@Colin Bagley Now how can we prompt engineer to get this right more often. (I'd guess we would all be trying to figure that out.). May be a limitation of the current LLMs. They are not currently designed to self fact check before answering. Maybe that is something to add to these master prompts. Say that it is an excellent fact checker who always sites its known reliable sources. I don't have time to play with that right now. However, if some one else has a chance. Would love to hear more about your discoveries.

Something like this helped a bit. "You are the smartest book keeper who has every book and always checks your facts before responding. I will ask some questions, your job is to answer with actual passages from relevant books to all those questions. Give your answers in a table format, example - Passage, book name, book publishing details, how to use it, key learning. Are you ready for my first question?" This seems to get better cites and stops answers before you ask a question.

There are really two or three reasons that these bans might make sense. 1. The inability to control AI hallucinations, and the concern about turning over agency to the models. This is similar to bringing any new employee on. One has to monitor, and bosses need to evaluate if the models are acting in alignment with organizational values. I believe that this will be human judgment calls. The person who puts it in place needs to be responsible for outcomes. However, organizations have not yet figured out these norms. This same kind of problem occurred when PCs came into organizations. It happened again when the internet came to organizations. 2. There is active litigation against some models like stable diffusion that claim copyright infringement. If an organization uses contentment created with these models, and copyrightable content was used in the creation of the model. Is it fair use to use creations based on these models. I’ve come to believe that just like the camera one can use a model to infringe copyright content, or not. The model is a tool with potential for original creation or infringement. But each work must be weighed on it’s own merits. However, organizations have not come to an understanding of these norms. 3. Finally, the content models are trained on will cause inherent biases. Models are intentionally bias reproducers. Understand those biases, and being ok with the alignment of those biases with organizational norms is an important organizational decisions. For example I was interacting with GPT chat and used an example with Children's toys. And when I change the gender of child in the example the toys changed to stereotypical boy or girl toys. From an organizational point of view, is that fair an appropriate. I would guess that most human workers would have the same biases. However, do we want to hold our technology to a different standard? Finally, a bonus consurn, come in the area of customer privacy. Most organizations do not have the ability to create models like GPT chat from scratch. If an employee uses one of these models, are the questions being used by the vendors creating these models. Are we allowed to do this. Are we giving an unwanted advantage to the providers of the models?

@Birgitta Granstrom I agree, I’m not clear that an individual employees can work around such biases. It is some where between hard and impossible for employees to know what they do not know. That is why diversity among staff working on such questions is very important from my point of view. Imagine a thought experiment regarding model construction. If we try to curate information going into the creation of models, we are going to induce biases and blind spots in the models. You could imagine several model who’s input content were curated by conservative Christians, Jewish groups, Muslim communities, and the atheist communities. And then make those models available to the other communities. If members of one community asks questions around hot button subjects of race, gender and sexuality. You could imagine that the results of the same question would be very different on each of these models because of the curation by each of the groups on input content. Would an organization like Planned Parenthood want to use a model that was curated by the Christian right or vice versa? These are some of the Biases that are inevitably going to come up. My inclination is that I’d like a model trained on all of these perspectives. However, I’d guess that some others would prefer the curated versions if they were in alignment with their norms and beliefs. The questions then are around awareness, making choices about resolutions of problems when they come up. Choices around which model(s) to use that are in close enough alignment with organizational norms. All models are going to be biased, just like all people have their biases. The question is are the biases in alignment with the values of humans involved, and can we make models that are gracious and kind toward other perspectives? Just a few thought.

@Etienne Reussner shortening the prompt should also keep the cost of using the resulting tool down. It appears that open ai is charging by the token. And if you are running a prompt thousands or 10s of Thousands of time this could get expensive.