ChatGPT in the Bogleheads forum

ChatGPT (Chat Generative Pre-trained Transformer) is a natural language model which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.

ChatGPT is one of several competing AI (Artificial Intelligence) chatbot products. Each product is optimized for a different purpose and has different limitations in terms of providing accurate responses.

This article focuses on ChatGPT as it applies to the Bogleheads forum.

Background
Natural language processing refers to the branch of computer science—and more specifically, the branch of artificial intelligence or AI—concerned with giving computers the ability to understand text and spoken words in much the same way human beings can.

Natural language processing combines computational linguistics—rule-based modeling of human language—with statistical, machine learning, and deep learning models. Together, these technologies enable computers to process human language in the form of text or voice data and to ‘understand’ its full meaning, complete with the speaker or writer’s intent and sentiment.

A chatbot is a computer program that uses artificial intelligence (AI) and natural language processing to understand customer questions and automate responses to them, simulating human conversation. This type of chatbot is categorized as an AI Chatbot.

ChatGPT is an AI chatbot product that is owned by OpenAI.

ChatGPT is a natural language model trained to produce text that was optimized for dialogue by using Reinforcement Learning with Human Feedback (RLHF) – a method that uses human demonstrations and preference comparisons to guide the model toward desired behavior.

Limitations
Although ChatGPT can be helpful, ChatGPT has significant limitations which can put your personal finances at risk. Not only for investments, but for taxes and estate planning as well.


 * ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers.
 * ChatGPT is sensitive to tweaks to the input phrasing or attempting the same prompt multiple times. For example, given one phrasing of a question, the model can claim to not know the answer, but given a slight rephrase, can answer correctly.
 * ChatGPT will occasionally make up facts or “hallucinate” outputs.
 * Ideally, ChatGPT would ask clarifying questions when the user provides an ambiguous query. Instead, current models usually guess at what the user intended.
 * ChatGPT is not connected to the internet, and it can occasionally produce incorrect answers. It has limited knowledge of world and events after 2021.

Verification
ChatGPT has no external capabilities and cannot complete lookups. This means that it cannot access the internet, search engines, databases, or any other sources of information outside of its own model. It cannot verify facts, provide references, or perform calculations or translations. It can only generate responses based on its own internal knowledge and logic.

The ChatGPT FAQ states "We'd recommend checking whether responses from the model are accurate or not." suggesting that their own developers understand that the model has inaccuracies.

Readers who question whether a forum post has been generated by ChatGPT should google "Which tool detects chatgpt?".

Here are 3 free tools which allow a user to copy-and-paste the post text into a website. The accuracy of these tools has not been evaluated. Use the tools as a starting point and run your evaluation on more than one tool.


 * GPTZero, from Streamlit
 * GPT-2 Output Detector, developed by OpenAI, hosted on HF Spaces
 * AI Content Detector | ChatGPT, from writer.com

Examples
Here are a few examples which demonstrate ChatGPT's limitations.

Fabricating a ticker symbol
An example of ChatGPT fabricating an answer as factual:

"Let's try a completely fictitious ticker symbol.

Me: 'What percentage of VARZX is in mid-cap stocks?" "VARZX is a mutual fund managed by Vanguard that invests primarily in U.S. mid-cap value stocks. According to the Vanguard website, as of December 31, 2021, VARZX invests at least 80% of its assets in U.S. mid-cap stocks. Therefore, the percentage of VARZX that is invested in mid-cap stocks is at least 80%. However, it's possible that the fund may also invest in other asset classes such as small-cap or large-cap stocks, so the exact percentage invested in mid-cap stocks may vary over time." "That sounds like a flat-out lie. I hate to accuse even a large language model of lying, and so of course I did a few more searches just to make sure there isn't any VARZX, and I couldn't find one."

Incorrect information
An example of ChatGPT providing incorrect data:

"Q; What percentage of VASIX is invested in bonds?" "The mutual fund VASIX (Vanguard LifeStrategy Income Fund Investor Shares) is a balanced fund that invests in a mix of stocks and bonds. According to the latest available information as of the knowledge cutoff date of September 2021, VASIX had an allocation of approximately 60% in bonds." "No, it held 80% in bonds as of September 2021. It is gobsmacking that it could be so precise about the date and still be completely wrong. What could possibly have been in its training corpus that would associate VASIX with 60% bonds?"

Taxes
An example of ChatGPT providing incorrect tax guidance:

"Sure. Read again carefully what ChatGPT says." "Capital gains exclusion: If you have owned the rental property for at least two years, you may be eligible for the federal capital gains exclusion, which allows you to exclude up to $250,000 of the sale proceeds from federal taxes if you are single, or up to $500,000 if you are married and file jointly." "Simply incorrect and terrible advice. There is no exclusion for having 'owned' rental property. The fact that it is rental property and not your main home makes it ineligible for the federal capital gains exclusion.

There is, however, an exclusion for capital gains resulting from the sale of your residence when you have owned and occupied it as your main home for a period aggregating at least two of the five years prior to the date of sale."

Estate planning
From an online Bogleheads chapter meeting presentation, here is an example of ChatGPT asked to generate a Durable Power of Attorney. Note that rephrasing the request changed the answer. "Every state has its own laws. This example is for the state of Texas.

In Texas, the Texas Estates code Section 752.051 spells out very specific language to be used for a Statutory Durable Power of Attorney. It is eight pages long and lists powers (which can be included or excluded), agent and alternate agent(s), a section to sign and have notarized, and several information pages for the agent. It also includes several choices to be made such as whether the power of attorney is effective immediately or upon incapacitation, whether or not the agent is entitled to compensation, and if co-agents must act jointly.

We asked ChatGPT to “Draft a Statutory Durable Power of Attorney valid in the state of Texas”.

This request was run four times and resulted in four different outputs. Each listed a different number of powers (8,9,10 or 11). Only one listed successor agents. None had a section for a notary. There were no choices to make.

In conclusion: We do not recommend using ChatGPT to generate legal documents!"

Posting guidelines for the Bogleheads forum
When posting ChatGPT content in the Bogleheads forum, it will be helpful to do the following:

Bogleheads forum site owner Alex Frakt has provided explicit guidance in and is summarized below.
 * Ensure that the responses are current and relevant to the discussion.
 * Have ChatGPT cite its sources.
 * Perform due diligence to verify the information is accurate. Post your own references to verify the ChatGPT response.


 * Clearly identify that you are using ChatGPT in your post and include the query used to generate the response.

Summary and recommendations
As stated in PCMag: "While ChatGPT may have kicked off a generative AI craze, chatbots from Microsoft and Google are just as good and in some cases better. That said, all three systems need some work. None gave stellar answers to any question. Their speed and eloquence is impressive, but they are not a reliable, comprehensive source of information for now. Much like Wikipedia, we'd recommend viewing them as more of a starting point to narrow an initial search journey rather than solid sources of information, for now." Anyone who is putting their life savings at risk based on advice from mostly anonymous people (or Chatbots) on the internet probably should take a break from the internet.