This article is about artificial intelligence and trust. That will require some patience, as the first half deals mostly with a challenge of a bygone, analog, age. But that foundation is necessary.
Early in my career, there was a volume of injured workers who alleged entitlement to a stream of benefits defined in the workers' compensation law. Some were measured in weeks or months, but permanent total disability was a stream that extended for years. In order to place a reserve value on those cases, the lawyer had to evaluate both the probability of an award of such benefits and the present value of the stream.
Present value was a term that I came to the practice already understanding. The idea is that the "cost" of paying someone $400 today is different than the cost of paying them $400 in 20 years. If I agree to pay you $400 tomorrow, that costs me $400. but if I am to pay you $400 on October 1, 2045, I can invest some money today and it will earn interest for years before I have to pay you that money.
It is important, therefore, to know the interest rate. And that is somewhat unpredictable. Sure, there are financial vehicles that may provide that predictability. A good example would be a 30-year fixed-rate mortgage. The day I wrote this post, that rate was 6%. The Internet told me that to make that payment in 30 years, I would need to invest "approximately $121.91 today."
That is the "present value" of that single $400 payment in the future. This seems relatively simple. However, if that $400 is one of a series of payments over the lifetime of an injured worker, then we also need to know the investment necessary for the $400 payment before that, and the one after that, and all the others.
Added to this complexity, the amount due for permanent total disability to a Florida injured worker is not constant ($400), but will usually increase annually. That increase is not always true, because certain maximum constraints might apply. Those constraints for this year are known, but can only be guessed at for tomorrow and for October 2045.
Therefore, as a young lawyer, I was faced with a series of challenging calculations, and there was great debate in the practice as to the appropriate assumptions (interest rate, life expectancy, and maximum rates) that had to be agreed upon. Then, various repetitive calculations would lead you to a present value of the benefit stream.
Under the statute, in some years, the interest rate was stated, 4% or 8%. Obviously, if you can earn 8% on the money, less is needed than if you can earn only 4%. For that example of $400 on October 1, 2045, according to Claude.ai, that difference means either
At 4% interest: PV = $400 / (1.04)^20.06 = $400 / 2.208 = $181.16At 6% interest: PV = $400 / (1.06)^20.06 = $400 / 3.281 = $121.91At 8% interest: PV = $400 / (1.08)^20.06 = $400 / 4.953 = $80.76
The reader will likely be asking what this has to do with artificial intelligence, but the foregoing quotes are illustrative. AI calculated those values almost instantly. Because they are relatively simple (one payment, fixed value, and known term), we could do so long-hand.
As an aside, some will find it curious that a statute would mandate the rate used in such a calculation. That represented a legislative conclusion that such rates were available to the injured worker so that they could invest their settlement and earn that return over time, replacing the actual value that would have been paid in periodic benefits, had the case not been settled.
Suffice it to say that calculating and predicting such a present value was tedious. It was slightly less so with a handy program like Lotus 123, Quattro Pro, or Excel. The spreadsheet made replication of discrete calculations far simpler and faster. Keep in mind that those tools were not intuitive; we spent many hours learning to use them.
But then came the entrepreneurs. By the early 1990s, several of them were marketing calculator programs (think of it like an "app"). Those programs did the math for you. There were inputs for known facts such as the worker's average weekly wage, the date of accident, the birth date (for life expectancy), and the interest rate. Those plug-ins resulted in a straightforward prediction of present value.
Unfortunately, each of the programs tended to yield different results from identical input data. Sometimes those distinctions were markedly different, and consistently so. One program tended to predict a higher value, another usually a lower. There were then arguments as to what assumptions were input, and what assumptions the program made about other variables, such as the predicted maximum compensation rate in future years that might limit the overall value stream.
The fact that programs yielded different outcomes sewed the seeds of distrust. Despite having a program at my disposal, I still spent hours manually calculating the present value of future benefit streams in order to advise clients and strategize about reserves and appropriate settlement values.
We tried the various calculator programs. We test-drove them, compared results, and did the manual work as an acid test. In time, we gained comfort. Not so much in the absolute inviolability of a particular program's output, but in our faith that this or that program produced results that were or were not reasonably valid based on our acid test manual calculations.
We came in time to believe one program or another was the most likely to be close to the correct number. This did not alleviate the arguments. The attorneys who represented injured workers tended to prefer the output that was predictably highest, and many defense attorneys preferred the lowest output. Others were more academically focused, but everyone had a preference.
The point of this post about artificial intelligence is that we all became more efficient. For a period, when those programs first appeared, we did both manual and computer calculations (with spreadsheets). Then we did both manual and spreadsheet and automated program calculations. The amount of time invested actually increased initially.
The increase was driven by our persistent engagement of the "old way," and the additional time for the spreadsheet or then the automated program. But, in a short time, we came to have faith in the automated. Then we stopped doing the spreadsheets. Eventually, we stopped doing the manual calculation.
Eventually, we became accepting of the output of one of the automated programs. Then, we were producing a present value in five minutes that had taken us a full day of work and patience.
We were more efficient. We were faster. And we delivered a better service to our clients at a greater value. Some attorneys still billed their clients for 8 hours of hard math and assumptions, despite working only 10 minutes. Others of us, instead, were honest and billed the 10 minutes. That is a discussion for another day.
The point here, however, is that we "tried and then we trusted." The technology initially cost us efficiency. We invested 10 hours in that eight-hour task, and the client was used to paying for eight. They did not understand why the cost would increase, and therefore, we tended to eat the extra time and bill the eight hours.
But as our proficiency grew, our confidence grew, and we gained faith in the technology. We became more efficient, with more hours in our day, because we could do eight hours' work in 10 minutes. The tech eventually resulted in efficiency, but first, we had to try before we could trust.
The example is only one, but it illustrates the challenge of artificial intelligence (AI). We may reach a day when blind reliance is appropriate for many AI outputs. But it is not this day. Today, we verify every AI output (does the picture it created have people with six fingers? Is the case law it cites real?).
In the short run, there will be increased work caused by AI. Over time, we may come to be more reliant, trustful, and confident. But that is not today. We can fear it and ignore it. But if you do, know that others in your profession are engaging it, leveraging it, and learning when and how to trust it.
The tool is new and novel. The challenge of evolving technology is not in the least. With all technology, there will be doubt, verification, and trust cycles. For more on cycles, see Is Gartner Helpful on AI? (December 2024). You must make decisions about your faith, trust, and process. You must grow with or without it and persist in a world that is constantly changing. Get used to that.
Know that there will be few, if any, perfect solutions. Know that tomorrow's innovation will always surpass today's. Know that this evolutionary process is cyclical and scary, and inevitable. And then, get back to work doing your best to adapt and verify, and adopt. It is a challenge, but that is life.
