Is a refundable payment a sunk cost?

The recoverable amount is relevant if recovery depends on the choice.

Is this the same as escalation of commitment?

They overlap; sunk costs can contribute to escalation, which can have additional causes.

Should an AI recommend cancellation whenever a project loses money?

No. The comparison concerns future options and consequences, not past losses alone.

AI Glossary / Governance, ethics and safety / Advanced

Sunk Cost Fallacy Bias

Definition

Sunk cost fallacy bias is the tendency to let irrecoverable past expenditure justify continuing an activity, even when its remaining costs and benefits do not support that choice. The expenditure can involve money, time, effort, or other resources. In an AI project, the bias may appear when a team argues that it must deploy or keep developing a system because it has already invested heavily in it.

What it means

A sunk cost cannot be recovered by choosing differently now. Spending more does not reverse that earlier expenditure. The current decision should compare what the available options can change from this point onward.
This does not mean every struggling project should stop. Previous work can create useful assets, knowledge, or a nearly finished capability that makes continuation worthwhile. Those present advantages matter through their effect on future outcomes. The mistake is treating the historical expenditure itself as a benefit of continuing, without a defensible forward-looking reason.

How it works

People may find it uncomfortable to acknowledge that earlier effort will not deliver its intended result. Arkes and Blumer's research describes increased commitment following prior investment and examines the desire to avoid appearing wasteful.
In organizations, the pattern can be reinforced when project discussions emphasize how much has been spent while giving little attention to changed requirements or alternatives. NIST identifies sunk cost fallacy as relevant to AI development decisions.
A useful review separates historical expenditure, remaining commitments, recoverable value, and costs that differ between options. Ask what evidence would justify the next stage if the same current assets were available without the emotional history. Preserve lessons from the past while evaluating the next commitment on its own merits.

Three practical examples

A company has paid to label documents for an AI extractor. New testing shows that its intended workflow is unsuitable. Labeling expenditure alone does not justify deployment; reusable data and alternative uses deserve separate evaluation.
A developer spends weeks building a complicated assistant. A simpler process now meets the requirement. Continuing solely to avoid admitting that the earlier work was unnecessary illustrates the bias.
A team retains an unreliable model because many staff attended training. Future retraining and transition costs are relevant, but the hours already spent in completed sessions cannot be recovered by keeping the model.

Think of it this way

Suppose you have walked a long distance toward a bridge and then learn that the bridge is closed. The distance already walked explains how you arrived there, but it does not make the bridge usable. Your next route should reflect the conditions ahead, including the real cost of turning back or taking a detour.

Worked example

Consider a fictional internal AI project that has already consumed 120,000 monetary units, all irrecoverable. For a simplified comparison, assume all future amounts use the same time horizon and valuation basis, and ignore taxes and financing effects.
Option A continues the original project. It requires another 40,000 and is expected to generate benefits worth 25,000. Its future net value is 25,000 minus 40,000, or negative 15,000.
Option B stops new development and uses an existing alternative. Transition costs are 8,000, and expected benefits are 18,000. Its future net value is positive 10,000.
On these assumptions, B is better by 25,000. Including the already spent 120,000 in both totals gives negative 135,000 for A and negative 110,000 for B; the ranking is unchanged. Saying that A must continue to recover the earlier expenditure does not improve its future value.
Now suppose credible new evidence raises A's expected future benefits to 70,000. Its future net value becomes positive 30,000, exceeding B's 10,000. Continuing could then be justified. The reason is the changed prospective benefit, not the sunk expenditure. These invented estimates illustrate decision logic, not a guarantee that project benefits can be forecast precisely.

Illustrative case study

Fictional company MapleRecords commissions an AI document classifier. After a long pilot, its workflow changes and many of the intended categories become obsolete. The project sponsor initially argues that canceling would waste the team's months of work.
A review group identifies three possibilities: continue the original design, adapt a reusable component, or retire the prototype. It compares remaining effort, expected usefulness, transition needs, and uncertainties.
The group selects a smaller adaptation because its expected future value is strongest under the documented assumptions. It acknowledges the historical expenditure openly instead of using it to justify the original scope. This fictional case shows that avoiding sunk cost bias can support modification rather than automatic cancellation.

Common misconceptions and mistakes

“All previous work is irrelevant.” Lessons and usable assets can affect present capabilities and future costs; irrecoverable expenditure is a different issue.
“Stopping means the initial decision was irrational.” New information can change a previously reasonable decision.
“Persistence proves sunk cost bias.” Persistence may be justified by current evidence. Examine the reasoning, not the outcome alone.
“Cancellation costs are sunk because the contract already exists.” A cost that differs depending on today's choice belongs in the comparison.
“Only money counts.” Irrecoverable effort and time can also influence commitment.

Business use

Use the concept in project reviews, vendor renewals, and decisions about replacing an AI workflow. Require a clear statement of what the next commitment is expected to achieve and what alternatives are available. Separate the original business case from the current one.
Record uncertainty and revisit important assumptions. Reviewers should be able to acknowledge learning without treating every change of direction as a personal failure. A decision log can preserve accountability for earlier choices while supporting a better next choice.

Check your understanding

Why does the initial 120,000 not change the worked example's ranking?

Answer: It is irrecoverable under both choices.

Why might continuation become sensible?

Answer: New evidence can improve its expected future value.

What should a reviewer ask after hearing “we have invested too much to stop”?

What remaining benefits and costs justify the next commitment compared with alternatives?

Related terms

Opportunity cost
Confirmation bias
Decision analysis
AI governance
Escalation of commitment
status quo bias
project evaluation

Learn next

Take a hypothetical stalled project. Separate irrecoverable expenditure from remaining costs, recoverable assets, and expected benefits. Compare continuation, adaptation, and stopping using the same assumptions.

Sources

NIST, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, sunk cost fallacy glossary entry: https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf Arkes and Blumer, The psychology of sunk cost, publisher abstract: https://www.sciencedirect.com/science/article/pii/0749597885900494 OpenStax, Identify Relevant Information for Decision-Making: https://openstax.org/books/principles-managerial-accounting/pages/10-1-identify-relevant-information-for-decision-making

Previous: Sunburst chart Next: Sunrise problem