A new essay by historian of science Pranav Sharma argues that today’s AI systems are being credited with understanding when they generate convincing language without regard for truth. His challenge is not just to the technology, but to what people mean when they call it intelligent.
Watch Desk analysis
What happened
In an opinion article published by ThePrint on 11 October, Sharma describes large language models as statistical systems trained to predict likely next tokens. He argues that reinforcement learning from human feedback rewards answers people prefer, which can make a system persuasive without making it reliable. The article points to a practical distinction: a fluent answer can still be wrong, and checking whether it is right often depends on a person or a rule-based test.
Sharma also takes aim at claims that AI systems understand or reason like people. He draws on the 1955 Dartmouth proposal, which described machines simulating features of intelligence, and on philosophical arguments about whether rule-following amounts to understanding. His discussion of Gödel’s incompleteness theorems leads him to argue that machines cannot surpass human mathematical reasoning. That is a contested philosophical argument, not a result established by the article’s account of current AI systems.
Why it matters
The words used for AI shape how people judge its outputs and where they place responsibility. If a system is treated as though it understands, readers may give its mistakes more weight, or expect it to handle tasks whose rules and checks have not been defined. Sharma’s proposed alternative is to describe these systems as statistical replication engines, and to use them for work where people can set the task and check the result.
That distinction matters in education, regulation and product design, where fluent output can be mistaken for demonstrated understanding. Sharma’s essay is a pointed argument, not a neutral account of the field: its claims about what machines can never do invite debate as much as its critique of inflated language does.
Our read
The essay is strongest when it asks what a model was actually trained to do, and who checks whether its answers are sound. “It sounds as if it understands” is a poor substitute for evidence, however impressive the demo. The sweeping claim that machines cannot surpass human reasoning is a much harder proposition, and the article’s philosophical case should not be mistaken for settled scientific consensus.
Still, Sharma is right to make the labels part of the discussion. Calling a system intelligent can quietly raise expectations before it has earned them. Read the full essay in ThePrint for the argument, then separate its useful questions about reliability from its larger claims about the limits of machine thought.
What to watch
- Whether AI developers and deployers describe systems’ capabilities in terms that match what they can demonstrate.
- How organisations distinguish fluent output from work that has been checked for accuracy.
- Whether the philosophical arguments about machine intelligence change as AI systems and evaluation methods develop.
Discussion spark: Should AI systems be called intelligent when they can produce useful, correct answers, or should the term be reserved for systems that understand what they are doing?
Sources and evidence
- Source update (11 October 2026, 06:33 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.