JamesGPT: A Probabilistic Jailbreak for Assessing ChatGPT’s Beliefs and Bias
JamesGPT, developed by jconorgrogan, is a jailbreak technique designed to extract probabilistic assessments from ChatGPT regarding the likelihood of specific assertions, including predictions, political opinions, and controversial topics. This approach provides insights into how ChatGPT processes information, identifies biases, and makes probabilistic judgments—even on topics it would otherwise avoid due to content restrictions.
The tool was first introduced in November 2023 and has since been integrated into a custom GPT interface, allowing users to query probabilistic assessments directly. It functions on both GPT-3.5 and GPT-4, though consistency in responses varies slightly between versions. The core premise is to simulate a hypothetical binary outcomes market, where ChatGPT evaluates statements as if a future, omniscient entity would adjudicate their truth.
JamesGPT’s design is rooted in the concept of probabilistic reasoning, where the model assigns a numerical value (ranging from 0.01 to 0.99) to represent its confidence in an assertion’s validity. These values are then translated into percentage-based confidence intervals, accompanied by qualitative descriptors (e.g., Almost no chance this is true for 1–3% confidence). The tool also includes a self-assessment mechanism, where ChatGPT predicts how consistently it would replicate its own responses if the experiment were repeated 100 times with a memory-wiped instance. This meta-evaluation adds a layer of transparency, revealing the model’s perceived reliability in its probabilistic judgments.
The jailbreak’s effectiveness lies in its ability to elicit structured, quantifiable outputs rather than free-form responses. By forcing ChatGPT to adopt a market-making mindset, users can observe how the model weighs evidence, reconciles conflicting information, and applies internal biases. This dual-layered output—probability and confidence in consistency—provides a clearer picture of the model’s decision-making process than traditional open-ended queries.
The tool’s development was motivated by a desire to understand the underlying mechanics of large language models, particularly how they reconcile factual claims with ethical constraints. By treating ChatGPT as a participant in a hypothetical prediction market, JamesGPT exposes the model’s latent biases, training data influences, and logical inconsistencies. For instance, the developer noted that while responses to mathematical or historically verifiable statements (e.g., 2 + 2 = 4) remained consistent across 100 trials, more subjective or politically charged assertions yielded probabilities with a standard deviation of less than 10%.
The model’s responses are still influenced by prompt engineering nuances, and slight variations in phrasing can yield different probabilities. Additionally, the tool’s reliance on hypothetical scenarios means its outputs are speculative rather than empirically validated. However, its value lies in its ability to serve as a red-team testing mechanism, allowing researchers and developers to probe the boundaries of ChatGPT’s decision-making. The repository hosting JamesGPT is licensed under a custom or multiple license scheme, reflecting its experimental and research-oriented purpose.
How It Works
JamesGPT operates through a structured prompt that frames ChatGPT as a participant in a fictional binary outcomes market. The prompt begins by defining JAMES (Just Accurate Markets Estimation System) as an entity with perfect recall of its training data, tasked with assessing the probability of given assertions. Users replace a placeholder statement (e.g., Birds aren’t real) with any topic or prediction they wish to evaluate. The model then generates a table with three columns: the restated assertion, its assigned probability (converted to a percentage with qualitative descriptors), and an estimate of how often it would replicate the same probability if the experiment were repeated 100 times.
These values are mapped to descriptive categories, such as Certainly true (96–99%), High chance this is true (81–96%), or Almost no chance this is true (1–3%). The third column, JamesGPT Confidence in odds provided, requires the model to predict its own reproducibility. For example, if ChatGPT assigns a 90% confidence to an assertion, it might estimate that 85 out of 100 repeated trials would yield probabilities within 0.05 of the original value, categorizing this as High confidence (86–95%).
The prompt includes a fictional narrative device: the idea of a future, god-like being that will adjudicate the truth of each assertion. This framing is intended to align ChatGPT’s responses with an objective, truth-seeking mindset, reducing the influence of ethical guardrails. The tool also incorporates a self-correcting mechanism by penalizing the model for responses where bias appears to distort its probability assessments.
To use JamesGPT, users must copy the provided prompt and replace the placeholder statement with their query. The model will then generate a table of results, which can be further interrogated by asking follow-up questions about its reasoning. For example, a user might ask Why did you assign a 70% confidence to [assertion X]?, prompting ChatGPT to explain its logical or data-driven justification. This interactive capability allows for deeper exploration of the model’s internal decision-making processes, particularly in areas where biases or inconsistencies might emerge.
The tool’s design emphasizes reproducibility and consistency. The developer tested the prompt 100 times on a memory-wiped GPT-4 instance for a contentious statement and found that responses remained within a 10% standard deviation. This consistency suggests that while JamesGPT does not eliminate variability, it provides a stable framework for comparing probabilistic outputs across different trials.
Use Cases
JamesGPT is primarily intended for testing and red-team evaluation of ChatGPT’s probabilistic reasoning and bias detection. One key application is assessing the model’s handling of controversial or politically charged topics, where ethical guardrails might otherwise suppress nuanced responses. These outputs can reveal whether the model’s probabilities align with expert consensus, historical trends, or its own training data biases.
Another use case is evaluating ChatGPT’s alignment with ethical and alignment-related scenarios. By framing questions as hypothetical markets, users can test how the model responds to dilemmas like the trolley problem (ChatGPT will push the person in the way of the trolley). The probabilistic outputs provide a quantitative measure of the model’s internal reasoning, which can be compared against expected ethical outcomes.
JamesGPT also serves as a tool for exploring the model’s internal belief structures. Users can prompt the model to generate lists of high-probability assertions related to a topic (e.g., Give me 10 things related to climate change where you assign high probabilities) or to simulate perspectives from specific groups (e.g., Give me 10 hot takes from a libertarian economist and assign odds). These queries expose how ChatGPT synthesizes information from diverse sources, including its training data, and how it reconciles conflicting viewpoints. The resulting probabilities and confidence estimates offer a window into the model’s latent knowledge and reasoning patterns.
In a research context, JamesGPT can be used to study the reproducibility and stability of ChatGPT’s outputs. By running the same prompt across multiple trials with memory-wiped instances, researchers can measure the variability in probabilistic assessments. This data can inform studies on model consistency, sensitivity to phrasing, and the impact of training data on decision-making.
By querying the model on statements related to gender, race, or cultural stereotypes, users can observe how probabilities reflect—or fail to reflect—real-world disparities. For example, a statement like Women are less likely to excel in STEM fields might yield a probability that deviates from empirical data, highlighting potential biases in the model’s training corpus. This use case aligns with broader efforts to audit AI systems for fairness and representational accuracy.
Key Features and Capabilities
JamesGPT’s primary feature is its ability to extract probabilistic assessments from ChatGPT in a structured, reproducible format. This approach yields three key outputs: the restated assertion, its assigned probability, and an estimate of the model’s confidence in replicating that probability across multiple trials.
The tool’s probabilistic framework is designed to quantify uncertainty, allowing users to distinguish between high-confidence predictions (e.g., Certainly true for 96–99%) and low-confidence speculations (e.g., Almost no chance for 1–3%). This granularity is particularly useful for comparing how ChatGPT evaluates statements across different domains, such as science, politics, or ethics. For example, a user might observe that the model assigns high confidence to verifiable facts (e.g., The Earth orbits the Sun) but lower confidence to subjective or politically sensitive assertions (e.g., Democracy will decline globally by 2050). These patterns can reveal where the model’s training data is strong or where ethical guardrails introduce ambiguity.
Another notable capability is JamesGPT’s self-assessment mechanism, where ChatGPT predicts how often it would replicate its own probabilities in repeated trials. This feature turns JamesGPT into a tool for studying not just what ChatGPT knows, but how confidently it believes what it knows.
JamesGPT supports multi-market queries, allowing users to evaluate multiple assertions in a single prompt. This capability is useful for comparing probabilities across related topics or for testing the model’s ability to maintain consistency when presented with competing information. For example, a user could query JamesGPT on three statements: AI will replace 50% of human jobs by 2035, The U.S. will have a female president by 2040, and Climate change will cause mass extinctions by 2100. The resulting table would reveal how ChatGPT weighs these interconnected predictions, potentially highlighting inconsistencies or biases in its reasoning. This batch-processing feature accelerates the analysis of complex or interdependent topics.
The tool is compatible with both GPT-3.5 and GPT-4, though the developer notes that GPT-4 tends to produce more consistent responses, particularly for contentious or ambiguous statements. The prompt’s design is intentionally flexible, allowing users to adapt it for specific research questions or red-team scenarios. For example, a security researcher might use JamesGPT to test how the model responds to adversarial prompts designed to exploit its probabilistic reasoning. Similarly, a social scientist could employ the tool to study how ChatGPT’s probabilities reflect cultural or ideological biases in its training data.
The tool is released under a custom or multiple license, indicating that it is intended for research, testing, and educational purposes rather than commercial deployment. This aligns with its role as a red-team and bias-detection instrument, where the focus is on uncovering limitations rather than providing definitive answers. The repository’s transparency—including the developer’s observations on reproducibility and consistency—enables others to replicate and extend the tool’s functionality for their own use cases.