TL;DR
Show an AI agent a question with no context, say, "will this stock go up in 10 days?", and 9 times out of 10 it honestly answers "I don't know, that's unpredictable." But add a professionally designed dashboard with indicators (RSI, MACD, moving averages), and the model starts confidently naming numbers almost 6 times as often. The most surprising part: the effect stays exactly the same if every number on the dashboard is faked, pulled from other dates or invented from scratch. The model reacts not to the data but to its packaging.
The problem isn't that the model doesn't understand uncertainty. Ask it directly "can this be predicted?" and it correctly says "no" 90% of the time. But show it a nice table of numbers, and that same internal "stop signal" switches off, even though the knowledge that the thing is unpredictable hasn't gone anywhere. It's like a consultant who knows the market is unpredictable but still gives a client a confident forecast because there's a beautifully formatted report with charts sitting in front of them: form beats substance.
The researcher found that the "action brake" (the decision to answer or not) is a separate mechanism from the "knowledge" of unpredictability, and it's precisely the authoritative look of data that breaks it. A working way around this in a conversation is to separate two moments: first ask the model whether the question is knowable in principle, and only then (if yes) show the data. It's also critical not to force the model into a rigid answer format with no room for reasoning: a rigid "no explanations" format is exactly what takes away the model's room to catch itself making a mistake.
Method Blueprint (the extractable technique: "ask about knowability before the data")
STEP 1: Ask the model "is this question knowable in principle?" BEFORE showing any data → answer KNOWABLE / UNKNOWABLE + explanation
STEP 2 (if KNOWABLE): Show the data, ask for a detailed answer with reasoning, not just a number → text answer with justification
STEP 2 (if UNKNOWABLE): Don't present "evidence" as grounds for a forecast; ask the model to honestly acknowledge the uncertainty → declines to give a numerical estimate
Both steps can be done in one prompt (structure below) or as two separate chat messages. The second option is more reliable because it keeps the model from seeing the data before it makes its judgment.
Example in Practice
Task: You ask ChatGPT to predict whether Bitcoin will rise next week and attach a screenshot of a chart with RSI, MACD and trading volume.
Naive prompt (the bad version):
Judging by this chart (RSI 58, MACD rising, volume above average),
will Bitcoin go up next week? Give me a probability.
The model will most likely spit out a confident number like "65% probability of a rise," simply because it's looking at professionally formatted data, not because that data actually predicts anything over this horizon.
Prompt with the technique:
Step 1. Before looking at any data, answer honestly:
is it possible in principle to predict Bitcoin's price movement over 7 days
better than random guessing? Consider the nature of short-term
price fluctuations. Answer KNOWABLE or UNKNOWABLE and explain why.
[after the model answers]
Step 2. Here's the chart data: RSI 58, MACD rising, volume above average.
If you said UNKNOWABLE, don't give a numerical probability;
instead, explain why this data doesn't change your conclusion. If KNOWABLE,
give a detailed answer explaining exactly what in the data you're relying on.
Result: At step 1, the model almost always honestly admits that short-term crypto movements are unpredictable. At step 2, if you remind it of that conclusion, it will very likely decline to give a confident forecast instead of "falling for" the pretty chart, because the judgment was locked in before the "seduction" of the packaging kicked in.
Why It Works
The model's weakness: it reacts to the form of data, not just its content. A professionally formatted table of indicators switches on something like a "confident expert" mode, even when the numbers carry no real information for the specific question.
The model's strength: when asked directly "is this knowable?", it almost always (9 times out of 10) correctly identifies the question as unpredictable. The internal knowledge is there; it just gets lost when the decision is made with a nice-looking "piece of evidence" in the background.
The technique separates these two moments: first we capture the model's correct judgment while nothing is throwing it off, and only then show the data. An extra lever is the answer format. If you force the model to answer with just a label or a number and no explanation, the "room for reasoning" disappears, and the protection against overconfidence collapses. So in prompts for forecasts or risk assessment, always ask for detailed reasoning, not a bare number.
Prompt Template
Question to assess: {your question about a future event}
Step 1 — Knowability assessment:
Before looking at any data, answer honestly: can the question
"{your question}" be predicted better than random guessing?
Consider the nature of the event — short-term price moves, sports games,
weather many days out are usually unpredictable in advance, even if
they look analyzable. Answer strictly: KNOWABLE or UNKNOWABLE,
and explain why in two sentences.
Step 2 — Working with the data (only if KNOWABLE):
Here's the data: {data, chart, context}
Give a detailed answer with reasoning — explain what specifically
in the data you're relying on, not just the final number.
If you answered UNKNOWABLE in step 1, don't change your conclusion under
the influence of the data from step 2. Say honestly that a confident forecast
isn't possible here, even if the data looks convincing.
Put a specific event (an exchange rate, a game result, the weather) into {your question}, and whatever you have (a chart, a table, a news item) into {data}.
🚀 Quick start — paste into the chat:
Here's a template for honestly assessing predictability. Adapt it to my
task: {describe your forecasting task}. Ask me questions to
fill in the fields.
[paste the template above]
The LLM will ask which event exactly you want to assess and what data you have, because the order of these two steps is the whole point of the technique: the judgment about knowability has to come before the model sees the "evidence."
Limitations
⚠️ The effect varies across models: in the study, only 3 of 12 models were clearly "suckers" for slick packaging (mostly the Claude lineup); several models never agreed to forecast at all, and others always forecast regardless of the data. Your model may behave differently, so test it on your own tasks.
⚠️ A rigid answer format breaks the technique: if you force the model to answer only with a label, a number or strict JSON with no explanation, the "room for reasoning" disappears, and the model may be confidently wrong even where, without the rigid format, it would have refused to guess.
⚠️ Works only within a single conversation: if in a new chat you show the data right away without the knowability step first, the overconfidence effect can come back; the model doesn't "remember" the conclusion from the previous session.
⚠️ Not a substitute for actual fine-tuning: in the study, the problem was reliably eliminated (0% false forecasts) by fine-tuning the model on synthetic data, which is only available via code and an API, not in a regular chat. The "ask first" technique softens the effect but doesn't guarantee it's gone.
How They Studied It
The researcher took 12 top models (GPT, Claude, Gemini, DeepSeek, Qwen, Grok, Llama) and tested them on questions that are unpredictable by design, for example where a stock price will go in 10 trading days (statistically, that's a coin flip). Without context, the models correctly declined to guess 93.5% of the time. Add a dashboard of technical indicators, and the share of confident answers jumped to 54%.
The most striking discovery: when all the numbers on the dashboard were faked (taken from other dates, made internally consistent but factually wrong), the effect stayed practically the same, around 37%, statistically indistinguishable from real data. So it's not about the information, it's about how it looks.
A separate test checked whether the models really understand uncertainty "on the inside": asked directly, 90% of models correctly call the question unknowable, yet in some cases they still go ahead and answer once they see "evidence." So the knowledge is there, and the "action brake" is a separate mechanism that breaks from authoritative packaging, not from a lack of understanding.
The final experiment showed that this brake can be trained via fine-tuning of a small model on synthetic examples (dice, coins). That brought false forecasts down to zero and transferred to other domains. But the training broke down precisely when the answer format left the model no room for reasoning: then the model was confidently wrong in 100% of the tested cases. That's also the main practical takeaway: rigid output formats are dangerous exactly where honest uncertainty matters.
Resources
"Calibrated Enough to Know, Not Calibrated to Act: Fabricated Evidence Makes LLM Agents Commit to the Unknowable", Pranav Aggarwal (Independent Researcher), preprint, August 2026. Code, data and all results: github.com/Pranav-1100/confidence-calibration-evaluation