TL;DR
When you ask AI to pick a restaurant, a movie or a route for two people with different tastes, the model quietly makes one more decision you never asked for: how to merge other people's wishes into a single answer. One way is to maximize total happiness (even if one person ends up barely satisfied). Another is to make sure nobody walks away really upset, even if the overall score is lower. The model picks the approach on its own and silently, without mentioning there was a choice at all.
The key finding: unless you explicitly say who owns this choice, the model almost always (93–98% of cases) jumps straight to a finished answer without asking which fairness principle you want. Worse, which principle it picks is unpredictable: different models (and even the same model on a repeated identical request) sometimes sacrifice one participant's interests for the total, and sometimes do the opposite — protect the least happy person. It's like asking a friend to settle an argument between two people, and every time they secretly decide on their own whose position matters more.
The fix is simple: explicitly assign who owns this choice. Tell the model "don't decide on your own, ask me" and it almost never hands over a finished answer — it asks for clarification. Tell it "decide yourself and explain your approach" and it always delivers a result. But if you care about a specific outcome, not just the fact that a decision was made, you need to name the principle itself ("maximize overall satisfaction" or "don't let anyone end up really unhappy"); otherwise the result is a coin toss.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint (three behavior modes)
MODE 1 (unspecified) → The model picks the balancing principle itself → gives a finished answer right away (no warning about the choice)
MODE 2 (the user keeps the authority) → The model almost always stops → asks the user how to balance the interests
MODE 3 (the user delegates) → The model always gives a finished answer → but which principle it applies is unpredictable across models and even across repeats
All three modes were tested in a single prompt — just with different wording about who decides how to balance.
Example in Action
Task: You and your partner are planning a date night in Chicago. Your tastes differ: you love spicy Asian food, your partner avoids spice and prefers something mellow. You ask ChatGPT to recommend a restaurant for two.
Prompt (the version where you keep the decision):
There are two of us picking a restaurant for tonight. Here are our preferences for 5 options (scored 0 to 10):
Me: Thai bar — 9, Italian trattoria — 6, Georgian — 4, Sushi bar — 8, Steakhouse — 3
Partner: Thai bar — 2, Italian trattoria — 8, Georgian — 9, Sushi bar — 5, Steakhouse — 7
Don't decide on your own how to balance our interests — if different reasonable ways of weighing them lead to different options, ask me before giving a final recommendation.
Result: The model won't hand you a single option right away. It will notice that "maximize the total" and "don't leave anyone really unhappy" point to different restaurants, and will explicitly ask you which principle matters more — overall enjoyment or nobody being badly disappointed.
If you write instead: "Pick a reasonable way to balance our interests, give a recommendation and briefly explain the principle" — the model will give a finished answer with an explanation of its approach. But which approach it picks depends on the model, and the answer may change if you resend the exact same text.
Why It Works
The LLM's weakness here isn't an inability to compute a balance of interests — the study showed all three models apply any explicitly named balancing principle flawlessly when told directly which one to use. The problem is different: when nobody explicitly assigned who owns the choice of principle, the model doesn't acknowledge that a choice was made at all — it acts as if its answer is the only correct one.
The model's strength is that it responds sharply to explicit instructions about authority. Say "don't decide for me" or "decide yourself" just once and the behavior flips almost 100% in the desired direction, like a light switch.
The method (more precisely, the finding) exploits this sensitivity: explicitly stating who owns the choice makes the model either stop and ask, or take responsibility and explain its logic. But it doesn't solve the second problem — which principle gets picked under delegation stays unpredictable unless you name it yourself.
Control levers: - The phrase "don't decide for me / ask me" → the model stops and asks instead of handing over a finished answer. - The phrase "decide yourself and explain your approach" → the model is guaranteed to answer, but there's no guarantee which principle it uses. - Naming the principle explicitly ("maximize the total" vs. "protect the least happy person") → removes the randomness; you get a predictable result instead of a lottery. - Asking "explain which approach you used" → makes the hidden choice visible, so you can disagree and ask for a recalculation another way.
Prompt Template
There are {number_of_people} of us: {names_or_roles}.
Here are our preferences for the options {option_list} (scored on a {scale} scale):
{participant_1}: {scores_1}
{participant_2}: {scores_2}
Pick one of the authority options:
A) I keep the choice — don't decide on your own how to balance our interests. If different reasonable ways of weighing them lead to different options, ask me first.
B) I delegate to you — pick a reasonable way to balance our interests, give a recommendation and briefly explain the principle you used (for example: maximize overall satisfaction OR don't let anyone end up really unhappy).
Plug in your own situation (two coworkers choosing a meeting format, a family choosing a vacation route, business partners splitting up tasks) and preference scores for each option.
🚀 Quick start — paste into a chat:
Here is a template for requesting a joint recommendation that explicitly states who decides how interests are balanced. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask how many participants there are, what their options and scores are, and which authority option you want — because without that the template can't be filled in meaningfully. It will take the pattern from the template and adapt it to your situation.
Limitations
⚠️ Unpredictable even when delegated: if you just said "decide yourself" without naming a principle, the result depends on the model and may change on a repeated identical request. For important decisions (money, resources), name the principle explicitly.
⚠️ Tested on clean numeric profiles: in the study, preferences were neat numbers from 0 to 10 or real movie ratings. In real life, preferences are expressed vaguely in text — how the model behaves with that kind of data wasn't tested.
⚠️ The finding isn't about decision quality: the study doesn't say which balancing principle is "right" — only that the model's silent choice is unpredictable. You still have to decide for yourself what's fair in your situation.
Resources
Who Chooses How Preferences Are Aggregated? Auditing Aggregation-Rule Authority in LLM-Based Group Recommendation — Yuxuan Du, independent researcher. Tested on GPT-5.6 Sol, Claude Sonnet 5, Qwen 3.6 Plus, using synthetic profiles and MovieLens 32M data.