The AI You Don’t See: Hidden Ways AI Influences Your Everyday Decisions

An automated decision does not always announce itself. A feed orders what you see. A lender evaluates an application. A retailer decides which products to place first. The important question is not whether every such system deserves the label AI. It is how a system changes your options, and what happens when its judgment is wrong.

The effects can be helpful: less time searching, fewer irrelevant recommendations, more consistent processing. They can also be consequential when the objective is poorly chosen or the system leaves people unable to challenge a decision. We need a way to discuss those differences without treating every algorithm as either an invisible villain or a neutral convenience.

A ranking is a choice about attention

YouTube’s 2021 explanation of its recommendation system describes multiple signals, including viewing behaviour and satisfaction surveys, alongside safety-related decisions. This is the platform’s account of its own system, rather than an independent assessment of how well it works.¹

The example shows why a single phrase such as engagement algorithm is inadequate. A service can pursue several objectives, and those objectives can conflict. What people click immediately may differ from what they later consider worthwhile. An organisation’s stated intention does not tell us whether its trade-offs are appropriate for a particular person or topic.

For users, a practical question is whether the service offers meaningful alternatives: a subscriptions view, a chronological option, controls over recommendation history or the ability to dismiss a suggestion. These controls vary by product. When they exist, test whether they change the experience rather than assume that a setting’s presence guarantees control.

Consequential decisions require more than convenience

Imagine a hypothetical applicant who is refused credit and receives only a generic explanation. The issue is not simply that software was involved. It is whether the explanation identifies the relevant reasons, whether the information was accurate, and whether there is a usable route to challenge an error.

US Regulation B contains specific adverse-action notification requirements for covered credit decisions, including requirements concerning reasons or the right to obtain them. This example is jurisdiction- and context-specific; it is not a universal entitlement to inspect every algorithm used by any business.²

The broader policy question is who carries the burden of discovering an error. If an organisation benefits from automation, I would argue that it should also provide a practical process for correcting the information and reviewing consequential outcomes. A support page that never reaches someone able to act is a weak substitute for that process.

Personalisation is not a complete picture of a person

A prediction is built for a purpose. A preference estimate suitable for arranging music recommendations is not necessarily appropriate for evaluating someone’s reliability, need or eligibility. Reusing a profile in a different context deserves scrutiny even when the original data were collected lawfully.

Consider another hypothetical example: a person looking up an unfamiliar health condition begins receiving a narrow set of recommendations. We should not assume that the resulting feed accurately represents their beliefs or needs. The relevant questions concern how the system uses recent activity, how quickly it updates and whether the person can change that representation.

These examples are illustrations, not accounts of named people. Their purpose is to show which questions a consequential automated choice should make answerable.

What users can ask

Start with the decision, rather than the technology label. What is being ranked, recommended or refused? Which information may be wrong? Does the organisation explain the result? Is there a human review process with authority to change it? Keep a record of consequential notices and use the official correction or complaint routes available for that service and jurisdiction.

For everyday recommendations, review available privacy and history settings, compare alternative views and seek information outside a single feed. These are habits for reducing dependence, not guaranteed ways to defeat an algorithm. A person should not have to become a technical expert merely to participate in ordinary life.

What organisations should make possible

My proposed standard is simple: identify consequential automated uses, explain their purpose in ordinary language, test failures in the actual workflow, and give affected people a usable route to correction. The explanation should distinguish what the system predicts from what the institution decides to do with that prediction.

The same standard should include restraint. More data are not automatically better for the person being evaluated. If a decision can be made fairly with less information, collecting additional information needs a justification beyond the possibility of improving a score.

Invisible systems become easier to govern when we make their decisions visible. The goal is useful automation whose purposes, limits and remedies people can understand—and whose operators remain responsible for the consequences.

Sources and notes

1. YouTube, On YouTube’s recommendation system (September 15, 2021). The platform’s own account, not an independent audit or a guarantee of current implementation. ↩
2. Consumer Financial Protection Bureau, Regulation B, 12 CFR §1002.9, notifications. Specific US credit-notification requirements, not a universal right to an explanation of every algorithm. ↩