An emergency demand forecast model trains partly on scraped social posts with no provenance or authenticity checks. What lifecycle risk does that create?
Select an answer to reveal the explanation.
Short Explanation
Scraped posts with no pedigree are open windows for poison—don’t let rumors train the forecast.
Full Explanation
An emergency demand forecast model that trains partly on scraped social posts with no provenance or authenticity checks illustrates untrusted scraped civic-adjacent data as a poisoning pathway into training. Manipulated narratives, coordinated fake posts, or injected samples can bias forecasts that emergency operations may use for staffing and resource posture. Without provenance, authenticity filters, and human review of high-influence samples, the municipal training corpus becomes an open window for training-time attacks and rumor amplification. The correct emphasis is closing that pathway with source allow-lists, authenticity signals, rate/anomaly filters, and integrity checks—not historical dispatch-log overfitting.
Historical dispatch logs overfitting to prior incident patterns is a model-generalization issue, not a training-set poisoning pathway from unauthenticated scraped social posts. Seasonal shelter demand shifts causing gradual performance drift is concept drift in live demand, not an integrity failure in the training corpus. Delayed batch ingestion pipelines producing stale forecasts during surge periods is a serving-latency problem, not a provenance or authenticity gap in training data. Each distractor describes a different ML failure mode, none of which is the untrusted scraped-social-data poisoning pathway.
Exam caveat: social data can still be useful for situational awareness products when clearly labeled and separated from automated training that drives critical forecasts. The risk is specifically unsupervised promotion of unauthenticated scraped text into the training set. Also distinguish collection for analyst review from silent nightly retrain on the raw scrape.
Operational check: inventory which emergency ML jobs consume scraped social feeds; require documented provenance fields, authenticity scoring or allow-listed accounts, and a dual-review gate before those rows enter training. Canary poisoned hashtags in a sandbox scrape and confirm they never reach the production training bucket. If they do, treat the pathway as open until filters and signing are enforced.