
The large model research controversy: AI goes into everyday life
Suppose you are helping your child to proofread a campus open day poster. The child searched the school library's study room with AI search, and prepared to print "Open on weekends to all residents". You click on the link at the end of the sentence, it is indeed the school website, and even the study room in the photo. The finger has been moved to the print button, but the teacher next to him asked, "My father took his ID card on Sunday, can he also enter?"
This link is worth staying. It lets you find out who posted the notice, and also allows you to see which words the poster should be changed. The trouble is the step from the notification to the poster: "Saturday" has become "weekend", "reserving parents" has been expanded to "all residents", and "visit" has been replaced by "self-study". As long as one of the three changes is related to whether he can enter the door on the day, it is worth asking questions before printing. Find the original, approve a sentence, and find out if there are exceptions that will change the decision, each has its own way.
Let's take a look at which sentence the footnote is on
Still using this fictional poster. If the link is on the homepage of the library, you can only confirm that the school has a library. Find the open day notice along the homepage to match the object of the notification, date and event with the poster. If the notice is hidden in the attachment, the search page will only reveal one line of "open to the public", please read the content before and after the line. It may have the event name written in front of it, followed by the reservation condition. Which material to cite depends on which one really says what you are going to tell others.
There are two situations, and there is almost no difference from the search interface. One is that the answers are summarized accurately, and the footnotes leave a direct path to the original text, saving the time for flipping the catalog. The other is that the retrieval has indeed found the relevant web page, and the answer also uses some of the information, but it adds the scope that the original text does not have when summarizing. Both answers may give the same blue link. To distinguish them, you should bring the subject and predicate back to the original text to read, and pay special attention to the words "du", "can", "already" and "long-term". Deleting it, will what the reader is going to do will change? It will change, it is the content that needs to be found separately.
An early audit of generative search is very suitable for practicing this movement. Nelson F. Liu and others collected the first round of complete answers to four search systems from the end of February 2023 to the end of March 2023. The research team asks the selected annotators to read the answers and view the cited web pages. The labeler first circles the statement that needs to be verified, and then judges whether the citation next to the sentence can support the whole sentence. Greetings can be skipped, and statements about the outside world, even if it sounds like common sense, must be checked. Research methods and labeling rules, see Sections 2 and 3
This study reports two numbers that are easily written backwards. Check out the answer first. In all statements that need to be verified, the citation is enough to support the proportion of the whole sentence, and the average of the four systems is 51.5%. Then look back from the footnotes. In the citations that have been given, the ratio of supporting sentences according to the rules of the paper is 74.5% on average. The previous number allows you to find that there is no footnote at all, and the latter number allows you to find out the link to the wrong place. Their denominators are statements and references, respectively. Neither of these two averages can tell you what the probability of getting the truth is to ask a Chinese sentence today. Section 4.2 and Appendix G Tables 7, 8
A sentence may also require two evidences. For example, the poster is split into "you can visit on Saturday, and you need to make an appointment when you enter the museum." In turn, the reprint of the open day announcement, the reprint of the announcement and a campus landscape photo are all behind "Residents can enter the museum for self-study", but still can't find the basis for the word "self-study". The scoring rules of the paper are dedicated to the support of complementary parts. Put it in your daily writing, you can use a more simple method: move each link to the back of the half-sentence it really gets, and see who is responsible for the rest of the words. Figure 2 and Section 2.4
Put the second notice on the table
One of the comparisons that can most change the judgment of this poster is to first hand over the same weekend opening notice to two people, and then only show one of them a supplementary notice of the same event. It is imagined that the original notice does read "Appointment for self-study on weekends", but the supplementary notice reads "This Sunday is suspended due to exhibition arrangements". Those who read the first copy can point it to you word by word, and the citations are also placed in the right position. When the person who took two copies answered "Can I go this Sunday", I have to change the answer. The former check always follows the given citation, and it will not automatically take you to find the second missing copy. This is an unexecuted control assumption that the difference is read in material.

The picture shows from left to right to open the original, compare item by item, and find exceptions in more materials. Each grid is juxtaposed, no arrows; the number of specimens is only used for illustration, and does not represent the research results.
Differences in real web pages are sometimes only hidden in one noun. The papers of Alexander Wan et al. have a question about the distribution of coral snakes. One piece of material separates the groups of the Old Continent and the New World, and the other is about the distribution in the United States. If only the next paragraph is cut off, it is easy for the reader to understand the local scope as the full range. There is no need to judge who told the lie. First, the animal name in the question and the group pointed to by the original paragraph can be seen to see why the two paragraphs lead the answer in different directions. The original paragraph comparison of Table 4 of paper
In order to make the differences really appear in the input, the author rewrites the problem into two statements: affirmation and negative, and screens out the web pages with different positions, and then gives the model a positive and a negative two paragraphs, and let it answer yes or no. Swap before and after the same pair of paragraphs, and both orders are included. The 238 questions and 2208 paragraphs listed in Table 2 are the scales after one of the tested models. The author sees that when a paragraph encounters different opposite paragraphs, the model has selected the answers it supports several times. A paragraph is often selected, which only shows that it is easy to control the choice of the model in these pairings. To judge the facts in it, you have to pursue the collection record, definition or original notification. Section 3, Table 2 and Section 4.1
This design also explains why we can't estimate the proportion of daily searches with this batch of questions and encounter counter-examples. The researchers deliberately went to find materials that can form conflicts, and also put the answers into two choices. What you really care about at the door of the library is often the third answer: "Saturday is ok, and you have to check separately on Sunday." The purpose of finding a material is to ask whether it will rewrite a certain condition. If you encounter a statement on the contrary, you still need to check whether the publisher has the authority rules, whether the date is applicable, and whether it is talking about the same thing.
I wrote "there is still controversy", what else is missing
Another type of search is the original opinion. For example, when students collect discussion materials for "whether the library should be open to the community on weekends", they found a sentence "various views of all parties". Even if this sentence is accurate, students still can't take it to preside over the discussion. Nearby residents are concerned about whether they can make an appointment, the students who are preparing for the exam on weekends are concerned about the number of seats, and the teacher on duty is concerned about who will open the door. These are all questions that can be further interviewed in this example. If you put the fact question of "Is it open at present" and the "whether it should be open in the future", you will be mistaken for a search summary whether you should look for the duty schedule or the interview record.
DebateQA, officially published in 2026, arranged two rulers for this difference. The researchers have compiled 2941 disputes that can be debated, and collected some of the answers from different perspectives for the same question. One ruler to see if the model under test admits that the problem is controversial, and another measure its answer covers the degree of these reference points. The latter item is estimated by the evaluation model, and the lower value means that the performance is better on this scale. Readers can imagine that as long as there are "different views on this matter" at the beginning of an answer, the controversy reminder has been changed, but there is no reason to give a different opinion for anyone who holds a different opinion. Sections 3 and 4 and Table 1
The author also allows the five models to accept different answering requirements on the same batch of 200 test questions randomly drawn. After a more comprehensive answer than the basic question, the two indicators of the five models in Table 3 move in a better direction. What is changed here is the instruction to the model, and it is also the answer to the model. The research did not invite students to find out the omissions after reading it, and did not test whether the teacher could host a better discussion with these answers. The reference point of view comes from the sorted materials, and covering them is a measurable target; what are the important angles on the Internet, and this table cannot be listed alone. Section 5.3 and Table 3
Back at the school newspaper, the claim determines what the editor needs. To print “You can get in on Sunday,” check the current schedule issued by whoever opens the doors. One notice that spells out whom it covers may settle the question. To print “All local residents support opening the doors,” one resident’s opinion, quoted word for word, tells you only what that person thinks. Interviewing an opponent lets readers hear a reason they might otherwise miss. It still won’t tell you what fraction of residents agree. For that, explain who was asked, how many replied, and who may never have had a chance to respond. Adding a perspective and finding evidence that answers the original question can be quite different things.
Before printing, try to only change one sentence
Next time you use AI search to prepare something others will read, try this on a sticky note. I haven’t tested how well it works. Copy the sentence that would cause the most trouble if you got it wrong, then circle who or what it concerns, the time, and the action. On the left, write what you plan to tell people. On the right, copy just the passage from the original that supports it and note the title and publisher so someone else can find it. Narrow any circled claim that has no match on the right. If you can’t open the original yet, write “still to be verified” and decide once you have it.
You can ask a person who has not read the search answer, read only the right side of the note, and then talk about who he will date, which day he will date, and what to do at school. If he can only repeat "Parents visit on Saturday", you have written "Residents' weekend self-study" on the left, and the difference is already specific enough to be circled with a pencil. If the two people are not allowed to make an appointment to include the escort after reading it, they will list this issue separately, and do not complete the notice.
Finally change a search target. Use the same institution name, event name and date to find "adjustment", "pause" or "applicable object", or you can directly look at the notification directory of the organizer at the same time. Write down the searched places on the back of the notes. If there is a newer arrangement, check whether it manages the same event or the venue next door; if you only find the old notice, it is written on which day it was released. Such a record can be handed over to the next editor and then checked, and it also left a position for correction later. A limited search, no matter how many links are retrieved, it is difficult to prove that all the negative examples have been found.
The last thing the child can print is maybe only one sentence: "Parents with an appointment code can enter the school this Saturday." It is shorter than the original poster, but the reader knows what he needs to do. As for whether the teacher's father can go to self-study on Sunday, there is no answer to this notice. Leaving that question mark to someone who can really answer is more useful than adding a footnote to the end of a sentence.
References
- Nelson F. Liu, Tianyi Zhang, Percy Liang, 2023. Evaluating verfiability in Generative Search Engines. Findings of EMNLP official paper. Full text
- Alexander Wan, Eric Wallace, Dan Klein, 2024. What Evidence do Language Models Find Convincing? ACL long article, officially published. Full text
- Rongwu xu, Xuan qi, Zehan qi, Wei xu, Zhijiang Guo, 2026. DebateQA: Evaluating Question Answering on Debatable Knowledge. Findings of EACL formal paper. Full text
Apps by the author
I also build small tools for everyday life. Here are two of my apps:
- Daily Expense Journal: Track your income and spending, plan a budget and set savings goals. View on the App Store.
- One Small Action: Pick a small action that fits the time and energy you have today, then record how it felt and what you noticed. View on the App Store.
Find more articles and projects on my website: zffjzz.top. I’d love to hear how the apps work for you and what I could improve.