AI Project Advice: Separate Past Effort from the Next Decision

Imagine a creator who has spent weeks building an online course. The lessons remain unfinished, interest is uncertain and other work is waiting. They ask an AI assistant to explain how to make the project succeed. A detailed completion plan might answer that question beautifully while leaving a more important one untouched: is completing this version still the best next step?

The hours already spent deserve an honest record. They do not, by themselves, establish what deserves tomorrow’s hours. When using AI to review a stalled project, separate the history of the work from the decision about what happens next.

What makes a cost sunk?

A sunk cost is a past expenditure that the present choice cannot recover. In their February 1985 paper, psychologists Hal Arkes and Catherine Blumer examined how prior investment could increase people’s willingness to continue. Their studies included hypothetical project decisions and a field experiment involving theatre subscriptions. The paper presented evidence linking this persistence with a wish to avoid appearing wasteful.

That is a research finding about particular decisions, rather than a diagnosis of every unfinished project. Persistence can have good reasons. The question is whether the argument for continuing depends on benefits and costs still ahead, or simply on the discomfort of leaving past expenditure unrecovered.

Using a purchase and extending a project are different

Roth, Robbert and Straus’s meta-analysis, published online on 19 September 2014 and in a 2015 issue of Business Research, examined monetary sunk-cost research. It found evidence of sunk-cost effects whose strength varied substantially across studies. The influence of factors such as elapsed time also depended on the type of decision. These results do not establish an equivalent effect for every kind of emotional or creative investment.

The authors distinguished decisions about using an already-purchased product from decisions about committing further resources to a project. For an AI-assisted review, make that distinction explicit. “Should I use the training I already bought?” and “Should I spend another month developing this course?” involve different choices. A description of yesterday’s payment cannot settle either one.

A recent study adds a useful caution

George Beknazar-Yuzbashev, Sota Ichiba and Mateusz Stalinski’s study in the January 2026 issue of the Journal of Economic Behavior & Organization reported two preregistered experiments. In an online study involving 1,806 participants, varying the price of entry to a real-effort task produced a sunk-cost effect on time spent playing.

In a YouTube field experiment involving 11,328 videos, the researchers varied how long viewers had to wait before skipping a pre-video advertisement. The intervention had no statistically significant overall effect on video engagement; longer waits also prompted more people to leave before the video started. These were usage decisions, not tests of business-project reviews or chatbot advice. They caution against assuming that a familiar psychological effect will operate identically in every setting.

Keep what the past has made available

Ignoring an unrecoverable expenditure does not require pretending the project never happened. Completed materials, acquired skills, audience feedback and remaining commitments can affect the options available today. Their present usefulness belongs in the review. The original amount spent creating them is a different matter.

Suppose an existing lesson can be adapted with little additional work. That may support reuse. Suppose a cancelled launch would leave customers needing a response or refund. Those future consequences also need attention. Calling everything connected with the past “sunk” would hide relevant work still to be done.

One course, three possible next steps

Consider a fictional illustrator who has spent forty hours preparing a course. Three lessons are usable, but the planned membership area is incomplete. The illustrator estimates another twelve hours to finish the original package. These figures are assumptions for this example, not observed results or reliable estimates for another creator.

The illustrator could finish the package, adapt one lesson into a smaller workshop, or pause development while checking what prospective learners need. Each option requires its own estimate of additional work, delivery obligations and useful evidence. The workshop has no guaranteed audience simply because it is smaller; pausing may also delay a worthwhile launch.

The forty hours cannot be reclaimed by choosing any of these options. However, the three usable lessons may make one option more attractive than it would have been from a blank page. A fair comparison preserves that advantage without treating the original effort as a reason to keep every feature.

Give the AI a comparison to organise

My proposed routine is to write a short account of the present situation before requesting advice. It is an editorial application of the research, not an experimentally validated method for removing sunk-cost bias. Include the following:

• Decision: The specific next commitment being considered, with a clear boundary.
• Available assets: Materials, skills and permissions that can still be used.
• Future requirements: Additional work, obligations and foreseeable consequences under each option.
• Evidence and uncertainty: What supports the expected benefit, and what remains an assumption.

Then ask the assistant to compare continuing, changing scope and pausing against the same objective. Request the strongest supported reason for each option, the main unresolved assumption and the evidence that could change the comparison. Leave unsupported demand, costs and completion estimates marked as unknown.

Inspect the response for arguments that rely mainly on what has already been spent. Also inspect arguments for stopping: abandoning a project does not become sensible merely because the phrase “sunk cost” appears in an explanation. Neither a polished defence of persistence nor a polished case for quitting replaces checking the underlying facts.

Review the next choice without rewriting the earlier one

The original decision and today’s decision can reasonably differ because the information and remaining work have changed. The MaryChuks guide to reviewing choices without outcome bias offers a complementary way to examine earlier reasoning without judging it entirely through the result.

Record what you decide now, why, and what would justify reconsidering it. If you continue, name the next bounded piece of work. If you pause, name the question the pause is meant to resolve. If you stop, identify anything still owed and anything worth retaining. Past effort can inform what is available today; the next commitment needs reasons of its own.

Sources

• Arkes and Blumer, The Psychology of Sunk Cost (February 1985).
• Roth, Robbert and Straus, On the sunk-cost effect in economic decision-making: a meta-analytic review (online 19 September 2014; 2015 journal issue).
• Beknazar-Yuzbashev, Ichiba and Stalinski, To the depths of the sunk cost (January 2026 journal issue).