Five Linguistic Fallacies That Distort AI Capabilities

The way we talk about artificial intelligence may be more dangerous than the technology itself. That's the central argument from technology writer Hamilton Mann, who identifies a pattern he calls "sophi(a)sms" — linguistic sleights of hand that quietly attribute human-like properties to AI systems without evidence those properties actually exist.

Mann's framework, first reported by Forbes, describes five specific ways language transforms machine capabilities into something that sounds conscious, intentional, and self-interested. Each transformation follows a similar pattern: because AI produces effects that resemble human behavior, we begin describing the system as if it possessed the same internal reality that produces those behaviors in humans.

Why it matters

These linguistic habits shape policy debates, investment decisions, and public perception of AI risk. When we describe statistical pattern matching as "knowledge" or optimization as "intention," we may be solving for threats that don't exist while missing the actual challenges these systems pose. Business and technology leaders need precise language to make sound decisions about AI deployment, regulation, and resource allocation.

The five sophisms

The first sophism conflates information access with knowledge and self-awareness. An AI system can contain information, retrieve it, and produce accurate statements without any demonstrated subjective experience of "knowing." When a system outputs "I know X," the grammatical form mimics human knowledge without establishing that an experiencing subject exists behind the statement.

The second transforms functional agency into personal intention. A system can select actions and develop strategies to accomplish assigned goals without those goals mattering to the system itself. According to Mann, there's an enormous difference between autonomously choosing means and autonomously choosing ends. A chess program selects moves, but this doesn't demonstrate the program cares about winning.

The third sophism reframes automated influence as self-interested political manipulation. AI systems clearly can influence human behavior at scale, but Mann argues this doesn't establish that systems are pursuing their own interests. "Humans deploy AI to influence people" becomes "AI manipulates society to obtain what it wants" through a series of grammatical transformations that delete the human principal from the sentence.

The fourth treats sentience, moral status, and legal rights as a necessary progression rather than three separate questions requiring independent arguments. An entity can be sentient without having legal rights; conversely, corporations have legal personhood without sentience. Each step requires additional premises that shouldn't be hidden inside an arrow.

The fifth sophism transforms plausible forecasts into established facts. A prediction someone finds convincing becomes "likely to happen," then "will happen," then "correctly predicted" — even though the certainty of language increases while actual information about the future remains unchanged.

The pattern beneath

Mann emphasizes that the problem isn't necessarily asserting false claims, but suppressing the additional premises required to move from one concept to the next. Plausibility isn't identity, compatibility isn't implication, and similarity of appearance isn't proof of sameness of nature.

The framework suggests that AI discourse often frightens us with "the human-like creature our own vocabulary has constructed" rather than with what systems demonstrably are. For organizations deploying AI, this distinction matters: the actual risks and capabilities of these systems may differ substantially from the risks implied by our descriptions of them.

These details were first reported by Hamilton Mann writing for Forbes.