- Nov 26, 2019
- 16,496
- 8,756
- 51
- Country
- United States
- Gender
- Male
- Faith
- Generic Orthodox Christian
- Marital Status
- Celibate
If you ask AI two of the same questions just worded differently, sometimes you’ll get two different answers.
That’s entirely correct, and as a result its possible to fish for questions that produce the answers you want, which is another reason why using AI in an appeal to authority is inherently an example of the Appeal to Unqualified Authority logical fallacy. It’s the same as citing a Wikipedia article - the article might be right, but it might also contain biased information put in by an opinionated editor, and in some cases, the editorial decisions of Wikipedia regarding major articles are themselves highly controversial. But Wikipedia and an AI system can be used to find sources, provided one is willing to follow the breacrumbs to the original source document and verify that it says what the AI system thinks it says.
By the way, because the answers and behavior of AI systems can vary so dramatically (for example, several of my custom GPTs do things which are impossible on Google Gemini or Grok), I think we should avoid speaking of AI as a singular entity, and we should also be careful about using the term AI to refer to LLMs, which are the specific subtype of AI that all of the popular AIs use for most of their processing (a few, like chatGPT, also engage in reasoning, and use routing, so that math questions for example are not processed by the LLM, which can solve them but is extremely inefficient compared to the arithmetic logic unit built into every CPU and the floating point capabilities of every GPU; a classic LLM does not perform reasoning in transparent steps like chatGPT o4, 5.1 Thinking or Grok 3, but rather, just relies on the LLM component.
AI extends to many systems beyond the current popular LLM-based systems, however, and also predates it by a number of years; there are much simpler AI systems going back to the “fuzzy logic” used in the 1970s to control elevators in large skyscrapers, for example, as an extremely primitive example. The most common AIs in use are the AIs used in video games to control the behavior of adversaries - for many years, gaming was the only field where any active work on AI was being done, during the “AI winter“ of the mid 1980s to the mid 2000s.
Upvote
0