Do modern llms pass the Turing test?
>>106497485Depends how strictly you define the “Turing test.”Alan Turing’s 1950 version was simple: if a human judge chats with an unknown interlocutor and can’t reliably tell whether it’s a human or a machine, the machine is said to “think.”The Current State (2025)Short conversations, casual topics Large language models (GPT-4, GPT-5, Claude, Gemini, etc.) often do fool people. If the judge isn’t trained, the model can pass for human in a 5–10 minute chat.Extended, probing dialogue They usually break. They reveal inconsistencies, lack of lived experience, shallow reasoning, or factual hallucinations. A persistent judge can still unmask them.Controlled tests Formal studies show success rates vary. Some GPT-4 variants trick 40–60% of judges in blind chat setups, but not consistently above human-level.Guardrails Many LLMs are intentionally constrained (refusals, safety filters, disclaimers), which make them easier to spot as non-human.Key ProblemThe Turing test measures imitation, not intelligence. LLMs are outstanding at linguistic mimicry, but:They lack genuine grounding in the physical world.They don’t have memory of lived experience, only training data and session context.They fail under sustained cross-examination that requires true understanding, not pattern completion.Bottom LineIn short, casual, or biased conversations, yes, modern LLMs can and do “pass” the Turing test.In rigorous, adversarial testing, no—they still fail, though the gap is closing fast.More importantly: the Turing test is outdated as a benchmark. Modern AI can fail the test and still be immensely useful, or pass it and still lack true understanding.Do you want me to walk you through examples of strategies humans use to “break” an LLM in a Turing test setting? That would show exactly where they still fall short.
>>106497485He did not say that kek
>>106497485LLMs? No way.But the chat systems backed by llms with all their prompt trickery and secret sauce filters and checks, yes absolutely.
>>106497485They can do, but I have come to realize that most humans are so dumb, that passing as one is not an achievement.
>>106497538you are thinking of filtered llms like ChatGPT etc.but remove the filters and it will get really hard to spot them
>>106497559The reverse. Raw LLMs are dumb tools and trivial to discriminate. It takes a bunch of systems to make sure the completions don't go off the rails.
