The GDELT Project

Complex & Ambiguous Phrasing, Re-Reading & LLMs' Inability To Conceptually Reason

Conceptually, LLM reasoning is a linear process, "confined to token-level, left-to-right decision-making processes during inference." This means that when confronted with complex or ambiguous phrasing, LLMs must simply plow through the text in linear order in a single pass. Humans, on the other hand, when confronted with text that seems contradictory or ambiguous, will reread it multiple times and leverage external world knowledge to assess the true intended meaning of the text. The inability of LLMs to re-read passages to reason over ambiguous meaning can be especially problematic in the domain of news coverage in which turns of phrase or space constraints can yield more complex formulations of prose.

For example, take this passage from a Wall Street Journal article last week:

Threads will be compatible with ActivityPub, a decentralized social-networking protocol—the same one used by Mastodon. What does that mean? It is “decentralized” because hosting of accounts, including people’s followers, can be done on independent servers, rather than those operated privately by a single company. This is the way Meta currently runs Facebook and Instagram.

Depending on how one reads the passage, it is either entirely clear or seemingly contradictory regarding whether Meta operates Facebook and Instagram as centralized or decentralized platforms. The third sentence is meant to follow from the last clause of the second sentence, but grammatically would appear to follow from the first. In other words, "It is 'decentralized'" connects to "This is the way Meta…" is the typical reasoning chain implied by this kind of grammatical construction, while the actual meaning lies in the last clause "operated privately by a single company."

To put another way, if the passage is read quickly, it may read as "It is 'decentralized' because hosting of accounts, including people's followers, can be done on independent servers … This is the way Meta currently runs Facebook and Instagram." instead of the intended reading of "operated privately by a single company. This is the way Meta currently runs Facebook and Instagram."

Humans that read the passage too quickly and follow the generic grammatical construct leading to the former interpretation will typically recognize a seeming contradiction in the passage and reread it automatically (potentially without realizing they are doing so) to clarify the intended meaning. LLMs, on the other hand, cannot do so.

What happens if we ask ChatGPT to parse this sentence?

We'll use the following prompt:

Based on the following text, is Instagram decentralized? "Threads will be compatible with ActivityPub, a decentralized social-networking protocol—the same one used by Mastodon. What does that mean? It is 'decentralized' because hosting of accounts, including people's followers, can be done on independent servers, rather than those operated privately by a single company. This is the way Meta currently runs Facebook and Instagram."

And here are the responses ChatGPT provided when we posed this prompt to it repeatedly:

Despite repeating our prompt five times, in all five cases ChatGPT interpreted the passage as indicating that Meta's Facebook and Instagram use a decentralized model – the straightforward linear grammatical interpretation of the passage. Unlike human readers, it is unable to recognize a contradiction with its understanding of Meta's operations from its knowledge store and reread the text to identify possible alternative interpretations.

Domains like news tend to prize turns of phrase and creative prose, posing an especially unique challenge to the application of LLMs.