Citation 5.0: Exploring why AI struggles to grasp the essence of quotes — and learning from it
It’s always fascinating to see what Artificial Intelligence is capable of — for example defeating a chess master in a few seconds — but at the same time how quickly it struggles to grasp simple instructions. At IPPEN.MEDIA, we encountered this issue when dealing with quotes in our Generative AI editorial tools.
Quoting correctly is elementary in journalism. Of course, no politician or expert wants to have put words in their mouth or wants their words twisted. Misrepresenting quotes in journalistic articles or — worse still — making them up would significantly undermine the recipients’ trust in the media. However, Artificial Intelligence, especially large language models (LLMs), cannot consider the impact of such actions on media trustworthiness — with potentially serious consequences.
Challenges and solutions in implementing citations in news articles
This became particularly clear with tools that were supposed to edit, elaborate on, summarize, or improve articles. It didn’t matter how many times we told the prompt not to rewrite, shorten, or delete quotes from the original article — the LLM we were using (GPT-4) would do it anyway. Adjusting the prompt over and over again had a very unreliable effect on the results.
A potential solution that we had already developed as part of Nick Diakopoulos’ Generative AI in the Newsroom Challenge: A step-by-step approach based on extracting all quotes and then generating a new version of the text incorporating the previously extracted quotes. This approach proved reliable, but presented new challenges in a context where we were trying to build scalable solutions for over 200 editors.
Quote-approach 1.0: Extracting the quotes programmatically
That’s why we started extracting the quotes programmatically and replacing them with placeholders. The idea behind this approach is that we would not pass the quotes to GPT-4 via OpenAI’s API. This would prevent the LLM from rewriting them.
We instructed our tool to look for all the quotation marks in the article and to assign a specific label to each one it found, starting with “Quotation-1”. We then created a list and assigned each label to exactly one quote, much like a numbered index. This way, the quotes weren’t submitted to GPT-4 via the API. The quotes were then reinserted into the original position in the generated text after the API call.
This approach initially proved quite reliable, as the quotes were not altered or shortened. However, after some testing, we discovered a problem: because GPT-4 was only given a version of the article with the quotes removed, quotes were occasionally taken out of context, and in rare cases they were even fabricated.
Quote-approach 2.0 and 3.0: Showing AI the quotes and adjusting the prompt
With this in mind, we decided to continue with the automated extraction of quotes, but to provide the AI with the full article and all quotes in their original place in addition to a special reference list of all the quotes. By reminding GPT-4 in the prompt to leave the quotes unchanged, and adapting the prompt to this new method, we were quite happy with the results — until we applied this updated approach to articles that contained no quotes.
Quote-less articles: A challenge in the citation process
To prevent the LLM from making up its own quotes in articles that didn’t contain any, we provided a list of fake “quotes” that usually appear in most articles (“is” and “it”). However, these words were often inserted in the wrong places in the text, causing the sentences to lose their meaning.
For example: GPT-4 changed the sentence ‘He used this time to steal gold jewellery from the bedroom’ to ‘It was during this time that he stole gold jewellery from the bedroom’ — just to include the pronoun ‘it’.
Sometimes the sentence ‘It should be noted that the use of the quotations “is” and “it” in this text happened unchanged and in full length’ even appeared at the end of the edited article.
Quote-approach 4.0: Where did the quotation marks go?
So, a new fix was needed: Quote-approach 4.0. Although we had learned a lot about handling quotes with GPT-4, especially what not to do, Quote-approach 4.0 proved to be the worst thus far. Even though the approach worked flawlessly in articles without quotes, in stories with citations, it had a tendency to remove all quotation marks and completely rewrite the quotes.
This happened because, to prevent GPT-4 from including “is” and “it” in the wrong places, we forbade it from using quotation marks in the article except for the programmatically extracted quotes. However, it seems to have misinterpreted and often deleted all quotes.
For example: GPT-4 turned the sentence ‘“Brain-computer interfaces have the potential to change lives for the better”, according to Neuralink’s website’, into a sentence without the quote: “Neuralink, Elon Musk’s neurotechnology company, believes that brain-computer interfaces have the potential to change lives for the better.”
Quote-approach 5.0: The perfect combination between 3 and 4
This all sounds very frustrating — and believe me, it was — but fortunately, during the whole process, we found one approach that worked quite well for articles with quotes, and we found one that worked really well for articles without quotes. So, our fifth and final approach is now a combination of the two: Approach 3.0 is used for articles with quotes and approach 4.0 is used for articles without them.
We have tested this version in different use cases with many articles, and so far, we are pleased with the result. No quotes are rewritten, deleted, or removed from quotation marks — and no new ones are invented.
The last problem remains with the German subjunctive, a verb form used to express indirect speech, but that’s a story for another day.
