What is a Mind?

By: Michael Boehmcke

Screen Readers, Large Language Models, and Artificial General Intelligence

There is a great deal of discussion today regarding the proliferation of Generative “Artificial Intelligence” (GenAI) and Large Language Models (LLMs) throughout society, especially within the realm of employment and the way that this technology interfaces with the public at large. There are a great many unanswered questions, ranging from the sensible and pressing, “What will happen to the countless workers who are employed in jobs LLMs are automating out of existence” and, “Are these jobs that are being automated actually being performed well by LLMs?” to questions that can more safely be put out of mind until the ethical concerns for humanity around AI are sorted. Nonetheless, these are most often the questions that are found to be the most interesting to people who discuss AI, as their nebulous qualities make these questions much easier to talk through rather than the definitive action that would be necessary to protect the worker from the tyranny of automation.

Foremost in the mind of many who discuss Large Language Models is the implications that artificial intelligence has for the experience of human consciousness, or, put another way, how AI calls into question the assumptions that we have about minds, ours and others. In his book, “Other Minds” Dennett says

“It is beyond serious dispute , however, that you and I each have a mind. How do I know you have a mind? Because anybody who can understand my words is automatically addressed by my pronoun “you,” and only things with minds can understand. There are computer-driven devices that can read books for the blind: they convert a page of visible text into a stream of audible words , but they don’t understand the words they read and hence are not addressed by any “you” they encounter; it passes right through them and addresses whoever listens to-and understands-the stream of spoken words . That’s how I know that you, gentle reader/listener, have a mind. So do I. Take my word for it.”

I think that Dennett touches on an interesting point for discussion, which is the question of the difference between screen readers and the LLMs of genAI. Functionally, the two appear to be incredibly different. Screen readers are, ironically, blind to the text that they read and simply pass the information from the visual to the audible without any additional processing. Generative AI is capable of the same feat, though many operate in a text-to-text format only with no audio interpretation, though there are those which can provide essentially the same end product as a screen reader. However, despite the seeming parity of genAI and screen readers, and perhaps even a promising gain of function which could come from generative AI based screen-reader technology through providing a linguistic description for images that were not given an alternative text description that a traditional screen-reader relies on for such purposes, I have not seen much push for the adoption of such technology into screen-readers.

Fundamentally, the difference between genAI and a screen-reader, the difference which may explain why they have not been a largely co-opted as a driving force for screen-reader development, is the degree of processing which is done to interpret the text being read out. For a screen-reader the goal is to have the minimum amount of processing necessary to re-emit the text as audio for the user, but that can never be the case for a generative AI model. It will always fully process every token that it takes in as data, vectorize it, translate it, rewrite it, and only then be able to send that data to the audio emission engine. This process may be a perfect copy ninety-nine percent of the time, but that one percent of error which remains renders every transcription from the AI into question, and only for a marginal increase in efficacy. Perhaps the AI will find its place in screen-reader technology as a system which can step in to interpret web objects that a traditional screenreader cannot, but the inherent unreliability will always call the systems into question.

As you will no doubt come to see if you read the entirety of this blog, I am of the firm opinion that generative AI based on the current LLM approach to vectorized data will never be able to be consider to have a mind of its own. I believe that any output from these systems which resembles thought, reason, or interpretation is a facsimile, a simulacrum, a pale imitation of the actual complexity that gives rise to the human cognitive capacity. Despite these beliefs, as well founded as I think they are, I have to admit that the degree of processing that an LLM performs on all data that it intakes is a basic component in any system which could be considered to have a mind. A human may be capable of a one hundred percent accurate transcription of an input corpus, but that person will always interpret the text being input as language, interpret it and think about what it is saying in the process. They may ignore that process but you, now, as you read this, even if you were to ignore the actual meanings your mind associates with each word, will still interpret the words as you transcribe them. In this, man and machine are one in the same, in regards to LLMS.

All of that said, however, there is one key difference between the way that an LLM and a human will interpret a text as they transcribe it. A person may make errors of judgement, scribal accuracy, typos, or any number of strange mistakes, but it would require a truly troubled mind to take a task of transcription and, in the process of interpreting a text and re-writing it word for word, with that explicit goal in mind with no ulterior motive, to make up a completely new sentence with similar words but still meaning something utterly different. In that regard, humanity bears no resemblance to an LLM. No matter how good these predictive models become, they will ultimately still be beholden to a degree of obfuscated processing which will always bear the possibility of a whole-cloth hallucination of something new. And no mind which can act without regard for continuity can truly be considered a mind like ours.

clankers beware