Do not limit recommended objectives solely to duration of stay.

Guess et al. conducted a large-scale experiment on Facebook and Instagram during an election period, changing the default algorithmic feed to reverse chronological order. While usage time and exposure content changed, there were no detectable changes in major political attitudes. [1]

The paper has two Errata. The 2024 correction says a processing error excluded deleted Instagram accounts, affecting descriptive exposure estimates but probably not randomized treatment effects; delinked data could not be recomputed. [5] The 2026 correction addresses carousel-post processing. [4] This article is limited to the record after both corrections.

This result is not a universal conclusion that recommendations do not influence attitudes; it is limited to the study period, consenting users, platform, compared feeds, and outcome metrics. [1]

Changes in exposure, content reading, belief alteration, and behavior are distinct stages of causality. Do not estimate subsequent effects from changes in a single stage.

In ranking experiments, link delivery logs, viewing, memory, attitudes, and behavior to the extent possible. Do not judge that learning or well-being improved solely because viewing time increased.

• Delivery: What became a candidate and what rose to the top
• Viewing: What was actually seen, read, or shared
• Cognition: Whether memory, knowledge, or beliefs changed
• Well-being: Whether time utilization and subjective goodness changed

Levy reported an experiment where users were randomly prompted to subscribe to news media with differing political leanings; exposure to opposing news reduced negative attitudes toward the opposing party but showed no clear change in political opinions. [2]

The same study also indicated that even if users subscribe to opposing media, Facebook's algorithm may make it difficult to supply those posts. [2] Explicit choice to subscribe is not the same as actual exposure via ranking.

Do not define diversity merely by left-right ratios. Define multiple axes including evidence quality, region, experience, expertise, and minority opinions. Do not call increasing low-quality information diversification.

Forced exposure to opposing views may cause backlash or burden. Separate the range users can adjust from public minimum diversity, and test the results.

Observe not only individual attitudes, but also source concentration, creator opportunity, and reach of low-frequency but important information. Do not hide supply-side disadvantages using only user averages.

Allcott et al. reported effects in different directions in a randomized experiment stopping Facebook for a period before an election, including improved subjective well-being, decreased news knowledge, and reduced polarization. [3] These results cannot be summarized as simply good or bad.

Service cessation is not the same intervention as a ranking change. [3] Nevertheless, it provides comparative material showing that maximizing usage volume does not necessarily maximize user well-being.

Even if users attach high value to the service, evaluations may change after cessation. Current clicks or intent to continue do not fully represent long-term preferences.

When usage time decreases, some may increase family interaction while others may become isolated. Investigate not only the reallocation of average time, but what it was replaced with.

Well-being metrics separate happiness, autonomy, regret, sleep, social contact, and information acquisition. Do not treat a single self-report score as a universal metric.

• Usage Value: Benefits perceived by the individual and substitutability
• Time Substitution: Where the reduced time moved
• Information Benefit: Whether knowledge or acquisition of important information was maintained
• Subjective Well-being: Measure mood, regret, and autonomy separately

Place relevance, reliability, diversity, novelty, user control, and well-being as ranking objectives. Collapsing all into a single weighted score too quickly hides major floor violations.

First, establish a floor to prevent degradation, then optimize other objectives within that range. For example, after keeping low-quality information exposure below a predeclared maximum, choose the balance between relevance and diversity.

Do not let the technical team decide weights alone. Users, domain experts, safety officers, and business owners must explicitly state what they consider compensable.

When changing objective weights via personalization, distinguish between user choice and estimation. Do not secretly alter the information environment by estimating political leanings or vulnerabilities.

Measure short-term click improvements and long-term user satisfaction over separate periods. Do not treat temporary effects from novelty as permanent improvements, and verify persistence after cessation.

Recommendations interact among users. Changes in display for one user may reach the control group through sharing or posting, causing effects to mix with individual randomization alone.

Consider allocation by network, region, and time zone, and analyze with interference as a premise. If complete separation is impossible, leave that limitation in the results.

Suppliers may adapt to ranking changes by altering post content. Distinguish initial display effects from supply-side reactions weeks later.

Establish stop conditions if significant harmful exposure increases during the experiment. Do not postpone user protection under the guise of research value.

Provide consent, opt-out, and data minimization. Do not perform unnecessary sensitive estimations under the pretext of abundant behavioral logs.

Even if average engagement is the same, usage time may increase extremely for a minority of users. Check distributions, upper tails, and nighttime usage.

Investigate whether exposure quality differs for youth, vulnerable users, minority languages, and specific political groups among pre-approved groups. Do not treat lack of difference in small samples as equality.

Do not measure filter bubbles solely by individual homogeneity; observe source concentration, topic diversity, and contact with falsification. Ensure metrics are not met by diverse low-quality information.

Even if the average rate of harmful information is low, concentration on important queries during crises is significant. Maintain a stress set weighted by impact rather than usage frequency.

Verify whether the benefits and burdens of personalization accrue to the same individuals. Do not count advertiser or creator profits as user well-being.

Providing feed choices such as chronological, relevance, or diversity is not control if settings are deeply hidden. Test discovery, modification, understanding, and retention.

Do not revert user-selected settings to defaults upon return or update. If reversion is necessary, explain the reason and make re-selection easy.

Show what 'Not Interested' inputs change. Control that affects only a single post, or where it is unclear whether it affects topics, sources, or ads, creates expectation violations.

Explanation screens should show the reason a post appeared and configurable factors, rather than general algorithm explanations. Designs that do not inadvertently disclose sensitive estimations are also necessary.

Do not treat low control usage as evidence of unneeded features. Investigate lack of understanding, modification costs, and weak effects, and audit actual exposure changes after operation.

For each ranking update, compare delivery, viewing, reliability, diversity, attitudes, time, and well-being before and after. Do not approve based solely on business KPI improvements.

External research effects depend on the platform and period. [1] [2] [3] Conduct small-scale experiments with your own user distribution, content, and objective functions, and verify reproduction over a separate period.

If a degradation floor is breached, stop not only the entire model but also identify the problematic objective term, target group, and content category, and revert to the previous version.

In result publication, include major outcomes, pre-registered hypotheses, no difference, opposite direction, attrition, and unmeasured factors. Do not craft success stories using only favorable click differences.

The responsibility of recommendation systems is not to increase what users reacted to. It is to measure the multiple impacts of allocating attention on information quality, choice, autonomy, and well-being, and to be able to stop degradation.