Chapter audiobook
एआई से गलतियाँ क्यों होती हैं
Listen to the complete teaching text, with spoken descriptions of diagrams. The lesson activities remain available below.
52:21 remaining at 1×
Press play to listen and follow the words here.
With this player focused: Space plays or pauses; arrow keys skip fifteen seconds.
Chapter sections and transcript
Download the transcriptHindi synthetic narration using Resemble AI Chatterbox-Multilingual-hi, MIT. Original Addaly teaching text adapted for listening.
Generation is not automatic verification
Here is the uncomfortable part of the last lesson, stated plainly.
The same generator can produce a true statement or a fluent falsehood. Training, the supplied context and the decoding procedure influence what it writes. A plausible continuation is not automatically a checked fact. Some systems add retrieval, tools or verification; their presence and effectiveness need checking too.
Unsupported or false generated claims are often called hallucinations. An invented citation need not be a software crash: normal generation can produce it. That does not make every error unavoidable or excuse a product that failed to check an important output.
What it looks like in practice
In the June 2023 Mata v. Avianca sanctions order, a New York court described six non-existent judicial opinions with fake quotations and citations produced using ChatGPT. It sanctioned two lawyers and their firm after they submitted and continued to defend them. The case establishes a failure to verify, not a measurement of the model’s internal cause.
The same thing happens with a bus route that sounds right for a city it has read about, a customer helpline number with the correct number of digits, a section of a country’s tax code with a convincing number, a research paper attributed to a real author who never wrote it.
Details that deserve a source check
There is a pattern, and it is useful.
Specific, poorly supported details deserve particular attention: names, numbers, dates, page references, quotations and local records. Their format can be learned without the exact fact being reliably available. Treat this as a useful risk category, not a universal ranking of error rates.
A familiar fact can be easier to answer, but repetition is not a truth guarantee. For example, pure water boils at about 100 degrees Celsius at normal atmospheric pressure. The conditions matter, and a model can still repeat a common misconception confidently.
A rare, arbitrary fact may have little evidence in the training data or the supplied material. A system can still compose a convincing-looking answer. It can also abstain, retrieve evidence or use a checking procedure; inspect what the actual product did rather than assuming it had to guess.
Both a well-supported answer and an unsupported one can sound confident. Tone alone does not distinguish them. Some uncertainty estimates are useful under tested conditions, but an emphatic sentence is not an independently calibrated confidence score.
What is not enough by itself
Asking are you sure. It can prompt a useful reconsideration, an unsupported reassurance or an unnecessary reversal. The reply needs checking. Research has found useful self-evaluation in particular models and formats, so “nothing can estimate reliability” is too strong.
Asking for a source. A model may name a real source, quote supplied material or invent a reference. Retrieval can help locate evidence, but a link is not proof of support. Open the actual record and compare the claim.
Assuming importance guarantees accuracy. You can explain the stakes and request checks, but urgency does not make an answer correct. Use an independent calculation, a current record or a qualified reviewer where an error would matter.
What does help
Supplying relevant documents or retrieval can improve answers on some tasks and gives you material to inspect. It does not guarantee that the right passage was found or read faithfully. Check both the source and what the answer claims it says.
Beyond that, a working rule: the more specific and checkable a claim is, the more you have to check it. Which is the opposite of the instinct. Specific claims feel more trustworthy. Here, specificity is the warning.
Methods and products continue to change. Judge a claim that hallucination has been solved by its defined task, test set and remaining errors, rather than by a confident sales statement or a blanket prediction that improvement is impossible.
Check your understanding
An answer states a familiar boiling-point rule and cites a particular 2019 study with a 34 per cent result on page 12, but gives no source record. Which claim most clearly needs its original record checked before you quote the finding?
Pick the one you would defend. Your answer stays private.
Your place and first choices stay on this browser. Classroom credit is recorded separately.
Stuck on something?Ask Meenal, your AI tutor