AI Advice: Give Every Human Adjustment a Reason

Imagine an organiser using an AI tool to estimate attendance at a workshop. It suggests 40 people. She changes the number to 48 because she wants spare chairs. Her decision may be sensible, but the edited number now mixes two questions: how many people are expected, and how much capacity she wants available.

Human control over AI advice deserves that level of precision. Being able to intervene can make advice easier to use. An intervention can also add missing information, express a preference or introduce an error. Before celebrating either human judgement or machine accuracy, identify what the adjustment was meant to accomplish.

Why even a small amount of control can matter

Berkeley Dietvorst, Joseph Simmons and Cade Massey investigated control in three forecasting studies, first published online on 4 November 2016 and appearing in the March 2018 issue of Management Science. Participants were more willing to use an imperfect algorithm when they could modify its forecasts, including when adjustments were tightly restricted. Their preference was relatively insensitive to how much modification was permitted.

The researchers also examined satisfaction and performance. Allowing some adjustment increased satisfaction with the forecasting process, while constraining adjustments could protect performance compared with allowing unrestricted departures. These were incentivised tasks using statistical forecasts, including predictions of students’ test scores. They were not evaluations of present-day conversational AI or evidence that adding an edit button will improve every decision.

The useful lesson is a question about participation: what does the opportunity to intervene change? A person may become more willing to use advice without every alteration making that advice more accurate. Acceptance, satisfaction and predictive performance need their own evidence.

People can also favour algorithmic advice

A separate March 2019 paper by Jennifer Logg, Julia Minson and Don Moore complicates the idea that people generally reject algorithms. Across estimation and forecasting experiments, participants often placed more weight on equivalent numerical advice when it was labelled algorithmic rather than human. The authors called this algorithm appreciation. The pattern varied with expertise and whether people were comparing the algorithm with their own judgement.

This matters when interpreting a changed answer. Someone might accept a recommendation because it appears computationally authoritative, or revise it because they prefer their own estimate. The research supports examining the decision context. It does not provide a universal rule that humans either distrust or defer to AI.

These older studies offer useful distinctions for thinking about AI advice today. Their findings should travel with their task boundaries. A statistical forecasting aid, a writing assistant and a conversational companion do not perform the same job.

Name the reason for the change

For everyday work, try completing one sentence before applying an adjustment: “I am changing this because…” The following distinctions are a proposed practical framework, rather than categories validated by those experiments.

• A correction: the advice used a wrong fact, calculation or source. Identify the specific error and the evidence supporting the replacement.
• Additional information: a relevant fact was unavailable to the system. State what was missing and how it affects the recommendation.
• A preference: you want a different tone, emphasis or balance of goals. Explain the criterion you are choosing to prioritise.
• A planning margin: you are allowing for uncertainty or consequences. Keep that allowance distinguishable from the underlying estimate.
• An unresolved concern: something seems wrong, but the reason is unclear. Treat that as a question to investigate rather than a confirmed correction.

The categories can overlap. A creator might reject a proposed headline both because it overstates the evidence and because its tone clashes with the publication. Recording both reasons helps distinguish a factual correction from an editorial preference.

Keep the forecast and the decision visible

Return to the fictional workshop. The organiser can preserve three separate entries: the tool’s estimate of 40 attendees, her decision to prepare 48 chairs and her reason for keeping eight spare places. No assumption about improved forecasting is needed to explain the capacity choice.

Suppose 42 people attend. For this one event, the original estimate was two people away from the outcome. Treating the capacity figure as a forecast would make it six people away. Yet the extra chairs served the planning purpose. Those statements can all be true because prediction and preparation were answering different questions.

The numbers are illustrative, not results from an actual event. One case also cannot establish which forecaster performs better over time. Its value is showing how an unlabelled adjustment could distort a later review: a sensible capacity allowance might be misclassified as a poor forecast, or presented as proof that human intervention improved prediction.

Ask what the intervention can actually change

For a small team choosing an AI workflow, inspect the practical meaning of control. Can a person correct an input, change a stated goal, edit an output, request another option or stop an action? These are different opportunities to intervene. A visible control is useful only in relation to the decision the person needs to make.

In the workshop example, changing the capacity allowance should not require pretending the estimated attendance changed. A useful record could retain the original suggestion and the final decision side by side. If a revised estimate is requested, supply the additional evidence and explain what the new estimate represents.

This also helps when work is shared. A colleague should be able to tell whether the final number came from the tool, a corrected calculation or a deliberate margin. Otherwise, the next person may reuse a planning choice as if it were a fresh prediction.

Review the reason as well as the outcome

Choose a small set of repeated decisions and keep brief notes on important adjustments. Record the original advice, the change, its reason and the outcome relevant to that reason. For forecasts, that might be the eventual observation. For writing, it might be whether the revised text meets the agreed editorial requirements.

Keep revisions that were helpful and revisions that were not. The purpose is to learn when intervention adds value, rather than build a collection of victories for either humans or AI. If the same information is repeatedly missing, improve the input or the task definition before relying on another last-minute correction.

For decisions that also need an independent initial view, the MaryChuks article on preserving judgement before AI advice addresses the sequence of review. The question here comes afterwards: what exactly did the human adjustment contribute?

Meaningful control includes the freedom to accept useful advice unchanged, challenge it for a stated reason or leave a question unresolved. Explaining the adjustment makes that control easier to understand and learn from. It also keeps a satisfying sense of involvement connected to the actual purpose of the decision.

Sources

• Dietvorst, Simmons and Massey: Overcoming Algorithm Aversion, Management Science, 64(3), 1155–1170. Published online 4 November 2016; issue dated March 2018. Author-hosted full text.
• Logg, Minson and Moore: Algorithm appreciation: People prefer algorithmic to human judgment, Organizational Behavior and Human Decision Processes, 151, 90–103, March 2019. Author-hosted full text.