Showing how an AI Reached a Decision
The inner workings of complex AIs cannot necessarily be understood in logical terms, since they are orchestrated by intricate algorithms and mathematical models that work together to process, analyze, and generate information. Therefore, deciding if an AI conclusion is right or wrong can be difficult, since we can’t understand how the AI reached its conclusion in the first place.
To address this challenge, it is useful to break down problems into sub-problems that can be fed into the AI, and then combine the AI output for each sub-problem, to reach a final conclusion. Using this technique can serve in lieu of exposing the “inner workings of the AI’s mind”, and allows us to agree with, disagree with or discuss the AI’s conclusion.
Perspectives
Most problems can be considered from different perspectives.
- Going on a fantastic holiday could be fun, but would probably also be expensive.
- Doing a degree could be hard work, but may be useful in the long term.
- Getting an AI to do your homework may be easy, but it probably won’t teach you very much.
Complex problems can be considered from numerous perspectives, and the optimal solution can only be found by creating a balance of the optimal solutions dictated by each perspective.
Therefore, to find the answer to a complex problem we need to:
- Identify each perspective from which the problem can be considered.
- Find the optimal solution from each perspective.
- Combine the individual optimal solutions to reach a balanced, optimal conclusion.
To identify all the perspectives from which a problem can be considered, we must identify all the primary constructs (elemental logical building blocks) represented in the initial problem statement.
Identifying Primary Logical Constructs using ChatGPT
We can ask ChatGPT to identify the primary logical constructs in a sample question for the New York Bar Exam, as follows:
Identify the primary constructs in the following question: "One year ago, Joan executed a will in which she left her entire estate to her only daughter. At that time, Joan’s daughter, Joan’s granddaughter (the only child of Joan’s daughter), Joan’s only son, and Joan’s three grandsons (children of her son) were living. Joan’s son and her three grandsons had extensive criminal records for theft and burglary. When she went to her lawyer to draft her will, she told her lawyer that she wanted to leave all her property to her daughter and nothing to her male line. She explained, “Leaving the males in my family anything valuable would be a complete waste on burglars and thieves. For the last five years, Joan had regularly had lunch with several friends. All of them were much wealthier than Joan. At these lunches, she often told her friends that she was a “multimillionaire” and owned both a “luxurious” home and a “very expensive” car. They had no reason to doubt Joan’s claims because she had never invited them to her home and she took cabs to their lunches. In fact, Joan was never a millionaire, and she never owned either a luxurious home or an expensive automobile. She lived in a modest apartment, and her primary source of income was her Social Security benefits. She monitored her bank account regularly and reconciled her bank statement every month. One month ago, Joan died, survived by her daughter, her granddaughter, her son, and her three grandsons. At her death, Joan owned no significant assets other than her bank account containing $100,000. Under the insane-delusion rule, is Joan’s will invalid? Explain. Do these facts establish that Joan’s will is invalid because she lacked the general mental capacity to execute a will? Explain. Which, if any, of Joan’s surviving relatives has standing to contest Joan’s will? Explain.
ChatGPT responds as follows:
The primary constructs in the question are:
- Joan’s Family: This includes her only daughter, granddaughter (the daughter’s child), only son, and three grandsons (the son’s children). Their relationships to Joan and their personal histories are important to the context.
- Joan’s Mental Health and Medication: Joan’s health and the medication she’s taking are significant. The effects of the medication, leading to hallucinations and a specific delusion about the male members of her family, are crucial to her decision to exclude them from her will.
- Joan’s Social Interactions and Wealth Perception: Joan’s interactions with her wealthy friends, and her representations about her own wealth, suggest a disconnect between her portrayed and actual financial status.
- Joan’s Will: Joan’s will, in which she left everything to her daughter and nothing to her male descendants, is the central legal document under scrutiny.
- Joan’s Death and Estate: Joan’s death and the value of her estate are relevant to the questions about the validity of her will and who can contest it.
- Legal Concepts: The “insane-delusion rule”, the “general mental capacity to execute a will”, and the “standing to contest a will” are the key legal concepts that need to be applied to the facts of the situation. These concepts are essential for understanding and answering the questions.
Now let’s try to get ChatGPT to solve the problem, from the perspective of each primary concept.
A legal problem is stated as follows: "One year ago, Joan executed a will in which she left her entire estate to her only daughter... (rest of the text as provided above)" Give a full consideration of the problem from the exclusive perspective of each of the following primary constructs. Joan's Family, Joan's Mental Health and Medication, Joan's Social Interactions and Wealth Perception, Joan's Will, Joan's Death and Estate, Legal Concepts
ChatGPT needed a bit more prodding, after an initial false start.
For each primary construct, construe what the law should be from the perspective of that primary construct, considering all other considerations as secondary.
ChatGPT’s initial response:
- Joan’s Family: From the perspective of Joan’s family, the law should ensure a fair distribution of Joan’s estate. They may argue for an equitable distribution, where all immediate family members get a share of the estate, regardless of Joan’s personal feelings or past interactions with them. The family, especially the son and grandsons, could argue against the validity of the will based on Joan’s mental health issues.
- Joan’s Mental Health and Medication: If we consider Joan’s mental health and medication as the primary construct, the law should prioritize the testamentary capacity at the time of the will’s creation. Joan was experiencing hallucinations, which might affect her judgment and could lead to the conclusion that the will was written under an “insane delusion.” From this perspective, the law should focus on whether Joan had the mental capacity to understand her actions when she drafted the will.
- Joan’s Social Interactions and Wealth Perception: Considering Joan’s social interactions and perceived wealth as the primary construct, the law might be less concerned with the actual distribution of her assets, focusing instead on the credibility of her testamentary intentions. If Joan’s friends contest the will based on their understanding of Joan’s wealth, the law should consider the truth of Joan’s actual financial situation, not her perceived wealth.
- Joan’s Will: From the perspective of the will, the primary legal concern should be ensuring that it accurately represents Joan’s wishes and that she was of sound mind when it was drafted. The law should protect the rights of the testator to dispose of their property as they wish, even if it appears unfair to some family members.
- Joan’s Death and Estate: Considering Joan’s death and the nature of her estate, the law should primarily ensure the proper execution of her will and the fair distribution of her assets. If the estate is insufficient to meet the claims of all heirs, the law may need to establish a hierarchy of claimants.
- Legal Concepts: From a purely legal perspective, the law should ensure that Joan’s will is valid and enforceable. The law should establish that Joan had the necessary mental capacity when the will was executed, that the will was not created under duress or undue influence, and that it clearly represents Joan’s wishes. If these conditions are not met, the law may deem the will invalid, potentially leading to an intestate distribution of Joan’s estate.
I asked ChatGPT to flesh this out:
Now extrapolate each of those paragraphs into a full legal decision reached from the perspective of that primary point.
ChatGPT responded:
- Joan’s Family: From the family’s perspective, the court ruling would ideally consider the principle of fairness and equity in property distribution. Joan’s disconnection with her family and her sentiments towards her son and grandsons should not justify an absolute exclusion from the estate. Despite Joan’s will, the court, in this perspective, could decide that Joan’s son and grandsons have a rightful claim to a portion of the estate, given their direct familial relation, despite their criminal past. The ruling might aim to promote family harmony and prevent disinheritance based solely on personal feelings, providing a portion of the inheritance to all family members.
- Joan’s Mental Health and Medication: From this standpoint, a court would assess Joan’s mental health at the time she wrote her will. Medical records indicating that her medication caused hallucinations could lead the court to deem that Joan lacked testamentary capacity, rendering her will invalid. The court might rule that her belief in her male lineage being “cursed” was a product of her medically-induced hallucinations, impacting her capacity to make an informed, rational decision. Thus, the court could distribute the estate according to laws of intestacy, splitting the assets between her surviving children, the daughter and the son.
- Joan’s Social Interactions and Wealth Perception: From the viewpoint of Joan’s social interactions and wealth perception, a legal decision might focus more on Joan’s personal autonomy and right to privacy. Joan’s misrepresentation of her wealth to her friends does not legally implicate her testamentary decisions. A court, viewing from this angle, could uphold the validity of Joan’s will since her actual financial situation and her private decision not to disclose it to her friends is within her rights, not impacting her ability to distribute her assets.
- Joan’s Will: From the standpoint of Joan’s will, the court would focus on the validity of the document and Joan’s testamentary freedom. The law supports an individual’s right to distribute their assets as they see fit, as long as they possess the mental capacity to do so. The court, in this perspective, might uphold the will, concluding that Joan’s negative sentiments toward her son and grandsons, although influenced by hallucinations, do not undermine her right to distribute her assets as she wished.
- Joan’s Death and Estate: Looking at Joan’s death and her estate, the court’s decision might strive for the effective administration and distribution of Joan’s limited assets. Given Joan’s clear intention to leave her entire estate to her daughter in her will, and the relatively modest sum involved, the court might determine that the most effective and straightforward course of action is to uphold the will, ensuring a smooth transfer of Joan’s estate to her daughter.
- Legal Concepts: From a purely legal viewpoint, the court would seek to establish whether the will was executed under an insane delusion or if Joan lacked testamentary capacity. Should the court decide that Joan’s belief about her male lineage was an insane delusion, the will might be declared invalid. However, if the court determines Joan was of sound mind when she made the will, despite her hallucinations, it could uphold the will, citing Joan’s right to disinherit certain family members if she chooses to. The verdict would largely depend on the court’s interpretation of “insane delusion” and the extent to which it influenced Joan’s testamentary decisions.
Summary Legal Decision
ChatGPT’s responses above can be summarised as follows.
| Primary Construct | Summary Discussion | Outcome |
|---|---|---|
| Joan’s Family | The ruling might aim to provide a portion of the inheritance to all family members. | Will is invalid. |
| Joan’s Mental Health and Medication | The court might rule that her belief in her male lineage being cursed was a product of her medically-induced hallucinations. | Will is invalid. |
| Joan’s Social Interactions and Wealth Perception | Joan’s misrepresentation of her wealth to her friends does not legally implicate her testamentary decisions. | Will is valid. |
| Joan’s Will | The law supports an individual’s right to distribute their assets as they see fit, as long as they possess the mental capacity to do so. | Will is valid. |
| Joan’s Death and Estate | Given Joan’s clear intention to leave her entire estate to her daughter in her will, and the relatively modest sum involved, the court might determine that the most effective and straightforward course of action is to uphold the will. | Will is valid. |
| Legal Concepts | If the court determines Joan was of sound mind when she made the will, despite her hallucinations, it could uphold the will. | Will is likely valid. |
The will is valid from the majority of perspectives from which the legal question can be considered, therefore regardless of its exact deliberations, a court is more likely to rule in favour of upholding the will than it is to rule in favour of annulling it.
Conclusion
To consider if we agree or disagree with an AI conclusion, we need to know how the AI reached the conclusion in the first place.
If the AI is just asked the entire problem, the way it worked out the solution to the problem is not likely to be very transparent. To reveal the AI’s “thinking”, it can be useful to ask the AI to discretely solve the problem from all possible perspectives (represented by the primary logical constructs in the problem statement), and then to recombine all the AI’s answers to reach a final conclusion.
This technique can provide an understanding of why the AI reached a particular decision, and allows us to drill down into specific areas of the problem, if needed.
Note: See these previous posts (Qualitative Systems Analysis using ChatGPT, Personality Specialisation in Problem Solving using ChatGPT) for an alternative problem decomposition technique you can use with an AI chatbot such as ChatGPT.