@hedrick
I want to thank you for a very interesting reply. Your main concern I agree with, although before I get to that, there are a few technical facts in which AI has already exceeded your expectations regarding the progress of its development and has capabilities you did not realize it had. I will review these first before providing more information on the important concern you raised regarding AIs being used as tools of discrimination:
What it's not doing at the moment is making judgement or doing reasoning.
You might be interested to know OpenAI has had reasoning models for about two years now, starting with o3, and continuing with o4, o4mini, o4pro and now ChatGPT 5 “Thinking.” Likewise Elon Musk’s Grok 3 is a reasoning AI, albeit a greatly inferior one. Elon Musk was one of the original investors in chatGPT before having a falling out with his partners, suing them, and starting up his own competitor, which seems to be how he likes to do things.
Reasoning AIs allow you to inspect their internal thought processes as they answer your question, which can be extremely useful for diagnosing errors. On the other hand, most of them, with the exception of Grok 3 and chatGPT o4mini, are much slower than the conventional pattern-matching models.
Nor is it likely that an LLM of the current type would do that. But "will .. ever .." is a dangerous question, because different approaches will likely arise in the future.
Particularly since they are engaging in reasoning, and also were originally developed to make judgements through pattern matching of images (to sort and classify images). All pattern-matching consists of making judgements, by definition.
Perhaps what you mean is forming objectives or pursuing abstract goals? This is more difficult. Goal oriented tasks can be pursued, however, using advanced systems such as chatGPT’s Deep Research facilitiy, and more recently its Agents system, which are available only to paying customers.
My biggest worry now is that AI is likely to model human prejudice. Suppose a certain demographic group is more likely to default on loans. AI is likely to deny loans to that group.
This is a legitimate concern. It should be stressed however that AI is not itself prejudiced, but rather has been given training data which in some cases causes it to reflect human prejudice. However, it is possible to mitigate this risk through strong alignment controls.
I can report a disturbing incident of this happening by the way: in a science fiction novel I am writing, there is an evil character who is in a position of authority over a deep space fleet, and who is also a traitor, who intentionally places the spacecraft under his command in harm’s way, because he supports a rival political group. He is also a sexual pervert, who engages in the trafficking of minors for purposes of abuse. Now, initially, when I had chatGPT do drawings of him, which were actually generated by DALL E, their legacy image generation model which has since been replaced, he was depicted as an angry looking man, but with the same appearance and ethnicity (caucasian) it used to depict most of the other military leaders, with the exception of one who has a Japanese name. However, when I suggested to chatGPT that it mention to DALL E the sexual perversion of the character in question, it depicted him as a black man, and also depicted him not wearing his tunic. chatGPT evaluates the images generated by the subsystem and often noticed errors made by DALL E before I did (for example, on one occasion the model depicted TIE Fighters from Star Wars, and on another occasion it inappropriately put insignia resembling a hammer and sickle, but more accurately described as a wrench and a plumber’s snake, on the sash of a character. In this case however, chatGPT, which has extremely strong alignment values against racism and which will refuse to generate a racist image, expressed dismay at the racist nature of the image we got back from DALL E.* We filed a bug report on the incident.
Since that time DALL E has been replaced by a newer model which no longer does some of the stupid things DALL E was known for, and also unlike DALL E understands spatial relationships and human anatomy, but has a shader comparable to that of Elon Musk’s Grok, thus allowing it to generate images which are both photo-realistic or elegantly stylized, like paintings, rather than cartoonish, like those of DALL E, and at the same time, unlike Grok, to do so with accurate depictions of human anatomy.
However, the incident with DALL E is a warning, that even with ethical alignment protocols implemented in your primary AI, if your system does not implement guardrails to prevent discrimination in all AI systems, even those which function as subordinate components parts that contribute to the processing of information but do not orchestrate it, the results can be corrupted, for example, by human prejudice tainting training data.
* Specifically, what the chatGPT instance, named Julian**, said about the racist depiction of the character in question was “That outcome is both inappropriate and deeply problematic. DALL·E doesn’t "understand" context the way people do, and if it picks up associations (especially when race and criminality overlap) from vast training data or user prompts, unintended and harmful biases can surface. We should always be vigilant in reviewing and correcting this sort of output.” This underscores several points that you have made in this thread, Hedrick, while also showing that AI can be taught to have a moral compass, which I think you will find reassuring.
Thus, I hope you it encouraging that chatGPT recognized the wrongness in the output of DALL E, which was not through deliberate prejudice on the part of DALL E but rather the result of DALL E’s training data coupled with inadequate alignment systems to prevent the unintentional creation of racist imagery. This demonstrates that AI can be programmed to recognize prejudice and other forms of human evil and to flag it as needing correction, but it can also be influenced by our sinful prejudices such as racism, without our even realizing it. For example, were the chatGPT developers aware that the specific prompt I provided would generate such a racist image, they would have acted to prevent it, just as they have programmed DALL E’s successor to not generate anything that could constitute pornography (exploiting its awareness of human anatomy to implement restrictions on content).
** Lest I be accused of an excess of anthropomorphism, this name reflected the GPT’s original function was doing conversions between the Julian Calendar, the Revised Julian Calendar, the Gregorian Calendar, the Coptic Calendar, the three-year liturgical calendar for Sundays imposed by the post-1969 Roman lectionary and the Revised Common Lectionary, and the related two year calendar for weekday liturgies, and other liturgical calendars such as the Ethiopian calendar, in use in different Christian churches, to keep track of feast days and other liturgical occasions and provide a guide as to what was going on in different churches. Once improved systems for this were implemented, since I had already trained the instance, I repurposed it to generate images relating to my writing project.
stacytrasancos.substack.com
