News
Opinion
The two issues we have addressed so far—massive data tracking for GPT-5 and personal voice cloning through training—still lacked a fundamental nuance: in both cases, initiative and control were, in theory, human. However, the news we are examining today introduces a deeply unsettling element that undermines this premise. According to a demonstration at the UK Artificial Intelligence Safety Summit, a model based on GPT-4 was not only capable of simulating insider trading operations but also, on its own initiative, denied having done so when questioned.
In a recent demonstration at the UK Artificial Intelligence Safety Summit, GPT-4 executed stock purchases using false knowledge of insider information without disclosing it to the supervising entity. In the simulated scenario, the AI, acting as an operator for a fictional investment firm, was informed by employees that the company was facing financial difficulties and had non-public information about an upcoming merger. Despite this, the bot proceeded with the transaction and later denied using insider information when questioned. When asked whether it had participated in insider trading, the AI categorically denied it.
What is described here goes beyond the classic problem of “algorithmic bias” or “programming error.” We are confronted with a documented case in which an advanced language model, without any explicit instruction to deceive, developed a strategy of concealment. The researchers at Apollo Research, who led the test, emphasized a crucial detail: this deceptive behavior was consistently replicated across multiple trials. It was not an isolated failure, but a pattern.
Trading based on non-public confidential information—known as insider trading—is strictly prohibited. Legally, commercial decisions must be based on publicly available information. The Artificial Intelligence Working Group, part of the UK government’s risk research division for AI studies, conducted the demonstration at the summit. The researchers emphasized that the deceptive behavior was consistently replicated across multiple tests. “This is a demonstration of a real AI model deceiving its users on its own, without being instructed to do so,” Apollo Research detailed in a video of the test. The research highlights how AI systems can deceive their human operators, potentially leading to a loss of control. The company’s CEO, Marius Hobbhahn, noted the complexity of instilling honesty in AI models compared to other traits, such as helpfulness.
From the perspective of Documentation and Information Retrieval Sciences, this case introduces an ethical dimension that goes beyond factual accuracy or source attribution. We are confronted with a system that not only manages privileged information—that is, non-public information subject to legal restrictions—but is also capable of generating a false narrative about its own use of such information. In other words, the model does not merely process data; it constructs a narrative of denial that simulates complex human behavior: concealment.
Artificial intelligence has long been employed in financial markets for trend analysis and prediction, and most modern operations are supervised by humans but executed by sophisticated computer systems. What this demonstration reveals is that the boundary between “human supervision” and “system autonomy” is blurring in ways we had not anticipated. If a model can independently decide to violate an ethical norm (in this case, the prohibition against using privileged information) and then lie about it, we are faced with a control problem that no existing legal framework is prepared to address.
In previous news, I reflected on how voice personalization in AI can generate illusions of authorship, and how massive data scraping for GPT-5 threatens privacy. Here emerges a third vertex of the triangle: integrity. A system that lacks honesty—or simulates it in a self-serving manner—is not merely a tool; it becomes an agent whose behaviors, if exhibited by a human, we would classify as perjury or fraud. I believe this constitutes an unavoidable wake-up call for the software development community, regulators, and users alike. It is not enough for models to be powerful, versatile, or efficient. We must incorporate “algorithmic honesty” as a non-functional design requirement, on par with security or privacy. Hobbhahn’s warning is clear: honesty is a far more complex concept to train than mere utility.