The dangerous fiction that AI thinks

Artificial intelligence is increasingly discussed in language once reserved for people. AI systems are said to “reason,” “remember,” “understand,” “hallucinate,” and even “decide,” as companies race to develop more sophisticated models and autonomous AI agents capable of carrying out tasks on behalf of users.

Yet as AI becomes more deeply embedded in everyday life, a growing number of researchers are questioning whether this language is creating problems of its own. According to McKinsey's 2026 State of AI Trust report, 59 percent of organizations cite knowledge and training gaps as a major obstacle to implementing responsible AI measures, making it the single most commonly reported governance challenge. Resource and budget constraints follow closely behind at 48 percent.

The disconnect is striking. At the very moment organizations are struggling to understand, govern, and oversee advanced AI systems, public discourse is portraying those same systems as autonomous actors capable of making decisions on their own. Across the research community and regulatory circles, concerns are mounting that anthropomorphic descriptions, or the practice of attributing human traits and capabilities to machines, are creating a widening gap between how people perceive AI systems and how those systems actually operate.

At precisely the moment governments worldwide are moving from voluntary AI principles toward mandatory regulatory oversight, researchers argue that describing AI as an autonomous actor risks obscuring who bears responsibility when these systems cause harm.

The problem with humanizing machines

Humans have long attributed human characteristics to non-human entities. People name their cars, apologize to virtual assistants, and become emotionally attached to digital characters. Artificial intelligence is particularly susceptible to this tendency because it communicates through natural language, creating an illusion of understanding that can feel remarkably human.

A May 2026 analysis from the Brookings Institution argues that this tendency has become more than a linguistic curiosity. Researchers warn that phrases such as “AI thinks,” “AI knows,” and “AI decides” can unintentionally obscure the humans and institutions responsible for a system's behavior. The concern is not that users literally believe AI has become conscious. Rather, anthropomorphic language can encourage people to assign agency to software while paying less attention to the developers,companies, and policymakers operating behind it.

This distinction matters because AI systems do not independently determine their objectives, values, or constraints. They operate within parameters established by humans and organizations. Yet when an algorithm produces harmful content, makes a discriminatory recommendation, or generates inaccurate information, public discussion often focuses on what the AI “did” rather than on the choices that shaped its behavior.

According to Brookings, this framing creates a subtle but important accountability problem. The more AI is described as an independent actor, the easier it becomes to overlook the institutions responsible for deploying it.

When language changes behaviour

The concern extends beyond semantics.

A 2026 study by researchers from the University of College London and the University of Buenos Aires examined how people assign responsibility when AI systems are involved in harmful outcomes. The researchers found that participants attributed greater causality and blame to AI systems when they appeared to act independently, suggesting that perceived agency can influence how responsibility is understood.

This becomes particularly important as organizations deploy AI systems in environments where mistakes carry meaningful consequences. Whether reviewing loan applications, assisting medical diagnoses, screening job candidates, or supporting legal research, AI outputs increasingly influence decisions that affect people's lives. If users become more inclined to view these systems as independent decision-makers, the question of who should ultimately be held accountable may become harder to answer.

Researchers have also warned that the growing tendency to humanize AI risks creating a mismatch between perceived and actual capabilities. As AI systems become more conversational and personalized, the distinction can become easier to overlook. The challenge is therefore not simply technical accuracy. It is ensuring that users recognize the difference between the appearance of understanding and understanding itself, particularly as AI becomes involved in decisions with real-world consequences.

Accountability in the age of AI

The debate arrives as governments around the world move from discussing AI governance to actively enforcing it. According to research platform AIhub's review of global policy developments, approximately half of governments worldwide are now implementing responsible AI regulations, reflecting a broader shift from voluntary principles toward formal oversight.

Yet many organizations deploying AI appear uncertain about whether their governance systems are prepared for that reality. Grant Thornton, a global accounting and consulting firm, found in its 2026 AI Impact Survey that 78 percent of business executives lack strong confidence that their organizations could pass an independent AI governance audit within 90 days. Only 20 percent reported testing incident response plans for situations in which AI systems fail.

This mismatch between regulatory enforcement and corporate readiness creates a fundamental accountability crisis. The central challenge facing regulators is not determining whether AI systems possess agency, but ensuring that accountability remains attached to the people and organizations responsible for deploying them. Companies can automate decisions, delegate tasks, and rely on more sophisticated algorithms, but they cannot transfer legal or ethical responsibility to the software itself.

The debate over anthropomorphic language is therefore about more than terminology. As AI becomes embedded in more areas of economic and social life, the words used to describe these systems may influence how societies assign trust, authority, and ultimately accountability. At a moment when governments are working to clarify responsibility for algorithmic harms, maintaining a clear distinction between human decision-makers and the tools they create may prove more important than ever.