherently align with stakeholder values and legislative benchmarks. Our focus group activities are based on RRI using the AREA+4P framework - with respect to definitions and operationalisation of agentic intervention design, transparency, personalisation, and culture (N=30).
(v) Security: Embedding secure-by-design principles that fortify human dimensions of cyber security within network environments.
HEARS delivers a vital blueprint and significant step towards realising the perfect fit between people and the design of intelligent systems. It establishes how embedding structural trust and mutual legibility will allow distributed human groups and AI agents to safely cooperate; transforming isolated automated tools into secure, resilient systems for collective intelligence. Our application domains including crisis (e.g. pandemic) response, environmental intelligence, financial stability, healthcare ecosystems, and smart-city design.

Aim: Human–autonomy teams (HAT) comprise humans and autonomous agents, such as Artificial Intelligence (AI), working together toward shared goals. As organisations increasingly integrate AI into teams, there is a need for systems Human Factors and Ergonomics approaches to inform optimal design. This study aimed to model how different approaches to AI integration impact teamwork in a university research setting.
Methods: Subject-matter expert (SME) input was used to develop an Event Analysis of Systemic Teamwork (EAST) model of a university-based research team, including task, social, and information networks. Once the traditional system (no AI use) was modelled, AI capabilities were defined through desktop research and SME input, and networks were modified to reflect AI introduction in line with two scenarios: (1) introducing AI with current capabilities as additional technological actors within the EAST network; and (2) replacing a human team member (Research Fellow) with an advanced AI agent according to anticipated future capabilities. The EAST outputs were quantitatively analysed using network metrics and qualitatively interpreted to identify differences between the traditional human-only team and the two HATs.
Results: Introducing AI into the EAST network altered teamwork, with effects differing between the two scenarios. In scenario 1, the number of nodes and interactions in social and information networks increased, reflecting increased coordination and information exchange demands. In contrast, scenario 2 more closely resembled the human-only team at the global network level. However, the betweenness centrality of key human actors in this scenario increased, suggesting these team members may take on greater responsibility when AI is used to replace existing roles.
Discussion/Implications: Findings suggest that the impacts of introducing AI differ depending on AI capabilities and integration choices. Using systems approaches can help organisations to identify potential unintended consequences (e.g. increased demands on human team members) prior to implementation and inform HAT design.

Introduction
Human-Robot Collaborations pose challenges for Human-Robot Interaction from a Human Factors (HF) perspective, many of which are only now being identified. Robot hardware capabilities and the sophistication of the algorithm protocols can impact the safety and efficiency of the robot interaction interface. Review of the literature identified limited evidence regarding the human capabilities, knowledge, and skills within human-robot collaboration, internationally or in Australia.
Aims:
This study used a structured method of analysis which examined the performance differences between inexperienced and experienced robot users who performed a specific task which was a representative of human-robot collaboration in industry.

Methods
The study was undertaken in a university research laboratory with a protocol developed by a research engineer, to evaluate performance of experienced and novice operators using structured observation of a simple task, including:
• A Cobot Task Protocol
• Questionnaire including the NASA Task Load Index survey’
• Observation Records
Novice operators (n=8) were recruited during two site visits to the research centre by professionals with an interest in Cobots in May 2023. Experienced cobot users (n=3) were recruited by the research centre staff. The majority (64%) had never interacted with a cobot, and the remainder (36%) had prior interaction with cobots.
Findings
Novice operators were more likely to adopt a bottom-up problem-solving approach whereas experienced operators were more likely to problem solve using a top-down method of problem solving.
Discussion
The expectation would be that experienced operators would overall perform better in all areas, performing the cobot task with greater reliability, efficiency, precision and with less mental workload when compared to novice operators. However, analysis of the NASA-TLX workload results did not support this assumption.
There is no perfect fit of humans working together with collaborative robots; however this study presents reliable methods of study for use in future research.

Aim: Early detection of clinical deterioration in hospitals is a complex sociotechnical challenge influenced by cognitive workload, communication variability, and inconsistent mental models across multidisciplinary teams. Traditional monitoring relies on manual assessment and subjective interpretation, creating opportunities for error and bias.
Our aim is to design, implement, and evaluate an AI enabled system that supports consistent, objective risk assessment and strengthens shared mental models between physicians and nurses, thereby improving early identification and prevention of clinical deterioration.

Methods: A multidisciplinary AI enabled system was deployed at Stanford Hospital. It included: (1) an ML model predicting clinical deterioration; (2) a dual role alerting mechanism for physicians and nurses; and (3) a structured interdisciplinary huddle supported by a shared checklist to align care plans. The workflow was intentionally designed using human‑factors principles addressing interruptive alerting, trust calibration, distributed cognition, and cross‑role communication. The intervention began in select units in January 2021 and expanded to all non ICU units in June 2022. Quantitative outcomes were assessed using a regression discontinuity design; qualitative insights were gathered through clinician surveys (n=134).

Results: From January 2021 to November 2022, the intervention was associated with a 10.4% absolute reduction in deterioration events (inpatient codes, rapid responses, ICU transfers; p=0.03). After expansion, huddle completion reached 80%. Survey results showed strong acceptance, particularly among nurses: 48% reported they always or often agree with model predictions, and 78% found the system valuable for patient care.

Discussion/implications: This work demonstrates that AI systems designed as human centred, team integrated components of clinical workflows—rather than standalone prediction tools—can enhance shared situational awareness and improve patient outcomes. It highlights the essential role of ergonomics in shaping safe, trustworthy, and resilient AI enabled care delivery