August 3 Leg Act News
- 5 days ago
- 4 min read
Highlights
AI Ethics: Trust by Mark J. Norton
More and more, AI applications are becoming the dominant way to access information. Do you need to know when the US Constitution was fully ratified? Ask ChatGPT (all 13 original states ratified the Constitution by May 29,1790). Want to know what causes the common cold? Ask Gemini (coughs and sneezes). Need a little application that converts English to Metric measurements? Ask Claude Code to write it. Whenever you seek an answer to a question and get an answer, another question arises: Do you trust the answer?
In the old days (prior to a few years ago) if you got your Constitution answer from Wikipedia, you might trust it because of the way Wikipedia worked. In Wikipedia, anyone can add information to update the file on a topic. If the answer matches other things you read, it is likely accurate. On the other hand, if you asked your doctor about head cold treatment, you’d probably trust his or her answer without much further investigation, given the doctor’s years of training and experience. In short, whether you trust an answer to a question depends on the source.
Large Language Models (LLMs) are usually trained on a very large body of source data (more likely, many sources). The result is a very large set of statistical relationships represented as weights between artificial neurons in an artificial neural network (ANN). In order to generate a response from an LLM, the prompt is used to compute a “seed” that is pushed into the network. The seed is used to “predict” the next token in the (eventual) response. This cascades through the ANN and eventually results in a response. [Warning: this is a very simplistic description of what actually happens.]
Because of the statistical representation of the network, there is a chance (very small in some cases) of taking a “wrong turn” resulting in a response that might not be factual. However, this is also the source of an LLM’s creativity. This is how it suggests “new” things. Sometimes a random process produces interesting (and new) results.
The likelihood of a factual response depends (in large part) on how the model was created. If many sources say something like “May 29, 1790, was the date the US Constitution was fully ratified,” then the model is very likely to generate that answer given certain kinds of prompts. However, it is not guaranteed 100%. There is a small chance (statistically) that it might respond with “May 28, 1790.” It is not lying; it just generated a false response.
Another thing to be aware of is source bias. For example, LLMs trained on the Internet and other sources in English are inherently biased towards Western Medicine. Models trained on Chinese sources might give very different responses based on the extensiveness of Traditional Chinese Herbal Medicine in China. Bias cannot be avoided, so the user must be aware of potential leanings and take them into account when evaluating the truth of a response.
If a chat application gives a false response, is it behaving unethically? In my example, no. The app has no thought process. There was no intent to deceive. Essentially, it relies on you to decide if the response is factual … or even merely useful. The bigger picture is more complicated. If a company releases an application that intends to provide answers to medical questions knowing that it has serious biases that the company chooses to ignore, THAT might be considered unethical. Remember, the source data may not have supplied enough information or there may be bias in the source data.
Sadly, there is little to prevent corporations from behaving unethically and consumers have little or no recourse. Keep that in mind the next time your chat application recommends that you buy a particular stock.
Action Alerts/Events
TCDC’s Premier Event: The Annual Picnic and Pig Roast.
Mark your calendars for August 20 from 5 to 8 pm, Hickories Park, Owego. Tickets are available for purchase via TCDC members and ActBlue. For Ticket purchase and auction items, checks are preferred due to limits on how much can be collected in cash donations. Comptroller DiNapoli, Congressional candidate Aaron Gies, and New York State Assembly candidate Andy Fagan will speak.
The Tioga County Fair - Aug 12 - 15, 12 noon until 6 pm (11 am until 6 pm on Aug 15).
A volunteer is still needed for Friday Aug 14 from 3:30-6:00 and Sat Aug 15 from 11-1:30 and 2 to 4 pm. Contact Mary Schwarz at msp5@cornell.edu or TCDC at tioganydems@gmail.com.
Meet and Greet fund raiser for Candidate Andy Fagan on August 16th, 2-4 pm at the Newark Valley Noble Room at 9 Park St. Questions: Jim Tornatore 607-205-2616.
The next TCDC meeting will be on Tuesday September 22 (the 4th Tuesday of September!) We will have the required biennial election of TCDC officers and Town Chairs before the general meeting. PLEASE MARK YOUR CALENDERS TO MAKE SURE YOU HAVE THE CORRECT DATE.
Call Langworthy, 202-225-3161, Schumer, 202-224-6542 and Gillibrand, 202-224-4451.
Demand that the truth about all casualties and deaths from the Iranian war be reported and counted including the death of four soldiers in Jordan. Demand that they vote to end the war in Iran.
If you are interested in joining the Leg Act Committee or want to submit something to our newsletter, send an email to TCDC at TIOGANYDEMS@GMAIL.COM or MARITA FLORINI at FNPMAF1@GMAIL.COM An educated citizenry is essential for a healthy democracy.
