This is a quick overview of a few of the issues.
Bias - Training data comes from the global north and focuses on the global West. Be aware when you ask for pictures of doctors, for example, and all it gives are white guys.
Crticial Thinking - Critical thinking is now more important than ever. Generative AI seems the easy way out, and it can help in many ways. However, it is a problem if used uncritically as the major source of information..
Copyrightt - Is the AI giving you copyrighted information without citing it? How would you know?
Currency - if your topic depends on the absolute latest information, what year was the dataset that was used for training?
Dependence on AI - If we become too reliant on AI, we can lose our creativity, critical thinking and human voice.
Economic Inequality - AI has the potential to further benefit the wealthy and the big corporations who have the money to develop AI, and the potential to eliminate "low-skilled" jobs. It will also create jobs, but who gets them?
Environment - Huge data centers take a lot of power, and they have to be kept cool.
Informed Consent - similar to copyright, but different. Some information is given on the condition it stays with the person it was given to.
Integrity
Low-Paid Labor - Who trains the AI? Who labels AI responses? Low-paid labor.
Misinformation/Hallucinations - AI gives a statistically plausible answer, not the correct one, and you have to know enough about the subject to spot them. If the data it trained on has misinformation, the AI will confidently give you an incorrect answer. It also gives wrong information in what is called "hallucinations." Claude has said that reindeer herding was a popular occupation for indigenous Mexican men. Why? Because Claude lacked the knowledge to know that a source on popular occupations for one group did not apply to another. ChatGPT has said that Halley's Comet came by in the 1930s when that is impossible because Halley's comes by every 72-80 years and had a famous appearance in 1910. Halley's Comet was a statistically reasonable answer but it was not the correct one.
Training data for Chat-GPT3 appears to be:
Common Crawl (a snapshot of the internet is captured and put into a bucket) - 60% of the training mix
WebText 2 database (includes Reddit with three or more hits!) - 22% of the training mix
Books 1 and Books 2 (internet-based books) - 8% and 8% of the training mix
English-language Wikipedia - 4% of the training mix
Source: "Language Models are Few-Shot Learners." by Brown, Tom B., Benjamin Mann, Nick Ryder, et. al.. From Advances in Neural Information Processing Systems 33 (NeurIPS 2020). Provided by Trudi Radtke.
If it's wrong and you know it, tell it so.
Privacy - What happens if I put in private information? (Don't). Where is the funding coming from and how does that impact the situation?
Security
Tool Transparency - none. How does the GenAI arrive at its output?
General
The audience includes business owners, but the issues are applicable to everyone.
