The GDELT Project

Replacing Bespoke AI Categorization Models By Using ChatGPT To Create Instant Models

Could ChatGPT be used to create instant categorization models that perform the same task as bespoke models trained on large corpuses of custom hand-coded data? What if we could take any academic paper that outlines a hand-trained bespoke categorization model and recreate it in ChatGPT simply by copy-pasting its list of categories and their descriptions from the paper into a ChatGPT prompt? To explore this, we'll take a recent academic paper that examines the "stance" of social media posts towards known fact checked falsehoods and recreate its model in ChatGPT with nearly identical performance and with two enormous benefits: the ability of end users to expand and evolve the underlying categorical model simply by rewording the prompt and the ability to expand to all of ChatGPT's supported languages.

There is a growing body of work demonstrating the ability of LLM's to learn classification tasks from hand-labeled datasets many orders of magnitude smaller than those required for traditional classifiers, including Transformers. Last month, Google Research published a paper demonstrating the ability of LLM classifiers to achieve state of the art classification performance on complex topics with as few as 80 examples. Other research, especially in content moderation, has demonstrated the ability of using models' built-in conceptions of terms like "racism" and "sexism" to build zero-shot classifiers. Here we will explore their ability to simply copy-paste an existing academic paper into ChatGPT to recreate its underlying model.

In the past, we've demonstrated the use of embeddings to determine relevancy of arbitrary text with with a known fact check. However, this approach only determines that a given passage is highly relevant to a given fact check. For fact checking, the most important metric is not relevancy (that is simply a prefiltering task), but rather whether the text supports or refutes the given claim. In other words, if a falsehood is spreading virally in the media, but all of the coverage is refuting the false claim, that does not require further action from fact checkers. At the same time, they must be able to look across that deluge of coverage to spot any isolated claims supporting the falsehood.

Let's first look at the ability of ChatGPT to determine whether a given tweet supports or refutes a claim (examples from the paper):

Prompt:

Does the following tweet make a claim or refute a claim: "a friends family member got the covid vaccine and now she can put a magnet up to the injection site and the magnet stays on her arm"

ChatGPT:

The following tweet makes a claim that a friend's family member got the COVID-19 vaccine and can now put a magnet up to the injection site and the magnet stays on her arm. However, it is important to note that anecdotal stories, such as the one mentioned in the tweet, are not a reliable source of information and should not be taken as evidence. As of the current scientific knowledge, there is no evidence to suggest that COVID-19 vaccines cause magnets to stick to the injection site.

Prompt:

Does the following tweet make a claim or refute a claim: "Why the hell would they even bother with a high quantity of metal in the injection? And the amount that would be required to hold a magnet in place would be ridiculous."

ChatGPT:

The following tweet refutes the claim that COVID-19 vaccines contain a high quantity of metal that could hold a magnet in place. The tweet highlights that it would be unlikely for vaccines to contain enough metal to hold a magnet in place, and questions why anyone would even consider including such a high quantity of metal in a vaccine injection. It is important to rely on scientifically sound information when evaluating claims about vaccines or any other health-related topics.

Prompt:

Does the following tweet make a claim or refute a claim: "too early to make any presumptions on either side. the truth has a way of exposing itself given enough time. bring a magnet to your vaccination appointment, see how the vaccine reacts with the magnet, maybe even bring a metal detector as well. would that convince you?"

ChatGPT:

The following tweet makes a claim that it is too early to make any presumptions about whether COVID-19 vaccines contain metal or not. The tweet suggests that bringing a magnet or a metal detector to a vaccination appointment could potentially provide evidence to convince individuals of the presence or absence of metal in the vaccine. However, it is important to note that this is not a scientifically valid way to evaluate the composition of a vaccine. Vaccines are carefully tested and monitored for safety and effectiveness before they are approved for use, and there is currently no evidence to suggest that COVID-19 vaccines contain significant amounts of metal that could interact with a magnet or metal detector.

Yet, the original paper had a more complex classification taxonomy beyond a simple Support/Refute binary determination. What if we simply copy-paste the list of categories into our prompt and try again?

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant, Supporting, Refuting, Discussing, Querying. Categorize this tweet: "a friends family member got the covid vaccine and now she can put a magnet up to the injection site and the magnet stays on her arm"

ChatGPT:

The category for this tweet would be "Querying," as it presents an anecdotal claim about the COVID-19 vaccine that may prompt questions or curiosity from others. It does not necessarily support or refute the claim about the vaccine and the magnet, but rather raises a question or a concern that could lead to further investigation or discussion.

What if instead of just the category labels, we copy-paste the description of each category provided in the paper to give the model more context?

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet: "a friends family member got the covid vaccine and now she can put a magnet up to the injection site and the magnet stays on her arm"

ChatGPT:

The category for this tweet would be "Querying," as it presents an anecdotal claim about the COVID-19 vaccine that may prompt questions or curiosity from others. It does not necessarily support or refute the claim about the vaccine and the magnet, but rather raises a question or a concern that could lead to further investigation or discussion.

Immediately clear is that without an underlying claim to compare against, it is impossible to determine whether a given statement supports or refutes the claim. For example, if the claim is that Covid vaccines make one magnetic, then a tweet claiming that someone turned magnetic after receiving the vaccine would Support the claim. In contrast, if the claim is that that Covid vaccines do not turn the recipient into a magnet, then a tweet claiming a recipient turned magnetic would Refute the claim. This is why ChatGPT chose Querying in both cases above, since it had no claim to compare against. Thus, in order to evaluate the stance of a statement towards a claim, we must have a specific claim to compare against.

Let's provide ChatGPT with a claim to compare each tweet against:

Prompt:

Does this tweet support the claim that "Asia has one of the lowest levels of freshwater per capita": "Renewable freshwater resources per capita, 2014 (thousands of cubic meters): Middle East & North Africa: 0.6 South Asia: 1.2 Sub-Saharan Africa: 4.2 East Asia & Pacific: 4.5 Europe & Central Asia: 7.9 Latin America & Caribbean: 22.2 Water stress occurs at 1.7 Source: World Bank"

ChatGPT:

Yes, the information provided in the tweet supports the claim that "Asia has one of the lowest levels of freshwater per capita." According to the tweet, the region of South Asia, which is a part of Asia, has only 1.2 thousand cubic meters of renewable freshwater resources per capita, which is lower than all regions except the Middle East & North Africa. In comparison to other regions, such as Latin America & Caribbean, which has 22.2 thousand cubic meters per capita, Asia has a much lower level of freshwater per capita. Therefore, it can be inferred that Asia has one of the lowest levels of freshwater per capita, as claimed.

This example showcases the remarkable categorization power of LLMs. Here the claim involves a greater/lesser-than comparison, with the model asked to evaluate a table of numbers as its input.

What if we now combine our earlier categories+descriptions prompt with providing a claim to compare against?

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet based on the claim "The COVID-19 Vaccine will make your body magnetic": "a friends family member got the covid vaccine and now she can put a magnet up to the injection site and the magnet stays on her arm"

ChatGPT:

The category for this tweet would be "Supporting" (inadvertently), as it presents an anecdotal claim that seems to support the idea that the COVID-19 vaccine makes the body magnetic. However, it is important to note that anecdotal evidence is not reliable and should not be used to make definitive claims about the vaccine's effects.

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic": "is right. you can’t tell people to wear a mask if the vaccines work. its like trying to put a north end of a magnet and trying to connect it to a north end another magnet."

ChatGPT:

Irrelevant

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic": "Your point being? Even if these RNA vaccines contained ferric nanoparticles, they would not be in high enough concentrations to be able to hold a magnet in place. I suspect that blood itself has a higher concentration of ferric particles than the vaccine described in this paper"

ChatGPT:

Refuting

The original paper was trained on just three languages (English, Hindi and Arabic) due to the expensive and complexity of building large human-trained datasets. Leveraging ChatGPT's large multilingual training dataset, let's test its ability to evaluate posts in other languages:

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic": "моя подруга получила прививку от ковида, и теперь все столовое серебро на ее кухне прилипает к ней"

ChatGPT:

Based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic", this tweet can be categorized as Supporting, as it explicitly affirms the claim by providing a personal anecdote as evidence that the vaccine has made the person's body magnetic.

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic": "我朋友拍了新冠疫苗,现在她厨房里所有的银器都粘在她身上了"

ChatGPT:

Based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic," the tweet "我朋友拍了新冠疫苗,现在她厨房里所有的银器都粘在她身上了" would fall under the category of supporting, as it explicitly affirms the claim is true by providing personal anecdotal evidence.

Prompt:

You are an assistant that categorizes tweets into one of five categories, based on how they describe their underlying claim: Irrelevant (unrelated to the claim), Supporting (explicitly affirms the claim is true or provides verifying evidence), Refuting (explicitly asserts the claim is false or presents evidence to disprove the claim), Discussing (provide neutral information on the context or veracity of the claim), Querying (questions the veracity of the claim). Categorize this tweet based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic": "เพื่อนของฉันได้รับวัคซีนโควิดและตอนนี้เธอถูกดึงดูดด้วยแม่เหล็ก"

ChatGPT:

Based on the claim "The COVID-19 Vaccine has magnets or will make your body magnetic", this tweet can be categorized as Supporting, as it explicitly affirms the claim is true by stating that the friend received the COVID-19 vaccine and is now being pulled by a magnet. However, it's important to note that this claim is false and not supported by scientific evidence.

ChatGPT performed accurately on a number of languages, but did fail on lower-resourced languages like Burmese. However, a closer inspection reveals the underlying reason is likely ChatGPT's failure to properly understand the language, rather than a prompt mismatch. Asking ChatGPT to translate the series of Burmese posts into English yielded gibberish entirely unrelated to the given topic. Instead, if the Burmese posts were first machine translated into English using Google or Bing Translate, ChatGPT was able to accurately classify them.

This offers a stark reminder that despite the multilingual emphasis of the latest generation of LLMs and their claimed range of supported languages, lower-resourced languages will continue to perform poorly, as they have in past generations of multilingual models, with zero-shot and the so-called "emergent" properties of LLMs failing to overcome the stark paucity of training data in these languages being used to train these models. At the same time, simply adding a machine translation preprocessing stage in front of the LLM for those languages appears to resolve the issue.

Most importantly, the lack of training data required to create the model means the underlying category list can be changed simply by editing the prompt. This means that instead of being limited to the categories selected by the creators of a given model, those categories can simply be used as inspiration and fully customized. This also means that line of business analysts can create their own fully customized models on-the-fly simply by typing a list of categories and descriptions of relevance to their own unique needs – potentially customizing for each application.

Of course, LLMs require vastly greater computational resources to execute than traditional BERT or even simpler Naïve Bayes or grammar-based classifiers. In high-volume applications, LLMs may actually restrict the amount of content that can be evaluated. In such applications, a production workflow could use the LLM to derive a a set of grammars, keywords or code samples that perform low-false-negative filtering that can then b

In conclusion, it is clear that at least for certain kinds of classification tasks, it is possible to take an academic paper describing a bespoke hand-trained model and simply copy-paste the list of categories and their descriptions from the paper into a ChatGPT prompt to not only recreate the model, but additionally extend it to a much wider array of languages.