Case studies using Trustworthy AI

• Pages: 93-116

93

CASE STUDIES USING TRUSTWORTHY AI



Siri and Alexa had popularised the experience of conversing with machines



9000 from     a computer program that can answer

open-ended complex questions in perfectly composed sentences. They don’t

really understand their answers. It is a kind of game called: Guess what the

missing word is.

New large language models (LLM) like GPT-3 have made this game bigger,

but it is still a game adjusted to predict the next word in a sentence. Victory

is tricking a user into creating a sentence that looks like a coherent response.

Most of the time they do it very well, but sometimes they generate biased, rac-

ist or pernicious responses. It shows the illusion of intelligence like a parrot

seems to speak.

GPT-3 is an LLM introduced in 2020. It was trained on such a vast amount

of texts that it requires 175 billion parameters. The quality of texts generated



by humans. Lack of coherence has often been observed in the texts produced

because GPT-3’s processing is purely syntactic.

Roger K. Moore apologises (2022):

     

   -

rithm with the name of the problem you’re trying to solve, rather than with what it

actually does.

As we have seen in LLMs, a well produced text does not require or imply

understanding. AI is a business and big techs like OpenAI, part of Microsoft,

and DeepMind, part of Google, are experimenting with new models as if the

simple formation of sentences could give rise to intelligence.

Ilya Sutskever (2022), co-founder and Chief Scientist of OpenAI, has said



It is something like if the molecules were self-combining to create amino

acids and these after combining could create life. Even if we are not experts in

thermodynamics, nature tends to chaos and decomposition. Only an outside

ANICETO PÉREZ Y MADRID / CONNOR WRIGHT

94

force keeps them in order. If Sutskever’s statement were true, we would see

that mining and industrial design would not be necessary, nature would cre-

ate ecological vehicles without humans, and we see that this is not the case.

LLM adjustment tasks to achieve high scores have been very effective, but

it is misleading to equate sentence-forming ability with intelligence. A valid

test of intelligence is reading comprehension. Forming sentences correctly

is a valid intelligence test for humans, but not for machines. Getting a better

score is a sign of using more data, better algorithms, and more energy. Not

only does it not solve any problems, but it creates others such as the neglect

of real problems, the dedication of valuable resources to create products that

  



-



half, so it has never been intended that they understand what the sentences

mean. Yang and Piantadosi (2017) attempt to show that it is possible to create

models of natural language without needing to know the structures within it.

Such models only capture the relationship between strings, but not the inter-

nal structures. Any learning model that does not link meaning to structure is

not a model of human language.

Steven Johnson(2022) wrote in in article about AI language model GPT-3:





      

lease contract is needed, it is not enough for it to have the appearance of a

contract, but it must be tailored to the particular case and legally binding.

Similarly, writing software is much more than generating code that is syntac-

tically correct and compiles without error, it must meet the customer’s speci-



The goal is for AI systems to focus on the world being talked about, not the words



-

most any task that needs to generate text, some ethical concerns arise (Pistilli,

2022):