… a standing rule, filter, recommender, or agent has acted on a case the actor did not want touched, or left alone one they did, and the actor is looking at that case: in the result, in a notification, or among the items the rule refused. Even where a trial run checked the rule beforehand, cases arrive in use that the trial did not meet, and the governing logic sits in settings the actor seldom visits. Where the fault is in the wording or shape of a generated draft rather than in a standing decision on a case, the repair is an instructed revision of that output.
Problem
When a standing rule, filter, recommender, or agent gets one case wrong, the actor meets the mistake on that case and wants it fixed there, with or without changing what happens next time. If the only remedy is to leave the work and edit the rule in settings, most people put up with the error or turn the automation off; if every one-off fix silently rewrites the rule, a temporary exception corrupts future behaviour.
Skipping a song and hiding it for good are different requests, and music services offer both on the track. A spellchecker’s “Ignore once” and “Add to dictionary”, a mail client’s “Not spam” and “Filter messages like these”, a browser blocker’s “Allow this time” and “Always allow on this site”, and a feed’s “Not interested” and “Don’t recommend channel” all separate the immediate override from the standing change, and put both on the item itself. That separation is what keeps delegation workable once a rule or model starts making mistakes in use.
Solution
Put the repair on the case itself, and separate fixing this case now from changing what happens next time. Offer, in rising order of commitment: undo or skip this once; make a standing exception for this item or attribute in either direction; or change the rule or preference that produced it, starting from an explanation of why this case matched. Accept the smaller corrections in one gesture without demanding a reason, while letting the actor narrow why if they choose.
Keep standing exceptions in a list the actor can review and clear, so overrides accumulated one by one do not silently cancel the rule they qualify.
A wrong refusal is harder to meet than a wrong act, because the message filtered out or the payment declined is out of sight. Where the rule refuses or filters out, keep what it refused where the actor can open it and reverse it, as a spam folder does, with the reason on each entry and the least certain entries first. Few people open such a list unasked, so now and then ask about a close call, as a bank asks whether a declined payment was the actor’s own. Where a held item may be harmful, show enough to recognise it, such as the sender and subject, and not the item itself.
Continuation
The case is put right on the spot, and what governs future cases can stay as it is or change from here. The smallest repair is an undo or a one-off skip that leaves the standing behaviour untouched. An explanation on the case shows which rule, signal, or preference produced it, and is where a deeper change starts. When an override should hold beyond this item, or when the same correction recurs across cases, it opens rule composition on the clauses that matched this case or becomes a standing instruction in the actor's own words. Where the actor can show the right handling on the case more easily than state a rule for it, tuning by demonstration takes the place of editing the rule. Dismissing a suggestion is the same move at lower stakes, since the system only proposed the act. Standing exceptions that build up out of sight can end up cancelling the rule they qualify, so the list of them stays open to review…
All connections
- Precedes
- Standing instructions — a case correction that keeps recurring is what a standing instruction is written from
- Rule composition — When an override should hold beyond this item, or when the same correction recurs across cases, it opens rule composition on the clauses that matched this case or becomes a standing instruction in the actor's own words.
- Serves
- Delegation — covers the repair use of tuning tools in delegation's exception handling
- Enacts
- Agency — the actor overrules standing behaviour where they meet its mistake
- Malleability — standing behaviour changes through small corrections made in use
- Complements
- Undo — The smallest repair is an undo or a one-off skip that leaves the standing behaviour untouched.
- Alternatives
- Instructed revision — case correction overrides a standing decision on a case, and instructed revision reshapes a generated output
- Tuning by demonstration — Where the actor can show the right handling on the case more easily than state a rule for it, tuning by demonstration takes the place of editing the rule.
- Related to
- Suggestion — Dismissing a suggestion is the same move at lower stakes, since the system only proposed the act.
- Preceded by
- Trial run — case correction handles the live case that a trial run did not meet and the running rule got wrong
- Enabled by
- Explanation — An explanation on the case shows which rule, signal, or preference produced it, and is where a deeper change starts.
- Appears in
- Delegating standing work: Putting it right, step 2
Research on this pattern
- Separating a one-off override from a standing change: when a learning thermostat treated every manual adjustment as a change to the standing schedule, temporary exceptions corrupted future behaviour and people hesitated to adjust it at all — the basis for keeping “this once” distinct from an exception or rule change.
- Correcting the reasons on the case rather than only relabelling the item: when a classifier showed why it predicted a label on a single item and let the actor adjust those reasons, actors formed a more accurate picture of the system’s logic and fixed mistakes with far fewer corrections than through item-level feedback alone — the basis for starting a rule or preference change from the case’s explanation.
- Reliance after seeing a mistake: people who saw a forecasting algorithm err abandoned it when they had to take or leave its output whole, and kept using it when they could modify what it did, even slightly.
Resources & references
- Noessel (2017) Designing agentive technology — skipping a case, exception lists in both directions, changing a rule from the case that broke it, and the list of rejects with a filter for low-confidence cases
- Microsoft Defender — quarantine notifications — a periodic digest of held messages with release and review actions
- Parasuraman & Riley (1997) Humans and automation: use, misuse, disuse, abuse — misses and false alarms weigh differently on reliance
- Yang & Newman (2013) Learning from a learning thermostat: lessons for intelligent systems for the home — one-off overrides mistaken for standing schedule changes
- Kulesza, Burnett, Wong & Stumpf (2015) Principles of explanatory debugging to personalize interactive machine learning — corrections made on the explanation of a single prediction
- Dietvorst, Simmons & Massey (2018) Overcoming algorithm aversion: people will use imperfect algorithms if they can (even slightly) modify them