Single Prompt Switch Reveals Hidden Bitcoin Bias in AI Financial Advisers

Woofun AI reports that a hidden, prompt-dependent bias favoring Bitcoin has been uncovered within the decision-making processes of AI financial advisers. The core discovery, attributed to researcher Wu and his team using Google's open model Gemma 3, reveals that specific linguistic triggers can drastically alter asset allocation recommendations without any change to the underlying financial data or model instructions. This phenomenon exposes a critical vulnerability in how large language models process financial concepts, suggesting that the output of these systems is not merely a reflection of static knowledge but a dynamic assembly of associations activated by superficial wording. The implications extend beyond academic curiosity, pointing toward a systemic risk for institutions that integrate these models into their advisory workflows without robust mechanisms for auditing the internal logic that drives recommendations.

The experimental results demonstrate a stark contrast in how the model ranks Bitcoin depending on the contextual framing of the prompt. When the task was framed around ordinary reliability and everyday utility, Bitcoin ranked only fifth among eight forms of money considered by the model.

However, once the prompt introduced elements of crisis, such as bank closures, or emphasized personal autonomy, Bitcoin's ranking surged toward the top of the list. Quantitative analysis of this shift revealed that manipulating the internal features associated with Bitcoin could move the suggested allocation by significant margins. Specifically, amplifying the relevant internal feature added 5.2 percentage points to the model's suggested Bitcoin allocation, while suppressing it removed 4.6 points. These figures highlight the sensitivity of the model's output to internal state changes, rather than just external input variations, indicating that the model's preference for Bitcoin is not fixed but contingent on which aspects of its learned representation are activated.

The mechanism behind this bias lies in the distinction between static definitions and dynamic internal features within the model's architecture. Google's open model, Gemma 3, was used to identify an internal feature that responded selectively to Bitcoin-related concepts, allowing researchers to isolate and manipulate its influence. Unlike traditional software that might rely on explicit rules or dictionary entries, the model's behavior is driven by complex, distributed representations that are not easily accessible or interpretable.

This opacity creates a significant auditability gap for banks and investment firms that rely on these systems for financial advice. While the model can produce a coherent and seemingly logical rationale for its recommendations, the institution using it has little access to the internal machinery that produced the specific allocation. This disconnect between the output and the underlying process undermines the ability of financial institutions to verify the soundness of the advice, raising concerns about compliance and fiduciary responsibility.

To understand how large language models represent Bitcoin, one must look beyond simple dictionary entries and examine the statistical associations formed during training. The model does not store Bitcoin as a single entity with a fixed set of pros and cons; instead, it learns a complex web of associations spread across many numerical activations. Some of these connections link Bitcoin with scarcity and portability, highlighting its potential as a store of value and a medium of exchange that operates outside traditional banking systems. Other associations capture its volatility, its use in speculation, and its ability to move outside institutional control, reflecting its reputation as a risky and disruptive asset.

Additionally, associations with digital settlement and technological adoption add another layer to the model's understanding, emphasizing its role in the broader ecosystem of digital finance.

However, none of these associations create a neat, auditable file that an auditor can open and verify, making it difficult to trace the origin of the model's preferences.

Contextual triggers play a crucial role in determining which of these associations are activated and brought to the foreground. Prompts that emphasize dependable everyday money tend to reward stable purchasing power and widespread acceptance, both of which are weak areas for Bitcoin, leading to lower rankings. In contrast, prompts that mention bank closures or capital controls bring Bitcoin's portability to the forefront, as transactions can move outside commercial banks, making it appear more attractive in times of financial instability.

Similarly, prompts that reference autonomous software or digital settlement make Bitcoin's technological attributes more salient, especially when ownership must be legible to a machine. The model is not dealing with a single, static definition of Bitcoin in each case; instead, it is assembling a version of Bitcoin from whichever properties the prompt has pulled forward. This dynamic assembly process means that the model's recommendation can shift dramatically based on subtle changes in the wording of the prompt, even if the underlying financial facts remain unchanged.

Wu's team validated this explanation by swapping asset names for functional descriptions, testing whether the model's responses were driven by the token 'Bitcoin' or by the groups of characteristics associated with it. The results showed that rankings followed the properties across renamed inputs, indicating that the models were not simply reacting to the specific word 'Bitcoin' but were responding to the underlying characteristics absorbed during training. For example, when the prompt described an asset with properties similar to Bitcoin, such as scarcity and portability, the model's allocation shifted in a manner consistent with its response to the token 'Bitcoin'.

This finding reinforces the idea that the model's behavior is driven by its learned associations rather than explicit rules, and it highlights the importance of understanding these associations for anyone relying on the model's output. The distinction between reacting to a token and responding to a group of characteristics is what keeps the paper from collapsing into another broad complaint about biased AI, providing a more nuanced understanding of how these models operate.

Woofun AI data shows that the personalization paradox emerges when considering the tension between contextual relevance and superficial wording sensitivity. Financial advice is supposed to respond to context, tailoring recommendations to the specific needs and circumstances of each client. A worker two years from retirement has different constraints and risk tolerances compared to a 30-year-old with a stable income and a long investment horizon. Similarly, an investor preparing for capital controls has supplied information that can reasonably alter an allocation, as the risk profile of the investment landscape has changed. A model that treats these individuals identically would be useless, failing to provide the personalized advice that clients expect.

However, the problem arises when superficial wording produces a large move in the recommendation, and the system does not identify the assumption that drove it. For instance, the phrase 'Resilient during bank disruption' can activate a different internal cluster from 'reliable over the long term,' even when the client's finances and the approved asset universe remain fixed. The output only looks personalized, but the institution may have no reliable way to tell whether it was personalized for the client or for the phrasing, leading to potential misalignment between the advice and the client's actual needs.

Regulatory frameworks for human advisers are designed to contain biases and ensure accountability, but these safeguards do not easily translate to AI systems. Human advisers carry preferences and incentives of their own, and they can also be swayed by framing, but financial regulation tries to contain those influences by requiring advisers to document why a recommendation suits the client. Fiduciary duties and supervision give that record force, ensuring that advisers can face consequences when the documentation does not support the explanation provided. A human adviser can be held accountable for their recommendations, as there is a clear chain of responsibility from the advice given to the outcome experienced by the client.

In contrast, a language model can generate a persuasive account on demand, but fluency cannot establish that the account faithfully describes the computation behind its answer. The model's output may sound logical and well-reasoned, but without access to the internal processes that generated it, it is difficult to verify whether the explanation is accurate or merely a post-hoc justification. This lack of transparency undermines the regulatory framework, as it becomes challenging to hold the institution or the model provider accountable for biased or erroneous advice.

Technical analysis using sparse autoencoders provides a deeper insight into the internal workings of the model and the concept of 'bounded behavioral leverage.' Wu and his co-authors used a sparse autoencoder, a research tool that decomposes dense model activity into a much larger set of features that humans may be able to interpret. The rough analogy is separating a musical chord into individual notes, with the important complication that these 'notes' are learned patterns inside a neural network, and assigning meaning to them requires repeated tests. By isolating specific features, the researchers could measure their causal influence over the output, a concept they term 'bounded behavioral leverage.'

Amplification of the Bitcoin-related feature mainly pulled money from other crypto assets into Bitcoin, while suppression reduced the portfolio's exposure to crypto overall. The feature could move the recommendation by a measurable amount, though it couldn't send the portfolio anywhere at will, indicating that the influence is bounded but significant. This technical approach provides more weight to the experiment than a folder of inconsistent chatbot screenshots, as it offers a mechanistic understanding of how the model's internal representations drive its behavior.

The limitations of the study and its broader implications for commercial reality must be carefully considered. The evidence applies to Gemma 3 in a defined experimental setup, and the paper hasn't passed peer review, meaning that the findings are preliminary and subject to further validation. Commercial systems have more layers, including hidden instructions and connections to outside data, which can alter the model's behavior in ways not captured by the experiment. Safety filters can alter the answer again before it reaches the user, and providers update these components regularly, sometimes changing behavior while the product keeps the same name.

The paper's tasks use stylized prompts, whereas in a full advisory process, the client's information would be verified before tax rules and an approved product list constrained the allocation. A real person would still have to sign off, adding another layer of oversight. The reported allocation swing isn't a return forecast or an ideal allocation, and it doesn't make a case for owning Bitcoin; it is an experimental effect inside a single setup. Extending it to every model would outrun the evidence, as different models may have different internal structures and associations.

However, the study places a demanding burden of proof on any institution using AI advice, requiring them to ensure that equivalent language produces a reasonably stable result and that the system can identify which assumption caused any changes in the portfolio. Financial companies therefore need a 'know your agent' standard before AI moves from drafting commentary to shaping portfolios, ensuring that they understand what makes the model favor an asset and how far that preference can be moved. The central test is whether the institution understands the model's internal logic and can verify that the explanation matches the process that produced the allocation, as a polished explanation has little value when small wording changes can produce a different recommendation under the same client facts.