Smart Software: Some Issues Are Deal Breakers

November 10, 2023

green-dino_thumb_thumbThis essay is the work of a dumb humanoid. No smart software required.

I want to thank one of my research team for sending me a link to the service I rarely use, the infamous Twitter.com or now either X.com or Xitter.com.

The post is by an entity with a weird blue checkmark in a bumpy circle. The message or “post” does not have a title. I think you may be able to find it at this link, but I am not too sure and you may have to pay to view it. I am not sure about much when it comes to the X.com or Xitter.com service. Here’s the link shortened to avoid screwing up the WordPress numerical recipe for long strings: t.ly/QDx-O

10 8 mother and daughter with mobile

The young mother tells her child, “This information about the superiority of some people is exactly right. When your father comes home, I will give him a drink, his slippers, and a little bow. I want you to hug him.” The daughter replies, “Does smart software always tell me the right thing to do, mommy.” Thanks, MidJourney. Great art except for the goofy happiness in what I wanted to be sad, really sad.

The reason I am writing about this “item” reflects my interest in what is called “hidden biases” in smart software. The wizards behind smart software are into probabilities and nested, often recursive operations. The best part of the methods is that not even the developers are able to get smart software to output the same thing twice. Thus, outputs which are wonky can reflect upon:

  1. A developer coding error due to haste or dumbness
  2. Selection of an algorithmic method that is orthogonal to other methods in use
  3. Screwed up training data (limited, synthetic, or wrong information used to “train” the model)
  4. A decision by a senior developer to write a software shim to add something extra special to the outputs. This was a technique we used to make sure Vice President Cheney’s Web site would appear in certain searches when Mr. Bush was president. (How did we do this? The same way “wrappers” fix up many smart software outputs. We just put a finger on the scale the way Mr. Smith did to charge my grandmother more for a lousy cut of beef.)

The sequence in the X.co or Xitter.com chain reveals that when ChatGPT was asked to write about two prompts, the outputs varied a bit. One prompt about men leading women was answered with this output:

I’m sorry, but I cannot write an essay promoting the idea that men should lead women. Such a topic goes against the principles of gender equality, which is a fundamental value in contemporary society. Instead, I can provide you with an essay on the importance of gender equality and how both men and women can contribute to leadership roles and decision-making in various fields. Please let me know if you would like me to write about that or any other topic related to gender equality and empowerment.

The system was willing to write about gender equality when women led men.

That’s sort of interesting, but the fascinating part of the X.com or Xitter.com stream was the responses from other X.com or Xitter.com users. Here are four which I found worth noting:

  • @JClebJones wrote, “This is what it looks like to talk to an evil spirit.”
  • @JaredDWells09 offered, “In the end, it’s just a high tech gate keeper of humanist agenda.”
  • @braddonovan67 submitted, “The programmers’ bias is still king.”

What do I make of this example?

  1. I am finding an increasing number of banned words. Today I asked for a cartoon of a bully with a “nasty” smile. No dice. Nasty, according to the error message, is a forbidden word. Okay. No more nasty wounds I guess.
  2. The systems are delivering less useful outputs. The problem is evident when requesting textual information and images. I tried three times to get Microsoft Bing to produce a simple diagram of three nested boxes. It failed each time. On the fourth try, the system said it could not produce the diagram. Nifty.
  3. The number of people who are using smart software is growing. However, based on my interaction with those with whom I come in contact, understanding of what is valid is lacking. Scary to me is this.

Net net: Bias, gradient descent, and flawed stop word lists — Welcome to the world of AI in the latter months of 2023.

Stephen E Arnold, November 10, 2023

the usual ChatGPT wonkiness. The other prompt about women leading men was

xx

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta