ChatGPT writes code, answers questions, drafts letters and produces a passable poem in the style of Jibanananda Das in ten seconds. In the eleven weeks since it launched, a great many people have concluded that it can do anything. It cannot, and the things it cannot do are more instructive than the things it can, because each one reveals what the machine actually is.
1. It knows nothing after 2021
ChatGPT is a language model: a program trained once, on a vast snapshot of text collected up to late 2021, and then frozen. Ask about the Qatar World Cup final, this year’s budget or yesterday’s news and it will either say it does not know or, worse, guess. It is not reading the internet. It is remembering a copy of it that stopped fifteen months ago.
2. It cannot look things up
For the same reason, it is not a search engine. It has no live connection, cannot open a link you give it, and cannot check a fact against a source. Microsoft’s new Bing, announced last week, bolts the same model onto a search index precisely to fix this — which tells you the limit is real.
3. It will not predict
Ask who wins Bangladesh v England next week and it declines. That is partly policy and partly honesty: a model that completes sentences has no basis for forecasting, and OpenAI has trained it to say so.
4. It stays out of politics
It will summarise positions but not take one, and it refuses to write attack pieces on named politicians. OpenAI added these guard-rails deliberately; users find them inconsistent, and the arguments about where the line should be will run for years.
5. It is wrong more often than it looks
This is the one that matters. The model produces the most plausible next word, not the true one, and it does so with the same fluent confidence whether the answer is right or invented. It fabricates citations, dates and quotations. Its Bangla is weaker than its English — the training data had far less of it — so Bangla answers are both less fluent and less reliable. Everything it tells you must be checked.
6. It falls over
“Network error”, “ChatGPT is at capacity”, answers that stop mid-sentence: the free service is running on more demand than OpenAI can serve, and the $20 ChatGPT Plus launched this month exists largely to buy priority access.
What this means in practice
For a student, it is a tutor that explains well and cites badly. For a freelancer, it is a first-draft machine for emails and code whose output must be read before it is sent. For a journalist, it is not a source. Two of the six limits — the knowledge cutoff and the lack of browsing — are engineering problems that Microsoft and OpenAI are already fixing. The fifth is not, and it is the one to remember.
Source: ZDNet, via Prothom Alo.
