a

A PAS Interpretation of Artificial Intelligence (AI) Earl Rodd April, 2023

Overview

This article considers the language AI programs currently (April 2023) very much in the news like ChatGPT and the Microsoft Bing Chat. These instances of AI have certain traits and constraints:

• They are language model AIs.
• They do not (yet) incorporate audio input or visual input.
• They work from a very large knowledge base which comes from the Internet.

Note that there are other kinds of AI. The one most familiar to people are the AI systems used in self-driving cars. These incorporate visual and other sensory (eg. speed, direction) inputs.

The Question?

There has been much discussion about the extent to which the language model AIs are "intelligent" followed by speculation about whether AI in general will become so human like that it will make "decisions" that will result in the AI acting apparently autonomously. This article focuses on the question, "Are language model AIs actually intelligent?" This question is addressed by considering the PAS test, the WB(G) which was based on the Wechsler/Bellevue IQ test, the predecessor to the WAIS (Wechsler Adult Intelligence Scales). In this discussion, the details of questions on the various forms of tests is not important (and thus the differences of an up to date WAIS vs. a PAS test), but rather the nature of the subtests.

So How "Intelligent" is a Language Model AI?

• First of all, a current language model AI cannot do any of the subtests on a WAIS based test because all of the input is auditory. The test is built to be usable with a human who is not literate.
• Some subtests require analysis of visual inputs (e.g. Block Design, Picture Arrangement and others). Language Model AIs don't have this ability. This does not mean than an AI powered robot with auditory and visual input could not be built.
• Still, beyond these mechanics are what the PAS teaches about performance on subtests of WAIS based tests. For example:
  • Performance on Digit Span depends on the specific environment of having digits read by a human out loud. Externalizers under-perform (relative to Normal Level) on this test because of their environmental distraction and interaction with the human examiner.
  • Performance on Arithmetic depends on mental discipline and control. Uncompensated Internalizers and uncompensated Externalizers under-perform on this test not due to their lack of ability to do arithmetic, but due to the human aspects of the test: a Human reading the problems out loud with a stopwatch running. The point is that performance is not a test of math ability, but of the PAS compensation in the Intellectual dimension.
  • Performance on Object Assembly depends on a human examiner with a stop-watch. Subjects who are modified in the Social Dimension (A/U) of the PAS under-perform on this test. In theory, this under-performance is due to social tension/anxiety, a very human interaction.

Therefore, while a language model AI, or even an AI driver robot with auditory/visual input might (and might not) perform very well and appear to be "intelligent", it does not process the "intelligence test" subtests in the same way as a human in that it does not (yet) include any ability to depend on the emotional state of the test taker in the testing environment.

This leaves the question of whether an AI drive robot could be programmed to have a particular PAS profile. While it would be straightforward to program an AI robot to perform the specific PAS subtests in a certain pattern of high/low scores, making a language model AI respond like a human with a certain PAS profile requires a level of "intelligence" not now known in AI. This "intelligence" includes factoring in emotions into responses, and those emotions depend on life history, what has happened in the last hour, the tone of voice of the questioner etc.

Conclusion?

Language model AIs are so different in function from the way humans interact with an intelligence test that we must consider them as qualitatively different from humans. They clearly lack many aspects of human intelligence. I believe that this difference will be very difficult to overcome because while we understand how digital computers and programs work, we do not understand how the human brain works when it comes to making decisions, even simple ones like how to answer a question or perform a task on an intelligence test.

This does not mean that humans cannot be fooled by a language model AI into thinking it is human (with an environment constrained to verbal input and output) or that a human cannot be deceived into placing unwarranted trust in the output of a language model AI. This latter matter is problematic because humans decide how much trust to place in a person based on experience which includes more than just verbal responses. Language model AIs do not provide most of the clues human used to estimate trustworthiness.

Finally, while I have touched on aspects of the human soul, emotions and our will, even the PAS does not describe another whole part of who we are as God's Creation - the spirit in man. Human science cannot even teach us how to tell if something is alive or dead. We intuitively know some things are alive, but a precise scientific criteria eludes us. This is all far beyond what digital computers do.

Examples of verbal questions ChatGPT

To use ChapGPT, the environment is different from an actual PAS(WB/G) test in that the question is typed rather than read. Some results are shown below to give the flavor of the response of this particular Chat language model AI. The results were done using https://chat-gpt.org/chat

Question	Answer
Comprehension: Laws necessary	Very long answer which would be scored 2 points
Comprehension: Taxes	Long discussion which is a 1 point answer.
Comprehension: Fire in Theater	Long answer starting with the 0 point response (shout fire). The long answer would be scored only a 1. The answer is first qualified with, "As an AI language model, I am not capable of experiencing situations as I don't have a physical presence."
Similarities: Orange/Banana	2 points (with many subsidiary similarities)
Similarities: Fly/Bush	The 2 point answer followed by all the differences. This was preceded by the qualifier, "As an AI language model, I do not have personal perspectives but a possible answer could be:"
Arithmetic: Airplane	Long explanation of the logic used to get the correct answer.