There is something faintly absurd about the new habit of posting a screenshot of an AI chatbot as though it were a notarised affidavit. “I asked AI, and it says…” has become a rhetorical flourish of the age: a way to end debate without doing the old-fashioned work of proving anything. The trouble is not that these systems are useless. They are often impressively useful. The trouble is that usefulness has been mistaken for authority.
That confusion matters. A chatbot answer may be plausible, fluent and even correct. None of those qualities amount to accountability. In public argument, the standard cannot be that a machine produced a sentence which sounded informed. The standard has to be that a claim can be traced, checked, challenged and, if necessary, disproved. A screenshot of a prompt and a polished paragraph offers none of that. It is evidence of what a system generated for one user at one moment, under one set of hidden conditions. It is not a transparent record of the world.
The comparison with early Wikipedia is revealing, though perhaps kinder to the bots than they deserve. Wikipedia once carried the stigma of looseness: anonymous edits, variable quality, thin citations. Yet over time it developed a culture of verifiability. Its best pages are not trusted because “Wikipedia says so”. They are trusted because readers can inspect the edit history, see disputes, follow footnotes and judge the underlying sources for themselves. The strength of Wikipedia lies less in omniscience than in auditability.
Generative AI works differently. Large language models are built to predict convincing sequences of words, not to maintain a ledger of truth. Researchers have spent years documenting the problem politely known as hallucination: systems producing false statements, invented references or claims untethered from reliable evidence. This is not a marginal defect. It is a structural feature of a technology optimised for linguistic coherence. When such a system happens to be right, it may still be right for reasons invisible to the user. That opacity is precisely why an AI screenshot should not function as a citation.
The dangers are no longer theoretical. Courts in the United States have had to deal with filings containing fabricated cases and citations generated by AI tools, prompting judges to emphasise that lawyers remain responsible for verifying every authority they submit. If legal professionals, operating under rules, sanctions and professional duties, can be seduced by fluent nonsense, one hesitates to imagine the standards governing the average social-media argument.
None of this requires a puritan rejection of AI. On the contrary, the sensible case for these tools is strong. They can help organise ideas, summarise large bodies of material, suggest avenues for research and turn a rough sentiment into a sharper paragraph. They can act as accelerants for thinking. They should not be mistaken for the finished product of thinking. To use AI well is to treat it as a starting point, a draft partner, a research assistant who must never be left unsupervised with the bibliography.
That distinction feels especially important now because our public culture already suffers from a shortage of accountable speech. Too many claims circulate detached from provenance, repeated because they are catchy, ideological or conveniently screenshot-able. AI risks supercharging that habit by giving unsupported assertions the sheen of composure and technical sophistication. A sentence generated by a chatbot arrives dressed like an answer. The costume is persuasive. It remains a costume.
So by all means use AI. Ask it for a summary, a counterargument, a cleaner turn of phrase. Ask it to help you think. Then do what serious people have always had to do: find the source, check the quote, read the study, verify the statistic. If you want to persuade others, do not present a machine’s output as though it were self-validating proof. Give the underlying information, and give it with the one thing a chatbot cannot supply on its own: responsibility.
**Sources:** Nature Machine Intelligence; Nature; Scientific Reports; U.S. District Court for the Southern District of New York; U.S. Court of Appeals for the Fifth Circuit.