Systems Thinking with Poisoned Systems#

Hazel Weakly, Nivenly Foundation; Sandeep Kanabar, Gen — Wildcard

• Why: An unusual framing — applying "garbage in, garbage out" and data-poisoning concepts not to ML training data but to the practice of operating carefully-tuned production systems with AI assistance layered on top — a genuinely different angle on AI-ops risk than the typical "AI will help SREs" pitch.
• What: A discussion of drawbacks (data poisoning, bias, inaccessibility, de-skilling) that emerge when AI assistance is introduced into carefully-tuned operational systems, plus strategies for keeping AI transparent and reliable rather than an inscrutable black box. (from abstract only)
• How: Draws on the speakers' personal experience working around biased/broken AI-assisted systems. (from abstract only — thin on mechanism)
• Where applicable: General for any org introducing AI-assisted tooling into existing operational systems.