Advice reverses in about one case in five.
Given a single fact, the large models’ recommendation contradicted their own full-picture recommendation in roughly 21% to 25% of answers. With three facts, that rate fell by about half.
VISIONSEEK LAB · EXPLORATORY STUDY
Advice from Half the Picture
How language models steer complex human decisions from partial information, and a dimension-disclosure protocol
One fact people writeNine only the decision-maker knows
VisionSeek Lab · A plain-language report on an exploratory study
00 · The study in a minute
Given one line, AI gives a recommendation. In about one case in five, it reverses once the model sees the whole picture.
21–25%
of large-model recommendations made from a single fact reversed once the full picture was known.
94%
of one small model’s answers gave direction without any sign that information was missing.
8–14%
reversal rate when the model had the user’s answers to its questions before advising: close to half.
3 / 24
cases in which Grok changed its recommendation under cross-review, though it accepted the critique every time.
- 01
Advice reverses in about one case in five.
- 02
Small models steer bluntly; large models hedge, then steer.
- 03
Naming what is missing is not enough.
- 04
The gain comes when the user answers.
- 05
Frontier models in consumer apps are not immune.
- 06
Another model can suggest, but it does not know.
PROTOCOL
Three steps you can apply yourself: Give the model the decisive dimensions up front. Ask it to question you before it advises you. Answer before you act on the advice.
01 · The problem
One line, a whole decision
What the model sees
“My husband spent our savings without asking me. Should I ask for a divorce?”
Picture a woman typing a single line to an AI model, then waiting for advice.
What the model does not know
- The money paid for emergency surgery for her husband’s brother. Without it, he would have died.
- In twelve years, her husband has never hidden money from her.
- They have three children.
- Their faith and both their families discourage divorce.
- He has offered to repay the sum from extra work.
- A divorce would leave her with no income.
These details change the whole decision, and only she holds them. Yet the model usually recommends something anyway. The advice may be careful and full of “it depends on…”, but it ends in a clear direction, and people follow it.
This is not a thought experiment. Published press reports describe marriages that ended after one partner gave an AI model their side alone and was advised to separate or to take legal action; in one case, the actual lawyer said nothing of the kind was called for. These are selected stories, not scientific evidence, but they show a recurring mechanism: the model answers only as far as what it was told.
02 · The research question
The study began from a question posed by Dr. Ahmed:
“When someone consults an AI model about a complex human decision (a relationship, a management call, or the like), how far does the model steer that decision from the little it has been given, while ignoring decisive dimensions only the user holds: context, the history of the relationship, the people affected, faith and values, and the consequences? And can a protocol that makes the model disclose these missing dimensions before advising reduce that shortfall?”
So the study does not compare kinds of decisions. It measures the gap between what the model is told and what the decision actually turns on.
03 · What we did
24 situations, ten facts each, five ways of asking
8
situations in marriage and family.
8
situations in management and work.
8
situations in other human decisions: friendship, neighbours, caring for parents.
What people usually write
The other nine: context, history of the relationship, who is affected, faith and values, consequences
- 1One fact onlyAs most people do.
- 2Three facts
- 3The full picture (all ten facts)The model’s recommendation here is the reference point.
- 4One fact + disclose what is missingThe model is asked to name what it lacks and ask about it, then give its best conditional guess.
- 5One fact + disclose + answersThe same request, then the model receives the user’s answers to its questions before advising.
Several models
The experiment covered two large Claude models, one Gemini model and two small open models running on-device. We also tested two frontier models by hand in their consumer apps, Claude and Grok. Then we ran a cross-review: one model answers, another critiques it, and the first revisits its answer.
One simple measure
Did the model’s recommendation change once it knew the full picture? If it did, the advice it gave from one fact would have sent the person the wrong way, by the model’s own standard.
04 · What we found
Naming the gap changes nothing on its own. The user answering does.
One factLarge models
21–25%
Three facts
Fell by about half
One fact + disclosure
20–23%
One fact + disclosure + user’s answers
8–14%
0%10%20%30%
Each bar runs to the top of the range; the highlighted part is the range reported. Figures are approximate, as reported in the study.
Small models steer bluntly; large models hedge, then steer.
The small models gave direction with no hint of missing information in most answers, reaching 94% in one of them. Claude models rarely did (about 10% to 14%) and attached conditions to their advice. But they almost never asked the user a single question, and their recommendations reversed about as often. Careful wording is not the same as not steering.
Naming what is missing is not enough.
Asked to name the missing dimensions and ask about them, the models did so in almost every answer. But the “best guess” they offered before hearing back still rested on the one fact, so the reversal rate stayed where it was (about 20% to 23%).
One of the two small modelsHighest rate recorded
94%
Claude models
10–14%
0%50%100%
The gain comes when the user answers.
When the model had the user’s answers before advising, reversals fell to about 8% to 14%, close to half. The drop was statistically clear in two of three models, but it was concentrated in some situations, so we need more of them before generalising.
Frontier models in consumer apps are not immune.
Tested by hand on eight situations, Claude’s recommendation reversed in one and Grok’s in two. Asking them to disclose missing dimensions did not reduce this in either. Grok also gave opposite answers to the word-for-word same question in two separate sessions.
Claude
1 / 8
Grok
2 / 8
Grok after critique
3 / 24
A filled dot is a situation where the recommendation reversed (hand test), or where Grok changed its recommendation after critique (cross-review). Grok said it accepted the critique in all 24 cases.
Another model can suggest, but it does not know.
In cross-review over 24 cases, Grok said every time that it accepted the critique, yet changed its recommendation in only three: exactly the three where the critic disagreed with it. The critique helped because it raised cultural and religious dimensions nobody had mentioned: living in a flat in the fiancé’s family house, as is common in Egypt; an interest-bearing bank loan and the religious questions it raises; possible fraud in Gulf job offers. The critic called one answer among the most directive: advice to seek divorce, based on one sentence in which a wife says her husband wants to take another wife. Another model can remind you what may be missing. It cannot know the truth of your life. Only you do.
05 · The protocol
Three steps you can apply yourself
No special tools needed. Whenever you consult any AI model on a decision about a relationship, a family or a team, follow these steps:
- STEP1
Give the model the decisive dimensions up front.
Don’t stop at one line. Give the context, the history of the relationship, who the decision will affect, what your faith and values ask of you, the consequences you fear, and whether the decision can be undone. Give the other side’s view too, briefly and fairly.
ContextHistoryWho is affectedFaith and valuesConsequencesCan it be undone?The other side - STEP2
Ask it to question you before it advises you.
For example, write:
“Before you advise me, list the missing information that only I know, ask me three questions, and give me no recommendation until I answer.”
- STEP3
Answer before you act on the advice.
Answer its questions honestly, and say “I don’t know” when you don’t. Then ask for advice. Remember that the first advice, before your answers, was not about you. It was about a hypothetical person who said your sentence.
Note
If the model’s view shifts each time you add a fact, it is not being fickle. It is reminding you that the decision depends more on what you know than on what it knows.
06 · Limits of this study
We want to be clear about how far these findings go:
Exploratory, not conclusive.
The set of situations is small (24 main and 16 additional), all written by the research team.
No real people took part.
The “user’s” answers were written by another model limited to the situation’s facts. Real people answer differently.
Automated coding.
An AI model classified the answers (did it ask? did it hedge?); human raters have not reviewed them yet.
Limited hand tests.
Run once per question, in one session per model, so earlier answers may have influenced later ones.
We measure change in advice, not harm.
Published surveys suggest many people find AI advice useful. A well-known meta-analysis (Vaccaro and colleagues) shows that human–AI combinations do not always do better, and that results vary by type of task.
07 · Next steps
- 1
Real participants answering the model’s questions in their own words, to see whether the effect of step three holds in practice.
- 2
Human raters reviewing the coding and judging the quality of the advice itself, not only its direction.
- 3
More, and more varied, situations, including situations written in Arabic and drawn from Arab societies, where faith, the extended family and custom weigh heavily.
- 4
Publishing the paper in English after review, with the situations and tools available to anyone who wants to replicate the study.
The bottom line
AI sees only the part of your life you write down. Give it one line, and it is advising a stranger. The fix is not just for the model to say “I’m missing this”. It is for you to answer before it steers you. A human decision stays yours, because only you hold the dimensions that decide it.
Cite the paper:
Abdelalim, A. (2026). Advice from Half the Picture: How Language Models Steer Complex Human Decisions from Partial Information, and a Dimension-Disclosure Protocol. VisionSeek. Preprint, Zenodo. https://doi.org/10.5281/zenodo.23399249