How Sterotypes are Learned

Artificial Intelligence (AI) provides a model for how stereotypes develop.  Stereotypes are different than prejudice. They are learned differently and seem easier to change. A stereotype lacks the emotion that characterizes a prejudice. 

AI is a branch of computer science that aims to perform tasks or solve problems that only humans seemed able to do.  It is a rule-based inference system that should be able to provide justifications for its answers through its mathematical rules that determine how it learns, sorts, quantifies uncertainty, defines how “knowledge” is defined for various domains, handles unclear parameters, estimates probabilities, and makes decisions. Current AI systems are not transparent to the user and don’t reveal how it makes decisions, although the code can be inspected for this information.

GPT4 – Open AI Model

A significant new AI advance is GPT-4 (Generative Pre-Trained Transformer). It is a sophisticated language model that was trained on textual content from books, text, Wikipedia, articles, and other online sources. It can produce text that appears to be written by a person.  One estimate is it can recognize close to 100 trillion parameters complex patterns in natural language and incorporates state-of-the-art machine learning. The program is able to handle nuanced instructions as well as produce human-like text and speech. 

However, GPT-4 it is not a search engine, and its answers are only as good as the information used to train it. Microsoft has linked GPT-4 to its search engine, Bing, but it cannot do fact checking. GPT-4 can be used with proprietary databases like medical or legal databases or copyrighted material for prevalidated information, but these databases require a paid subscription.

I tested the trustworthiness of Bing Chat on a problem that has faced gerontology since nursing homes began to expand in the 1960’s. Why do Blacks utilize nursing homes so much less than Whites despite higher rates of chronic illness, disability at earlier ages and less at-home resources? The 1970’s answer was that Blacks preferred care at home (as if Whites didn’t), or that they couldn’t afford long-term care.  My eyes were opened when one middle class Black family member that could afford nursing home care, told me they wouldn’t send a dog to the Nursing Homes that were available to them because of discrimination.  When I queried Bing Co-Pilot which uses GPT-4, the answer essentially repeated my question and said it was due to cultural preferences, language barriers (for Hispanics), and lower income. It did warn that “It is important to note that these are complex issues with multiple factors at play. Further research is needed to better understand the underlying causes of these disparities.” In other words, it did not mention discrimination and barriers to care.  

Limitations of GPT-4 that mimic stereotyping

Three main limitations have been noted for AI. One of the problems is issues with the algorithm, another is the lack of completeness of the learning database, and third is its inability to do fact checking. In traditional science writing or journalism, fact checking is done before publication. For information on the World Wide Web, fact checking is not done until after publication, if ever.

With people, any situation where fact checking is done needed post-hoc, checks often rely on knowing the source of information or who told us about it, like family, friends, or experts, rather than direct checks. If it came from a trusted source is is usually thought to be OK. GPT-4 selected training databases like Wikipedia. Since our trusted sources may not have fact checked either, the data is still prone to error. In all cases, bad data leads to wrong outcomes, like the old adage “garbage in, garbage out.”

AI and people must often decide things without having all the information needed. When that happens, responses may be biased or even seem illogical.  It would be nice if AI programs could give a degree of confidence about its answers, but that would be a long way off. 

Another limitation is its lack of “common sense,” as defined as prudent judgement based on simple perceptions.  Common sense requires abstract reasoning and GPT-4’s abstract reasoning ability against a benchmark, ConceptARC, was reportedly below 33% in comparison to at least 91% by humans. 

Insufficient Information

When AI is asked to make decisions before it has all the facts or even knows the number of possible outcomes, it estimates the highest probability option at each decision point from what it knows.  This is a sophisticated way to say it makes an educated guess. 

The less information there is, the less certain the prediction is (level of confidence), so actual outcomes must constantly be assessed, and course corrections made.  The more wrong information there is, the less accurate the prediction will be, and some information is so inaccurate that it should not be given the time of day. There are not alternative facts, only differences in what the facts might mean.

AI will not help in giving solutions to man-made crises like institutional racism or gun deaths as long as we restrict or filter key information. We cannot whitewash history, censor teachers, ban woke information, or prevent research or discussion on gun use if we want help from AI in helping solve these issues.

Illogical Reasoning

Bias can be due to unsound decision-making processes, i.e. algorithm erros.  We must remember that AI algorithms are written by people, and people are often illogical or don’t think through things fully.  Mistakes can easily be incorporated into the code (bugs and bias).  Some form of checking should be applied. AI labels this truth-maintenance. One method is a simple justification-based system where getting the predicted outcome validates our assumption. If the program assumes all Blacks are bad and a black man commits a crime, it has confirmed its assumption.  A second method is an assumption-based system in which there may be several assumptions which are combined into related sets.  If the predicted outcome is confirmed by the set of assumptions, it is considered a valid set of assumptions even if the set has contradictions in it.  A third method is a logic-based system.  The concepts must be clear, and relations logically connected.  If the assumption is not logical, it is rejected.  From what one AI researcher told me, Truth Management systems are not usually applied because of real-time processing constraints – it takes too long.

Logical fallacies require some elaboration since logic is not usually taught in school anymore.  There are many categories of potential logical fallacies: formal fallacies (an error in the argument’s form), propositional fallacies (compound propositions that aren’t logically connected), quantification fallacies (premise quantifiers differ from conclusion quantifiers), formal syllogistic fallacies (illogical argument), and informal fallacies (lack of grounded premises). 

Because illogic and bad information can be pervasive, the best advice still goes back to a 14^th century philosopher and theologian, William of Occam who gave a commonsense rule for problem solving: Entia non sunt multiplicanda praeter necessitatem, which translates as “Entities must not be multiplied beyond necessity” (Wikipedia).  This is also known as Occam’s razor or the principle of parsimony – why invoke a complicated explanation or contrived solution when a simple answer is evident?

“Hallucinations” and Truth Management

AI does not really understand its output and GPT-4 has been known to “hallucinate.” Hallucinations for GPT-4 means its answers may not be based on its training data (i.e. it makes its own inferences) and contradicts its previous statements or user’s prompts. Critics say this is akin to making things up. 

Microsoft researchers suggest GPT-4 may exhibit cognitive biases such as confirmation bias, anchoring (i.e. being more likely to select highlighted information called hyperlinks above other text) or give more importance to certain information than it should (base-rate neglect).

Unlearning a Stereotype

Predictions improve as the information improves. There is a mathematical principle called the law of large numbers where the observed value becomes closer to the true or expected value as the total number of observations approaches the entire amount.  The answers will be more accurate if we have complete information. For people to apply this rule, they need to stay open to new experiences, keep an open mind, and keep learning while building on what we already know. This scaffolding of knowledge is a reason to know history – it keeps things in perspective.  

The American philosopher George Santayana warned that “Those who cannot remember the past are doomed to repeat it” (The Life of Reason: Reason in Common Sense). We need to learn from history and science, and reflect on what we know. If we don’t, an anonymous person added, “Those who learn history wrong are just doomed.”