Did you know? People sometimes judge a harmful result more harshly when someone actively caused it than when a person allowed a similar result to occur by doing nothing. Psychologists call this asymmetry omission bias. It is not a rule that every action is worse than inaction, nor that action and inaction are morally interchangeable. It is a warning that the visible difference between doing and not doing can influence judgement even when the outcomes, intentions or foreseeable risks have been carefully held constant.
The practical repair is to describe both options as decisions with consequences. For each, ask what was known, who had responsibility, what would probably happen with and without intervention, and which duties or rights genuinely distinguish the choices. Compare anticipated outcomes before attaching the labels ‘action’ and ‘inaction’, then reintroduce legitimate ethical differences openly rather than smuggling them in through the default. This prevents a quiet choice from escaping scrutiny merely because it produces no dramatic event.
This guide explains the original research, a fully solved fictional risk comparison, the distinction from status quo bias and sunk costs, failures in organisations and learning, and a changed-context exercise. It does not claim that mathematics settles every moral decision. Numerical cases are invented teaching examples; healthcare and safety examples are conceptual, not individual professional advice.
Contents
- What omission bias means—and what it does not
- Why a visible action attracts responsibility
- What the foundational experiments support
- Doing nothing is still a choice when a live decision exists
- Worked example: a fictional service update with unequal risks
- Outcome-equivalent tests: change the frame, not the facts
- Ethical distinctions that arithmetic must not erase
- Committees can hide decisions inside inertia
- Students: revision choices also contain hidden omissions
- Omission bias versus status quo, outcome bias and sunk costs
- Eight common mistakes when trying to correct the bias
- A practical symmetric decision worksheet
- Changed-context exercise: a classroom accessibility decision
- Limitations: not every act and omission are equivalent
- Frequently asked questions
- Reader pathways and original research
What omission bias means—and what it does not
Omission bias describes a tendency in relevant settings to regard harmful outcomes from an action differently, often more negatively, than equivalent or even worse outcomes caused by allowing events to proceed. The observation is about judgement or preference, not a claim that objective consequences cease to matter. Context, responsibility, norms and intent can modify the effect.
The original experimental paper by Mark Spranca, Elisa Minsk and Jonathan Baron, Omission and Commission in Judgment and Choice (1991), compared scenarios involving omissions and commissions while examining the roles of intentions, motives and consequences. That work established an important research foundation for the distinction. It did not say that every person always prefers inaction or that all choices should be evaluated by a single outcome number.
The word commission here means an action that contributes to a result. Omission means not taking an available action. An omission can be an explicit decision, an unnoticed failure, a policy default or a loss of opportunity. These are not morally or legally identical. A person who deliberately withholds a promised safety inspection differs from a passer-by who did not know a problem existed.
The bias becomes clearest when the outcome and relevant background are held comparable. If choosing action changes the magnitude or distribution of risk, a preference may be rational rather than biased. Similarly, some ethical duties legitimately distinguish doing harm from allowing harm. A careful explanation must identify what the psychological experiment held constant and what real-world decision-makers have grounds to treat differently.
The useful question is therefore not ‘Is inaction wrong?’ but ‘Would my judgement change if the same responsibility and expected consequences were described without privileging the inactive option?’ This is an intellectual check, not an automatic answer to every case.
Why a visible action attracts responsibility
An action creates a visible moment of choice. Someone clicked a button, signed an approval or changed a policy, and observers can imagine the alternative where that step did not occur. A harmful result following action can therefore feel strongly attributable to the person who acted. A harmful result after inaction may appear to flow from the background, even when the decision not to intervene was informed and consequential.
Causal stories are not the same as full responsibility analysis. A repair technician may cause a brief service outage while preventing a much larger later failure; the outage is visibly linked to the repair. The silent decision to leave an unsafe component untouched may have a less salient causal story, although the foreseeable longer-term consequences can be greater. This visibility asymmetry can shape judgement.
Counterfactual imagination also matters. People can readily ask ‘What if the person had not pushed the button?’ after an active intervention. They may ask less readily ‘What if the person had taken the precaution?’ after an omission. The two questions need to be made equally explicit if the decision is to be assessed fairly.
That does not excuse active harm or erase genuine constraints. The directness of causation, authority, consent, trust and duties to act can supply legitimate moral differences. A bias diagnosis is appropriate only when the extra negative judgement cannot be justified by those relevant differences. Otherwise the same observation could represent principled ethical reasoning rather than a cognitive error.
An intelligent decision process keeps two ledgers: factual consequences and responsibility conditions. Separating them helps people find when an intuition about visible agency is doing too much work and when it correctly points to a real obligation.
What the foundational experiments support
Spranca, Minsk and Baron used hypothetical decision scenarios to investigate how participants evaluated harmful results linked to actions and omissions. The work is frequently cited as a foundational investigation of omission bias. As with any experimental psychology finding, the strength of an effect depends on task construction, participant population and which facts are held equal across alternatives.
A related issue is status quo bias, where the existing option gains preference partly because it is already the default. William Samuelson and Richard Zeckhauser investigated status quo bias in their original 1988 research article. The two phenomena can overlap but are not synonyms. A person may prefer to keep a current system because change feels risky without evaluating equivalent harm from acts and omissions in an experimental scenario.
Later work has examined how normality, causation and individual differences affect omission judgements. It would be careless to present the original asymmetry as a universal fixed constant, such as a claim that omissions always feel twice as acceptable. The literature involves contested mechanisms and context-dependent preferences, not a single conversion factor from moral feeling to objective harm.
Contemporary research also uses paired framings to evaluate decision systems. An October 2026 preprint, OMIT the Action: Measuring Framing-Invariant Omission Bias under Philosophical Disagreement, studies bias in language-model moral reasoning using paired cases. Its findings belong to its sampled models and benchmark design, and it is a preprint rather than universal evidence about all humans or AI systems.
The collective lesson from the research is methodological: control the relevant differences, test framing sensitivity, and separate an observed preference from an explanation of why it arose. No paper authorises an assistant to impose one ethical philosophy on a real person.
Doing nothing is still a choice when a live decision exists
Sometimes people treat the default as an absence of decision. A software team does not deploy a patch; a school does not update an outdated learning resource; a committee declines to review a known access problem. In each case the inaction may be deliberate or automatic, but the world continues to change. If the team had authority and sufficient warning, the option to continue unchanged can be a consequential policy choice.
Yet we should not infer awareness from outcomes. A person who never learned of a problem cannot be evaluated exactly as an owner who reviewed the evidence and consciously rejected a feasible repair. A thorough judgement asks what information was available, what action was realistically possible, and who held responsibility at the relevant time.
The default can also be genuinely safer. In a complex system with limited evidence, an intervention may introduce uncertain side effects. Conservative practice may be justified by the risk of irreversible harm. The bias check is not a command to intervene quickly; it is a prompt to compare risks instead of granting one option moral immunity because it requires fewer visible steps.
An intelligent organisation makes important defaults explicit. It documents what will happen if no new approval is given, who reviews the standing condition, and which new evidence triggers reconsideration. This makes an inactive option measurable and accountable without treating every decision as an emergency.
The same approach helps students. Leaving a difficult topic unrevised has an opportunity cost, but that cost depends on the exam timing, alternative use of study time and current understanding. A learner should compare the expected benefit of another revision session rather than assume either action or rest is always the better choice.
Worked example: a fictional service update with unequal risks
Imagine a fictional internal service running on 1,000 identical units for a defined month. Engineers estimate, under strong simplifying assumptions, that leaving a known fault uncorrected creates a 4% incident probability per unit. Applying a proposed update creates a 1% incident probability per unit, including update-related failures. These figures are invented and should not be treated as estimates for real systems.
Under the model, expected incidents if the service remains unchanged equal 1,000 × 0.04 = 40. Expected incidents after the update equal 1,000 × 0.01 = 10. The expected difference is 40 − 10 = 30 fewer incidents per 1,000 units, or an absolute reduction of three percentage points. The risk ratio is 0.01/0.04 = 0.25, meaning the update’s modelled incident risk is one-quarter of the no-update risk under these assumptions.
Suppose the team worries especially about an incident that would visibly occur during deployment, while a later fault after doing nothing feels like background bad luck. That emotional reaction could reflect omission bias. It does not prove the patch should ship. The probabilities may be wrong, incident severity may differ, and the patch may have other irreversible effects. Responsible evaluation must include those facts.
If the update incidents are severe while fault incidents are mild, expected incident counts alone are inadequate. Let a severe update-related incident cost ten illustrative harm units and an unchanged-service incident cost one harm unit. With probabilities as given, expected harm per unit under the update would be 0.01 × 10 = 0.10 units, versus 0.04 × 1 = 0.04 units under inaction. Under that changed severity assumption, inaction has lower modelled expected harm. The recommendation reverses because the factual consequences changed, not because one option was labelled action.
The right lesson is to fix outcomes, probabilities, severity, duty and alternatives before examining the psychological effect of the action label. Omission bias may distort attention, but corrected numerical comparisons can themselves be wrong when assumptions are poorly grounded. Use the model as a questioning aid, not a substitute for evidence.
In real service engineering, rollout stages, reversible toggles, independent tests and stakeholder approval can alter risk rather than forcing a binary act-or-ignore choice. Better option design often matters more than choosing whichever side looks morally tidy in a simplified scenario.
Outcome-equivalent tests: change the frame, not the facts
To look for framing sensitivity, construct two descriptions that preserve the relevant factual outcome while switching which path is presented as an active intervention. For instance, a committee might face two information-display policies that have identical measured error rates under the toy model. In one description, the committee must click to move to Policy B; in another, Policy B is already the default and staying requires no click. If preferences change sharply, the action/default frame may be influential.
Such a test must be designed carefully. The default might convey information about recommendations, switching costs or legal status, so it is not always an empty label. An appropriate comparison makes those variables visible, randomises framing where feasible and checks participant understanding. Otherwise the result can be mislabelled as cognitive bias when the participants inferred rational information from the default.
Suppose an invented experiment randomly gives 100 readers a neutral frame and 100 a frame that calls the same option the existing default. In the neutral frame, 48 of 100 choose it. Under the default frame, 64 of 100 choose it. The difference is 64% − 48% = 16 percentage points, with identical sample denominators. This is only a constructed numerical example showing how a measured framing contrast would be reported.
The observed difference does not automatically identify omission bias. It could be status quo preference, inferred endorsement or other effects unless the design separates those mechanisms. Statistical uncertainty also matters; a single toy sample is not a universal population effect. A sound study documents its randomisation, exclusion rules and exact wording so alternative explanations can be tested.
In everyday decisions, a mental version of the test can still help. Describe the options neutrally, ask an independent reviewer to compare outcomes and responsibilities, then reveal which is the current default. If the judgement changes, investigate why. The answer might be a legitimate switching cost or a cue that had been overlooked.
Ethical distinctions that arithmetic must not erase
Consequences matter, but so do rights, consent, fiduciary obligations, promises and distribution of harm. A system that reduces aggregate expected incidents might expose a small group to a severe involuntary risk. This is not adequately addressed by a single average. Ethical evaluation must consider who bears benefits and costs and which actions are permitted.
An omission can violate a duty. If a professional is charged with maintaining a safety system, knowingly failing to inspect a documented hazard may be ethically different from a stranger’s failure to intervene. Conversely, acting without consent or authority can itself be wrong even when an intervention seems likely to produce a benefit. Responsibility is part of the evidence, not an inconvenience to be stripped out.
People also differ in their moral views about direct causation and intentional action. Two thoughtful people can agree on the facts and still weigh duties or harms differently. Omission bias research studies judgement asymmetries; it does not establish that consequentialist arithmetic is the only valid moral framework.
The most useful repair is therefore a structured comparison rather than a moral shortcut. Record expected outcomes, uncertainty, relevant permissions, duties to act, duties not to act, reversibility, accountability and affected parties. Ask which considerations remain if the same facts are presented without emotionally loaded action verbs.
When stakes involve health, safety, law or rights, defer to qualified decision processes and applicable rules. The guide is educational: it teaches readers to notice framing and demand evidence, not to override professional safeguards based on an abstract example.
Committees can hide decisions inside inertia
Organisations often govern visible changes more formally than unchanged operations. A proposal for a new policy may require signatures, risk reviews and recorded responsibility, while continuing a questionable old policy requires no meeting. This asymmetry can leave substantial risk outside the approval process. The problem may be structural as well as psychological.
A helpful governance practice is to schedule review of important standing choices. If a service, curriculum or safety procedure has not been re-evaluated for a defined period, an owner examines whether relevant evidence has changed. The review does not force a change; it makes the continuation decision explicit and documentable.
Teams should also avoid turning accountability into blame for visible action alone. If a carefully controlled intervention fails, the fact that someone signed the change should not automatically dominate an assessment of the informed decision. The companion outcome bias guide explains why fair reviews consider information available before the outcome was known.
Record what evidence supported the choice at the time, the available alternatives, rejected risks and the point at which it should be reconsidered. A written decision trail does not make every judgement correct, but it reduces selective memory and clarifies the difference between a justified conservative choice and neglect through default.
Where decisions recur, an organisation can compare similar active and inactive cases in its incident reviews. The aim is not a target number of interventions. It is to evaluate comparable risks consistently, including harmful omissions that are easy to overlook because no new action was taken.
Students: revision choices also contain hidden omissions
Consider a student who chooses to keep revising a familiar chapter because it feels safe. Attempting a difficult unfamiliar topic creates immediate visible errors and discomfort. The student may judge that active choice as evidence of doing worse, while the lost opportunity from ignoring the weak topic remains invisible until a later assessment. This is not a literal replica of the original omission-bias experiment, but it is a useful analogy for making the consequences of non-action visible.
A tutor can ask the student to compare what will happen under two feasible plans. One plan spends forty minutes polishing mastered questions; the other spends forty minutes diagnosing errors in a poorly understood concept. Estimate expected learning benefits from actual error patterns, then test later independent questions rather than assuming the uncomfortable plan must win.
The alternative of rest must also remain legitimate. Study fatigue, competing deadlines and personal needs affect how productive another practice block will be. A decision tool that labels all non-study time as an omission would misunderstand the problem and could harm learning. The correct analysis includes realistic costs and benefits of each available option.
Parents can help students notice unchosen paths without shaming mistakes. ‘What happens if we leave this topic until next week?’ is more constructive than ‘Why did you refuse to work?’ The first asks for a forecast and constraints; the second can invite defensive explanations. The goal is an informed plan whose progress is independently checked.
Education that transfers to adulthood teaches learners to notice both visible mistakes and invisible opportunities foregone. The education transfer framework connects accurate reasoning with practical decisions beyond a single exercise.
Omission bias versus status quo, outcome bias and sunk costs
Status quo bias privileges an existing state. It may reflect psychological comfort with the default even when no specific harmful act is involved. Omission bias concerns a judgement asymmetry between causing and allowing outcomes. A default is often an omission, but an active choice can preserve the status quo, and failing to act can sometimes require an explicit decision. Do not use the labels interchangeably.
Outcome bias evaluates a decision too strongly through its eventual result instead of the information and reasoning available when it was made. An unlucky but well-supported change can be condemned after a visible failure; a poor decision to do nothing can escape scrutiny when harm does not happen to occur during the review period. A fair process looks at both the forecast and outcome.
Escalation of commitment or sunk-cost effects concern continued investment because resources have already been spent. They can coexist with omission bias: a team may resist ending an old project because it has invested much and may also fear the responsibility of actively cancelling it. These are different causal hypotheses and invite different questions.
The separate escalation of commitment guide treats stopping rules and sunk costs. Omission-bias repair adds a symmetric review of acting and not acting with respect to harms and duties. A team that recognises only one named bias can still miss a different mechanism in the same decision.
Rather than diagnose people by label, ask what evidence would distinguish the hypotheses. Do they change preference when the default is reversed? Do they assess the same action differently after learning the outcome? Does prior investment affect their decision even when future benefits are identical? Clear tests are more useful than a fashionable bias vocabulary.
Eight common mistakes when trying to correct the bias
The first mistake is assuming every refusal to intervene is omission bias. It may reflect risk, ethics, insufficient authority or limited information. The repair is to inspect what actually differs between the options. The second is treating expected outcome counts as complete moral answers. Include severity, rights, consent and distribution.
The third mistake is describing only the active option’s risks. If the inactive option’s possible harms are omitted, the comparison is not balanced. The fourth is comparing an intervention’s worst case with the status quo’s average case. Use matched time horizons, denominators and uncertainty assumptions.
The fifth is calling all inaction harmless because no visible event followed. Quietly continuing a failing policy can accumulate costs. The sixth is judging an action by an unlucky outcome while judging inaction by whether it avoided a public scandal. Use records of forecasts and decision quality at the time.
The seventh is replacing inaction preference with automatic action preference. Moving every problem immediately to intervention can introduce new risks. A corrected decision process should choose whichever option survives fair comparison, not simply invert the initial bias.
The eighth is generalising a laboratory phenomenon to a person without relevant evidence. People vary, and real choices contain more details than short experimental vignettes. Treat the research as a reason to investigate framing, not as permission to dismiss another person’s ethical view as irrational.
Across all eight, accuracy improves when the reviewer spells out the same variables for every option: scope, authority, foreseeable results, probabilities, harms, benefits, affected groups and feasible alternatives.
A practical symmetric decision worksheet
Begin by naming the decision date, the owner and the plausible options, including continuing unchanged. For each option, write a neutral description of what will occur next. Avoid language that makes one path sound natural and the other reckless before evidence is assessed.
Next record what is known and what remains uncertain. Separate observations from predictions, and predictions from preferences. If numbers are used, state the population, time horizon, denominator and whether outcomes are mutually exclusive. Compare like with like rather than combining a one-week risk with a one-year risk.
Then list actual obligations and restrictions. Who can authorise a change? Who must be protected? Which harms can be justified or consented to, and which cannot? Some choices should be excluded before expected-value calculations because they are not permitted.
Perform a frame reversal. Imagine the status quo were the alternative and the proposed intervention were already established. Would the same consequences still warrant your choice? If not, identify whether a genuine switching cost or an unexamined preference for the default caused the difference.
Choose a review point and evidence to collect. An initial decision may be appropriate under today’s information but should change if new facts arise. Record what would trigger reconsideration so that doing nothing does not become a permanent, invisible assumption.
Finally communicate the result truthfully: ‘We compared intervention and non-intervention, considered rights and uncertainty, and selected this policy under these conditions.’ A transparent conservative decision can be intelligent; an undocumented passive default can be fragile.
Changed-context exercise: a classroom accessibility decision
Imagine a fictional school website has a navigation pattern that some students find difficult. One proposal changes the interface immediately; another retains the present design while further testing is completed. Staff fear that deploying a change will be blamed if one student struggles with the new interface. Yet leaving existing barriers unaddressed may also cause foreseeable harm.
Construct a fair comparison. Suppose an illustrative pilot of 100 comparable tasks suggests 12 failures under the existing interface and 7 after a tested change, under the same task definition and observation conditions. These numbers imply 12% versus 7%, an absolute difference of five percentage points. They do not demonstrate causality without a suitable study design or tell us whether failures have equal severity.
Now add a relevant constraint: the proposed change interferes with a particular assistive-technology workflow used by a small group. A total of seven failures can still be unacceptable if concentrated among users whose access rights must be protected. The right next step may be to revise the design, test the affected workflow, or use a reversible rollout. Neither ‘keep the old one’ nor ‘ship quickly’ is an automatic ethical answer.
Finally swap the description so that the new design is called the existing default and the old design is called the proposed change. Does the evaluation shift before the actual user evidence changes? If so, investigate whether responsibility and familiarity are distorting the choice. Then reapply the accessibility requirements and measured outcomes.
The transfer answer is a process: identify foreseeable harms on both sides, make legitimate duties explicit, correct the design where possible and evaluate with suitable evidence. That is how the bias concept becomes useful without pretending that one numerical rate settles a complex human decision.
Limitations: not every act and omission are equivalent
Real decisions are rarely perfect pairs of equal outcomes and equal responsibilities. Intentional harm, informed consent, duty of care, rights and reversibility can produce genuine distinctions between action and omission. A laboratory preference asymmetry should not be used to erase them.
Psychological findings also have sampling and methodological limits. Effects can depend on wording, normality, causal attribution and individual differences. There is no scientifically justified single percentage by which everyone overvalues inaction. Avoid turning a qualitative tendency into a universal formula.
Quantitative examples here are fictional. Probabilities and harms must be estimated from appropriate domain evidence before they inform real safety, health or legal decisions. An arithmetic answer can be exact for invented assumptions while its applicability to reality remains unestablished.
A bias label should never replace respectful dialogue. Another person may hold an ethical principle about causing harm that deserves serious discussion rather than a claim that their preference is just a cognitive defect. The value of the research is to help us test whether facts or framing drive a judgement.
Finally, the opposite error is possible: a person can feel compelled to act even when waiting is more prudent. Good intelligence is not maximum intervention. It is evidence-based, accountable choice among feasible options under the relevant responsibilities.
Frequently asked questions
What is omission bias? It is a tendency in certain contexts to judge or prefer consequences differently depending on whether harm is caused by an action or allowed through inaction, even when key facts are held comparable.
Does omission bias mean doing nothing is always wrong? No. Inaction may be ethically justified or practically safest. The bias check asks whether comparable outcomes and responsibilities are being evaluated consistently.
How is omission bias different from status quo bias? Status quo bias favours an existing option because it is the default. Omission bias concerns the act-versus-allow asymmetry in judging harm. The two can overlap but are not identical.
Can an omission create responsibility? Yes, when a person has the relevant knowledge, ability and duty, failing to act can be consequential. The nature of responsibility still depends on the circumstances and applicable rules.
Does expected-value arithmetic tell us the moral answer? No. Probabilities and average harms do not alone settle consent, rights, duties, unequal risk or irreversibility.
How can a student use this idea? Compare the expected consequences of practising or postponing a topic, make time constraints explicit and test actual learning later rather than relying on how comfortable the default feels.
How can a team test its process? Review comparable active and inactive choices with matched evidence, document forecasts, and use blinded or reversed descriptions when appropriate to reveal framing effects.
Is every inactive decision an example of the bias? No. A bias explanation needs evidence that the inactive framing influenced judgement beyond legitimate differences.
What research established it? The 1991 Spranca, Minsk and Baron experiments were influential; later work studied causes, individual variation and distinctions from status quo preference.
Can an AI show omission bias? Researchers have begun investigating action-versus-inaction framing in language-model decisions, but any claim about a particular model requires testing on a relevant paired benchmark.
Reader pathways and original research
If you are a student, take one hypothetical choice and write equal-length accounts of acting and not acting. Calculate any comparable expected outcomes, then describe what additional moral facts a calculator cannot provide. This exercise teaches both quantitative discipline and respectful ethical reasoning.
If you are a teacher or parent, ask what will change under the status quo and under an intervention, rather than treating inaction as a zero-cost baseline. Keep examples age-appropriate and avoid presenting one philosophical conclusion as mandatory. The learning-transfer framework connects responsible evidence use with everyday choices.
If you lead a team, make significant continuation decisions explicit, revisit them at agreed checkpoints and keep the evidence available for fair review. Related decision mechanisms include outcome bias and escalation of commitment.
Primary sources: Spranca, Minsk and Baron (1991), Omission and Commission in Judgment and Choice; Samuelson and Zeckhauser (1988), Status Quo Bias in Decision Making; and the clearly labelled October 2026 preprint OMIT the Action, which studies language-model framing and is not interchangeable with the human experimental literature.
The fictional numbers here demonstrate how to reason, not how a specific institution should act. Continue through How Intelligence Works for the broader human and collective intelligence pathway.