Prompting & interaction burden
How much input is actually necessary for a high-quality response?
Research pillar 02
Human–AI Interaction
How people formulate requests, evaluate machine-generated answers, allocate cognitive effort and make decisions together with AI.
Scientific question
When does AI genuinely reduce human effort without reducing quality?
The Human–AI pillar studies the full interaction loop: how a person frames a task, how an AI system produces an answer, how the person evaluates and verifies it, and how the combined system supports a final judgement or action.
The focus is therefore broader than prompt engineering. CoHMI is interested in trustworthy assistance, validation, cognitive cost and the distribution of decision functions between person and machine.
Current programme
MSP: Minimum Sufficient Prompting
3experiments
24tasks
360LLM responses
P0–P2prompting conditions
MSP asks whether task-relevant prompting can preserve answer quality while reducing unnecessary user input and interaction burden. The evaluation includes human audit and is being strengthened with direct human ↔ machine agreement measures.
Project detail →Connected research
Four layers of Human–AI interaction
Human validation
How do people inspect, annotate and verify AI-generated outputs?
Decision support
How should AI information be represented inside intelligent control and sociotechnical systems?
Trusted AI
Agreement, disagreement and failure analysis between machine evaluation and human judgement.
EEG with LLMs
Future study of cognitive and neurophysiological response to LLM outputs and verification tasks.
Student-scale “Blank” experiment
Large-sample study of solution quality, checking behaviour and use of MSP.
Selected outputs
From intelligent assistants to formal Human–AI models
Human-AI Interaction in Intelligent Control Systems: An Information Model
RusAutoCon 2026.
Personalization of an Intelligent Assistant for Complex Sociotechnical Systems
RusAutoCon 2025 · IEEE.
DOI ↗Minimum Sufficient Prompting (MSP)
Three experiments, 24 tasks, 360 responses, human audit and evaluator-agreement analysis.
Igor V. Petukhov
Liudmila A. Steshina
Evgenii N. Bogdanov
Alexander Taran
Mariia Petukhova