Recent dates
How can AI help people recognize unsafe workflow behavior, intervene effectively, and verify that unwanted actions have stopped without disrupting legitimate work?
What is the problem?
AI systems can help people carry out multi-step tasks, from preparing messages to updating records and coordinating work. But when a workflow departs from someone's intent or permitted boundaries, that person needs to recognize the problem, intervene effectively, and understand what actually happened.
The challenge is to make that process clear and dependable, including when several tools or background tasks are involved. How can AI help people maintain meaningful control without overwhelming them with warnings or approvals?
Who experiences it?
Employees and business owners using AI tools, colleagues supervising their work, and customers or others affected by the resulting actions. Security, compliance, and governance teams also need understandable evidence of what happened.
The challenge invites contributions from computing, design, business, law, psychology, ethics, and other disciplines.
Why does it matter?
Unwanted actions can expose information, change records incorrectly, create financial costs, or damage customer trust. Excessive warnings and approvals can also undermine the usefulness of AI and encourage people to ignore alerts.
Better human control could help organizations adopt AI more confidently, reduce avoidable mistakes, and preserve its productivity benefits while making responsibility clearer.
What is the existing solution and/or what is hard about solving it today?
Approval prompts, permission limits, activity logs, and pause or cancel controls are useful starting points. The difficulty is connecting them into an intervention people can understand and verify. Stopping queued actions, interrupting work already underway, and addressing completed effects are different problems. A reassuring message from an AI system is not, by itself, evidence that an action stopped. Teams should also consider unclear warnings, incomplete visibility, delayed interventions, and unnecessary interruptions of legitimate work.
What would a better solution make possible?
A successful project would demonstrate a clearer path from recognizing a concern to intervening and checking the result. Students might explore AI-assisted decision support, interfaces, escalation processes, or evaluation methods; the model, architecture, and design are theirs to choose. In a safe demonstration, teams could measure missed concerns, unnecessary interruptions, time to intervene, and users' understanding of what stopped, what remains in progress, and what already happened. The goal is a practical improvement in human control, not a claim to eliminate every AI risk.
SnowCrash participation scope
SnowCrash’s initial participation will consist of one shared 30-minute kickoff, two shared 45-minute mentoring sessions, and up to one hour reviewing final deliverables or demonstrations. These are aggregate commitments across participating teams, not per team. Our selection of “test prototypes/feedback” refers to feedback on final deliverables or demonstrations, rather than a separate prototype-testing commitment.
To explore this challenge, contact:
Matt O'Brien: CEO, SnowCrash Labs Inc.
Email: matt@snowcrashlabs.com