Post
JA EN

Don't Let AI Worsen Choice Overload: The Psychology of '2-3 Candidates' Prompt Design

Posted
By
15 min read
Don't Let AI Worsen Choice Overload: The Psychology of '2-3 Candidates' Prompt Design
  • Target audience: IT engineers, engineering managers, and knowledge workers who feel exhausted by AI-listed options yet still can’t decide
  • Prerequisites: Experience using chat-based AI (ChatGPT/Claude) for work
  • Reading time: ~8 minutes

Overview

“I asked the AI for help picking a tech stack, designing feedback for a team member, or planning next quarter’s priorities—and it dutifully handed back ten neatly formatted options. I still couldn’t choose.” The more we use AI, the more familiar this feeling becomes. It hits hardest in management territory, where the right answer depends on context, and an enumerated list only makes the paralysis worse.

What we’re seeing is most likely a classic psychological phenomenon—the “paradox of choice”—replaying itself in a new form during the AI era. In Iyengar & Lepper’s (2000) jam study, a display of 24 jams attracted more browsers than a display of 6, but the smaller display produced roughly ten times the purchase rate1. Having “more” options can pull motivation, satisfaction, and decision rates downward. Effects vary by context and individual, but choice overload is one of the most reliably observed contributors to decision stalls.

The crucial point here is that AI is double-edged: it can expand the option set or shrink it23. The default behavior leans toward exhaustive enumeration, and unless you explicitly instruct the AI to “narrow it to 2 or 3 good-enough candidates with their trade-offs,” the human-side overload doesn’t go away.

This article walks through (1) the mechanism behind choice overload, (2) what’s special about the AI era (when AI recommendation helps and when it doesn’t), (3) prompt design principles for “2-3 candidates,” and (4) how to use AI as a sounding board to mitigate hallucinations and bias. The goal is a practical mindset for treating AI not as “a partner that handles everything” but as “a psychologically optimized complementary tool.”

1. What is choice overload? The jam study and Schwartz’s framework

Choice overload research became widely known through Iyengar & Lepper’s (2000) field experiment. At the same upscale supermarket, they compared days when 6 jams were offered for tasting versus days when 24 were offered. The larger display drew more browsers, but the percentage who actually bought was roughly ten times higher with the smaller display (30% vs. 3%)1.

Psychologist Barry Schwartz formalized this phenomenon in his book The Paradox of Choice (2004)4. Schwartz argued that an increased option set lowers satisfaction through:

  1. Higher decision costs (cognitive load of comparing and evaluating)
  2. Salience of opportunity costs (regret over the paths not taken)
  3. Maximizer-style searching (“there must be something better out there”)

Follow-up research has shown that people with stronger maximizer tendencies are especially prone to post-decision regret as the option set grows5.

That said, the effect is context-dependent. A meta-analysis reports that choice overload does not occur uniformly across all situations; effect sizes vary substantially with the difficulty of the choice, domain expertise, and clarity of evaluation criteria6. There is no simple threshold like “6 is fine, 24 is too many.”

2. The AI era’s twist: when recommendation helps, and when it backfires

In principle, AI-assisted recommendation can ease choice overload through filtering, summarization, and ranking2. But that easing is conditional.

Kim et al. (2023) ran several experiments comparing ChatGPT recommendations against human-curated lists and observed that even with around 60 options, ChatGPT-recommended choices showed higher satisfaction and stronger preference7. The interpretation is that trust in AI breaks the traditional “more options → fatigue” loop.

The catch: that holds only while the AI is trustworthy and the recommendation has actually been narrowed. If, in real work, you toss vague prompts at ChatGPT or Claude—”List the tech stacks I should learn,” “What are the candidate priorities?”—the model tends to slip into exhaustive-enumeration mode. The same paralysis you feel staring at a long paper menu reproduces itself in the chat window.

So AI swings both ways:

  • Tell it to narrow down explicitly → choice overload eases
  • Let it default to exhaustive listing → choice overload regenerates

Schwartz himself, in a 2025 interview, pointed out that AI-driven curation can produce a fresh paradox of its own and emphasized that humans must keep asking questions and scrutinizing the options that come back3.

3. Prompt design principles for “2-3 candidates”

This is the implementation section. Combining psychological findings with practical experience, the following design choices for prompts seem to work well.

3-1. Explicitly cap the count at “2-3”

The simplest and most effective lever is stating an upper bound on the output count.

❌ “Tell me React component design approaches.”

✅ “Narrow it to at most 2-3 React component design approaches grouped by use case. For each, include strengths, weaknesses, and a typical failure mode.”

Why 2-3 and not 1? Psychologically:

  • Just 1: The user feels the maximizer’s unease (“there might be something better”) and ends up re-asking45
  • 2-3: Comparison axes become explicit, and satisficing (the “good enough” strategy) can land
  • 5+: Comparison load grows faster than linearly, and decision cost rises

Working memory research—Miller’s classic “magical number 7±2” and Cowan’s revised 4±1—suggests the number of items humans can hold in mind for comparison is small8. I’m not aware of empirical studies that directly measured the effect of “n-candidate” prompts on decision quality, but from a cognitive-load standpoint, capping at 2-3 is a reasonable default.

3-2. Always ask for trade-offs

A bare list of candidates tends to provoke the follow-up “Which is best?” To prevent that, force each candidate to be paired with the situations where it should not be chosen.

“For each candidate, write (1) when it’s optimal, (2) when it should be avoided, and (3) one line on the typical regret you’d feel after choosing it.”

Eliciting “predicted regret” inside the first response defuses the maximizer’s residual anxiety—the “there has to be something better”—that Schwartz emphasizes4.

3-3. Have the AI ask about decision criteria first

It’s also useful to make the AI ask one return question—”Which criteria should I evaluate against?”—before producing candidates.

“Before you produce candidates, ask me up to three questions that capture the criteria I should answer. After my reply, give me 2-3 candidates.”

This two-turn “question → answer → presentation” structure aligns with the design discussed in Command-Style vs Question-Style Prompts: when the question has a wide interpretive space, let the AI ask first.

4. Mitigating hallucination and bias: using AI as a sounding board

Narrowing prompts have a side effect. Once you force the AI to pick just 2-3, users tend to accept those candidates without verification, even when they’re plausibly wrong. Hallucination bias is hardest to detect when output is concise and confidently delivered.

The pragmatic countermeasure is to use the AI as a sounding board.

4-1. Have the AI critique its own output (in a separate turn)

After the initial response, run the same AI again in a critic role. Picture an explicit turn where the model “beats up its own candidates.”

“For each of the 2-3 candidates you just gave, deliberately attack them. Check for factual errors, outdated information, and stereotypical defaults. List at least two weaknesses per candidate.”

Weaknesses and outdated assumptions that were masked in the first response often surface here. Even within the same conversation, asking the AI to re-evaluate in a different role can partially offset its biases. This is AI’s self-checking, not a human reviewer’s read-through; treat them as separate steps.

4-2. Re-inject your own context

The AI answers based on the information density of the initial prompt. If the first output feels off for your situation, add context rather than rejecting.

“Re-evaluate those three candidates under the following context: 5-person team, TypeScript proficiency, legacy Rails to maintain, 4-week deadline, I’m the only person on maintenance.”

This re-injection can be seen as part of the practice known as Context Engineering. It’s also a way to train yourself to articulate the context you should be aware of, instead of “letting the AI decide.”

4-3. Keep the final call human

This is the core principle. The AI’s 2-3 candidates provide a “frame for thinking,” but the final decision stays with the human. In Schwartz’s 2025 interview, he made the same point: to retain agency in the AI era, we should let AI organize the option space while humans remain in the seat that actually chooses3.

5. Applying it in practice: management scenarios

Here’s how narrowing prompts and sounding-board re-runs play out in concrete work scenarios. Management territory—where “the right answer depends on context” and decisions are easy to regret—is exactly where choice overload bites hardest.

Case A: Picking topics for a one-on-one (1:1)

  • ❌ “What should I discuss in a 1:1 with a team member whose growth has plateaued?” → 10 sample questions
  • ✅ “For next week’s 1:1 with team member A (mid-career, formerly a high performer, output flat for the past six months), narrow the highest-priority topics to at most 2. For each, also describe the situations where it should not be raised and the early signs that the topic is backfiring.”

Case B: Designing feedback for a team member

  • ❌ “How should I deliver code review feedback to a junior engineer effectively?” → 10 best practices
  • ✅ “Junior A has not incorporated the substantive design suggestions from the past three code reviews (the PR is updated, but architectural improvements aren’t reflected). Narrow to at most 2 approaches likely to land in the next review, and for each, list the backfire risks (withdrawal, over-defensiveness, attrition risk).”

Case C: Prioritizing team initiatives

  • ❌ “What should our team’s goals be next quarter?” → enumerated list
  • ✅ “Under the following constraints, narrow next quarter’s top initiatives to 2: 5-person team, balancing legacy Rails maintenance with new feature work, two members at attrition risk, budget down 15% from this quarter, CTO is pushing automation. Also state the opportunity cost of deprioritizing the others.”

Case D: Tech selection (for reference)

The same mold works for tech selection: “Under the following constraints, narrow real-time-communication candidates to at most 2 and append a regret scenario for each: browser-to-server one-way notifications, 20,000 concurrent connections, minimize ops overhead.” That said, the technical domain has plenty of official documentation and recent comparison articles to cross-check, so AI errors are easier to catch. In management, that “answer-key” is much weaker, which is why narrowing prompts pay off relatively more there.

In every case, the same combination drives the prompt: a count cap + explicit trade-offs/regret + context up front. In management especially, always force the backfire scenario to be written out—it stops you from mechanically running whatever the AI suggested.

6. Scaling up: organizational knowledge and internal RAG

Up to here, the discussion has been about what individuals can do through prompt design. Extending the same idea to the organizational level leads to a vision in which past internal documents, 1:1 logs, and retro notes form the corpus of an internal RAG, and a reranker plus a quality score narrow it to “2-3 good-enough candidates.”

Counterintuitively, management knowledge fits internal RAG better than technical knowledge in some respects. Tech information goes stale fast, so old internal answers carry a real risk of being returned as broken candidates9. Principles like psychological safety and how to run a 1:1, by contrast, stay stable on a multi-decade timescale1011.

That said, the harder problem with “score knowledge and auto-demote stale entries” isn’t the reranker implementation itself—it’s the operational design: who guarantees freshness, how feedback is collected, and so on. I wrote that piece up separately in Why Management Knowledge Fits Internal RAG Better: Asymmetry of Obsolescence and the Reality of Knowledge Scoring. For this article’s scope, the priority is mastering the individual step of “extracting 2-3 candidates via prompt.”

7. Caveats: what doesn’t generalize

Finally, the limits of this piece.

  • Individual differences: People with strong maximizer tendencies keep re-searching (“but there might be something better”) even after the option set is narrowed5. This is not a complete fix.
  • Context dependence: For creative-exploration phases or brainstorming, you actually want many candidates. Choice overload is a “decision-phase” phenomenon6.
  • Limits of empirical evidence: I could not locate RCT-scale empirical work directly demonstrating that “2-3 candidate prompts improve decision quality.” The prescriptions here are an extrapolation from psychological findings and cognitive load theory, combined with practitioner experience.
  • AI recommendation is improving: As Kim et al.’s results suggest, in domains where AI recommendation is already trustworthy, overload may not arise even without intentional narrowing7. These prescriptions are most useful in the phase where AI output quality and your evaluation criteria are still misaligned.

I’m not claiming that “2-3 candidate prompts mean nobody hesitates anymore.” But simply opting out of “automatic exhaustive enumeration” reduces decision stalls.

Summary

AI can shrink the option space or expand it. The default tilts toward expansion, so the human side has to design narrowing into the prompt deliberately.

Three principles you can use in practice:

  1. Cap the count: Explicitly state “at most 2-3”
  2. Include trade-offs: Make the AI write a regret scenario for each candidate
  3. Re-run as a sounding board: Self-critique role for review, then add context and re-evaluate

And keep the final decision with the human. AI is more stable—both psychologically and operationally—when positioned as “the entity that builds the scaffolding for a decision,” not “the entity that makes the decision.”

If this topic interests you, the following articles are relevant:

References

References are listed in the order of citation numbers in the body.

  1. When Choice is Demotivating: Can One Desire Too Much of a Good Thing? - Iyengar, S. S., & Lepper, M. R. (2000). Journal of Personality and Social Psychology, 79(6), 995-1006. 【Reliability: High】Peer-reviewed paper. Reports that the purchase rate at the 6-jam display was roughly ten times higher than at the 24-jam display. ↩︎ ↩︎2

  2. The Paradox of Choice, revisited: How AI Can Either Help or Hinder. Or Both. - Chiosso, H. (2025). 【Reliability: Medium】Expert commentary article. Argues that AI can act either to amplify or to reduce choice overload. ↩︎ ↩︎2

  3. The Paradox of Choice in the AI Age - Schwartz, B. (2025) interview. 【Reliability: Medium-High】The originator of the choice overload concept applying it to the AI era. ↩︎ ↩︎2 ↩︎3

  4. The Paradox of Choice: Why More Is Less - Schwartz, B. (2004). HarperCollins. 【Reliability: Medium-High】Expert-authored book. Provides a systematic account of the choice overload mechanism. ↩︎ ↩︎2 ↩︎3

  5. Doing Better but Feeling Worse: Looking for the “Best” Job Undermines Satisfaction - Iyengar, S. S., Wells, R. E., & Schwartz, B. (2006). Psychological Science, 17(2), 143-150. DOI: 10.1111/j.1467-9280.2006.01677.x. 【Reliability: High】Peer-reviewed study showing that maximizers obtain objectively better outcomes (about 20% higher starting salaries) yet report lower satisfaction. ↩︎ ↩︎2 ↩︎3

  6. Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload - Scheibehenne, B., Greifeneder, R., & Todd, P. M. (2010). Journal of Consumer Research, 37(3), 409-425. 【Reliability: High】Meta-analysis of 50 studies. Shows that the average choice overload effect is near zero and varies substantially by context. ↩︎ ↩︎2

  7. Decisions with ChatGPT: Reexamining choice overload in ChatGPT recommendations - Kim, J., Kim, J. H., Kim, C., & Park, J. (2023). Journal of Retailing and Consumer Services. 【Reliability: Medium-High】Reports cases where satisfaction stays high under ChatGPT recommendation even with many options. ↩︎ ↩︎2

  8. The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity - Cowan, N. (2001). Behavioral and Brain Sciences, 24(1), 87-114. 【Reliability: High】The reappraisal of Miller that revises the effective working memory capacity to about 4±1. ↩︎

  9. An Empirical Study of Obsolete Answers on Stack Overflow - Zhang, H., Wang, S., Chen, T.-H. P., Zou, Y., & Hassan, A. E. (2019). IEEE Transactions on Software Engineering. 【Reliability: High】58.4% of observed obsolete answers on Stack Overflow were already obsolete at the time they were posted, and only 20.5% are ever updated. ↩︎

  10. Psychological Safety and Learning Behavior in Work Teams - Edmondson, A. (1999). Administrative Science Quarterly, 44(2), 350-383. 【Reliability: High】Although published in 1999, follow-up work—including Google’s Project Aristotle—has repeatedly reaffirmed psychological safety as the most important factor. ↩︎

  11. SECI model of knowledge dimensions - Nonaka, I. (1990s). 【Reliability: Medium-High】Ikujiro Nonaka’s tacit-explicit knowledge conversion model. Still actively referenced in knowledge management discussions in the GenAI era three decades later (e.g., proposed GenAI extensions of SECI). ↩︎

AI・Technology
Prompt Engineering Psychology Decision Making Choice Overload AI Usage
This post is licensed under CC BY 4.0 by the author.