“Lying” in computer-generated texts: hallucinations and omissions

“Lying” in computer-generated texts: hallucinations and omissions

There is huge excitement about ChatGPT and other large generative language models that produce fluent and human-like texts in English and other human languages. But these models have one big drawback, which is that their texts can be factually incorrect (hallucination) and also leave out key information (omission).

In our chapter for The Oxford Handbook of Lying, we look at hallucinations, omissions, and other aspects of “lying” in computer-generated texts. We conclude that these problems are probably inevitable.

Omissions are inevitable because a computer system cannot cram all possibly-relevant information into a text that is short enough to be actually read. In the context of summarising medical information for doctors, for example, the computer system has access to a huge amount of patient data, but it does not know (and arguably cannot know) what will be most relevant to doctors.

Hallucinations are inevitable because of flaws in computer systems, regardless of the type of system. Systems which are explicitly programmed will suffer from software bugs (like all software systems). Systems which are trained on data, such as ChatGPT and other systems in the Deep Learning tradition, “hallucinate” even more. This happens for a variety of reasons. Perhaps most obviously, these systems suffer from flawed data (e.g., any system which learns from the Internet will be exposed to a lot of false information about vaccines, conspiracy theories, etc.). And even if a data-oriented system could be trained solely on bona fide texts that contain no falsehoods, its reliance on probabilistic methods will mean that word combinations that are very common on the Internet may also be produced in situations where they result in false information.

Suppose, for example, on the Internet, the word “coughing” is often followed by “… and sneezing.” Then a patient may be described falsely, by a data-oriented system, as “coughing and sneezing” in situations where they cough without sneezing. Problems of this kind are an important focus for researchers working on generative language models. Where this research will lead us is still uncertain; the best one can say is that we can try to reduce the impact of these issues, but we have no idea how to completely eliminate them.

The above focuses on unintentional-but-unavoidable problems. There are also cases where a computer system arguably should hallucinate or omit information. An obvious example is generating marketing material, where omitting negative information about a product is expected. A more subtle example, which we have seen in our own work, is when information is potentially harmful and it is in users’ best interests to hide or distort it. For example, if a computer system is summarising information about sick babies for friends and family members, it probably should not tell an elderly grandmother with a heart condition that the baby may die, since this could trigger a heart attack.

Now that the factual accuracy of computer-generated text draws so much attention from society as a whole, the research community is starting to realize more clearly than before that we only have a limited understanding of what it means to speak the truth. In particular, we do not know how to measure the extent of (un)truthfulness in a given text.

To see what we mean, suppose two different language models answer a user’s question in two different ways, by generating two different answer texts. To compare these systems’ performance, we would need a “score card” that allowed us to objectively score the two texts as regards their factual correctness, using a variety of rubrics. Such a score card would allow us to record how often each type of error occurs in a given text, and aggregate the result into an overall truthfulness score for that text. Of particular importance would be the weighing of errors: large errors (e.g., a temperature reading that is very far from the actual temperature) should weigh more heavily than small ones, key facts should weigh more heavily than side issues, and errors that are genuinely misleading should weigh more heavily than typos that readers can correct by themselves. Essentially, the score card would work like a fair school teacher who marks pupils’ papers.

We have developed protocols for human evaluators to find factual errors in generated texts, as have other researchers, but we cannot yet create a score card as described above because we cannot assess the impact of individual errors.

What is needed, we believe, is a new strand of linguistically informed research, to tease out all the different parameters of “lying” in a manner that can inform the above-mentioned score cards, and that may one day be implemented into a reliable fact-checking protocol or algorithm. Until that time, those of us who are trying to assess the truthfulness of ChatGPT will be groping in the dark.

_{Featured image by Google DeepMind Via Unsplash (public domain)}

OUPblog - Academic insights for the thinking world.

Real patterns and the structure of language

Real patterns and the structure of language

There’s been a lot of hype recently about the emergence of technologies like ChatGPT and the effects they will have on science and society. Linguists have been especially curious about what highly successful large language models (LLMs) mean for their business. Are these models unearthing the hidden structure of language itself or just marking associations for predictive purposes? 

In order to answer these sorts of questions we need to delve into the philosophy of what language is. For instance, if Language (with a big “L”) is an emergent human phenomenon arising from our communicative endeavours, i.e. a social entity, then AI is still some ways off approaching it in a meaningful way. If Chomsky, and those who follow his work, are correct that language is a modular mental system innately given to human infants and activated by miniscule amounts of external stimulus, then AI is again unlikely to be linguistic, since most of our most impressive LLMs are sucking up so many resources (both in terms of data and energy) that they are far from this childish learning target. On the third hand, if languages are just very large (possibly infinite) collections of sentences produced by applying discrete rules, then AI could be super-linguistic.

In my new book, I attempt to find a middle ground or intersection between these views. I start with an ontological picture (meaning a picture of what there is “out there”) advocated in the early nineties by the prominent philosopher and cognitive scientist, Daniel Dennett. He draws from information theory to distinguish between noise and patterns. In the noise, nothing is predictable, he says. But more often than not, we can and do find regularities in large data structures. These regularities provide us with the first steps towards pattern recognition. Another way to put this is that if you want to send a message and you need the entire series (string or bitmap) of information to do so, then it’s random. But if there’s some way to compress the information, it’s a pattern! What makes a pattern real, is whether or not it needs an observer for its existence. Dennett uses this view to make a case for “mild realism” about the mind and the position (which he calls the “intentional stance”) we use to identify minds in other humans, non-humans, and even artifacts. Basically, it’s like a theory we use to predict behaviour based on the success of our “minded” vocabulary comprising beliefs, desires, thoughts, etc. For Dennett, prediction matters theoretically!

If it’s not super clear yet, consider a barcode. At first blush, the black lines of varying length set to a background of white might seem random. But the lines (and spaces) can be set at regular intervals to reveal an underlying pattern that can be used to encode information (about the labelled entity/product). Barcodes are unique patterns, i.e. representations of the data from which more information can be drawn (by the way Nature produces these kinds of patterns too in fractal formation).  

I adapt this idea in two ways in light of recent advances in computational linguistics and AI. The first reinterprets grammars, specifically discrete grammars of theoretical linguistics, as compression algorithms. So, in essence, a language is like a real pattern. Our grammars are collections of rules that compress these patterns. In English, noticing that a sentence is made up of a noun phrase and verb phrase is such a compression. More complex rules capture more complex patterns. Secondly, discrete rules are just a subset of continuous processes. In other words, at one level information theory looks very statistical while generative grammar looks very categorical. But the latter is a special case of the former. I show in the book how some of the foundational theorems of information theory can be translated to discrete grammar representations. So there’s no need to banish the kinds of (stochastic) processes often used and manipulated in computational linguistics, as many theoretical linguists have been wont to do in the past. 

This just means that the methodological chasm between theoretical and computational linguistics, which has often served to close the lines of communication between the fields, can be surmounted. Ontologically speaking, languages are not collections of sentences, minimal mental structures, or social entities by themselves. They are informational states taken from complex interactions of all of the above and more (like the environment). On this view, linguistics quickly emerges as a complexity science in which the tools of linguistic grammars, LLMs, and sociolinguistic observations all find a homogeneous home. Recent work on complex systems, especially in biological systems theory, has breathed new life into this interdisciplinary field of inquiry. I argue that the study of language, including the inner workings of both the human mind and ChatGPT, belong within this growing framework. 

For decades, computational and theoretical linguists have been talking different languages. The shocking syntactic successes of modern LLMs and ChatGPT have forced them into the same room. Realising that languages are real patterns emerging from biological systems gets someone to break the awkward silence…

_{Featured image by Google DeepMind Via Unsplash (public domain)}

OUPblog - Academic insights for the thinking world.

What can Large Language Models offer to linguists?

What can Large Language Models offer to linguists?

It is fair to say that the field of linguistics is hardly ever in the news. That is not the case for language itself and all things to do with language—from word of the year announcements to countless discussions about grammar peeves, correct spelling, or writing style. This has changed somewhat recently with the proliferation of Large Language Models (LLMs), and in particular since the release of OpenAI’s ChatGPT, the best-known language model. But does the recent, impressive performance of LLMs have any repercussions for the way in which linguists carry out their work? And what is a Language Model anyway?

 At heart, all an LLM does is predict the next word given a string of words as a context —that is, it predicts the next, most likely word. This is of course not what a user experiences when dealing with language models such as ChatGPT. This is on account of the fact that ChatGPT is more properly described as a “dialogue management system”, an AI “assistant” or chatbot that translates a user’s questions (or “prompts”) into inputs that the underlying LLM can understand (the latest version of OpenAI’s LLM is a fine-tuned version of GPT-4).  

An LLM, after all, is nothing more than a mathematical model in terms of a neural network with input layers, output layers, and many deep layers in between, plus a set of trained “parameters.” As the computer scientist Murray Shanahan has put it in a recent paper, when one asks a chatbot such as ChatGPT who was the first person to walk on the moon, what the LLM is fed is something along the lines of:

Given the statistical distribution of words in the vast public corpus of (English) text, what word is most likely to follow the sequence “The first person to walk on the Moon was”?

That is, given an input such as the first person to walk on the Moon was, the LLM returns the most likely word to follow this string. How have LLMs learned to do this? As mentioned, LLMs calculate the probability of the next word given a string of words, and it does so by representing these words as vectors of values from which to calculate the probability of each word, and where sentences can also be represented as vectors of values. Since 2017, most LLMs have been using “transformers,” which allow the models to carry out matrix calculations over these vectors, and the more transformers are employed, the more accurate the predictions are—GPT-3 has some 96 layers of such transformers.

The illusion that one is having a conversation with a rational agent, for it is an illusion, after all, is the result of embedding an LLM in a larger computer system that includes background “prefixes” to coax the system into producing behaviour that feels like a conversation (the prefixes include templates of what a conversation looks like). But what the LLM itself does is generate sequences of words that are statistically likely to follow from a specific prompt.

It is through the use of prompt prefixes that LLMs can be coaxed into “performing” various tasks beyond dialoguing, such as reasoning or, according to some linguists and cognitive scientists, learn the hierarchical structures of a language (this literature is ever increasing). But the model itself remains a sequence predictor, as it does not manipulate the typical structured representations of a language directly, and it has no understanding of what a word or a sentence means—and meaning is a crucial property of language.

An LLM seems to produce sentences and text like a human does—it seems to have mastered the rules of the grammar of English—but at the same time it produces sentences based on probabilities rather on the meanings and thoughts to express, which is how a human person produces language. So, what is language so that an LLM could learn it?

A typical characterisation of language is as a system of communication (or, for some linguists, as a system for having thoughts), and such a system would include a vocabulary (the words of a language) and a grammar. By a “grammar,” most linguists have in mind various components, at the very least syntax, semantics, and phonetics/phonology. In fact, a classic way to describe a language in linguistics is as a system that connects sound (or in terms of other ways to produce language, such as hand gestures or signs) and meaning, the connection between sound and meaning mediated by syntax. As such, every sentence of a language is the result of all these components—phonology, semantics, and syntax—aligning with each other appropriately, and I do not know of any linguistic theory for which this is not true, regardless of differences in focus or else.

What this means for the question of what LLMs can offer linguistics, and linguists, revolves around the issue of what exactly LLMs have learned to begin with. They haven’t, as a matter of fact, learned a natural language at all, for they know nothing about phonology or meaning; what they have learned is the statistical distribution of the words of the large texts they have been fed during training, and this is a rather different matter.

As has been the case in the past with other approaches in computational linguistics and natural language processing, LLMs will certainly flourish within these subdisciplines of linguistics, but the daily work of a regular linguist is not going to change much any time soon. Some linguists do study the properties of texts, but this is not the most common undertaking in linguistics. Having said that, how about the opposite question: does a run-of-the-mill linguist have much to offer to LLMs and chatbots at all?   

_{Featured image: Google Deepmind via Unsplash (public domain)}

OUPblog - Academic insights for the thinking world.