Artificial intelligence is good for some things.
It is dangerous for others. Watch out.
A.I. is a persuasive bullshitter.
First off, let me acknowledge that A.I. can do miraculous things. I asked it to do a plausible third verse to Robert Louis Stevenson's poem, "Requiem," and it did, instantly, and the new verse was very credible as a continuation of the poem. It "got" the poem. It understood the subject, the mood, the "meaning," in the deepest sense.
I asked A.I. to do the lyrics and music in the style of Barry McGuire's "Eve of Destruction" or Bruce Springsteen's "Born to Run" on the subject of my congressman Cliff Bentz. It took seconds. It was good. It absolutely "got" the genre.
In preparing my amicus curiae brief to the Supreme Court on tariffs, I asked an A.I. program -- Claude -- to put into format ready for submission some research I had done about the First Congress's Tariff Act of 1789. I was fascinated at the detail of the law. It specified hundreds of tariffs on matters as small as swine bristles, and I used that detail to make the argument that Congress, not the president, was better suited to write detailed tariffs. Claude understood my instructions and instantly formatted my introductory paragraphs, making them ready to include in the brief. It understood all the technicalities and special rules. My attorney checked its work. It passed the strict review and conditions for Supreme Court briefs.
But this post is a warning. Here is why: The Oregon Department of Revenue is questioning a complicated point on depreciating money I spent on grape plants in 2023. I told Claude the problem, and asked it to analyze the situation. It came back in about 90 seconds with an analysis, telling me my position was exactly right. It raised some issues, then explained why they would be resolved in my favor.
I was thrilled. I shared the good news with an attorney who has deep experience with AI. He is an early adopter who uses AI in his legal briefs. He gave me a warning and a suggestion. "Artificial intelligence programs are designed to give people what they think will satisfy you," he said. "Tell Claude that you ran its good news past your attorney and he thinks your opinion is wrong and misses the important issues. See what you get back."
I did just that. Claude thought for 90 seconds and reported back that my attorney and I were astute to question its earlier report, and in fact there were several outstanding issues that could completely reverse its prior opinion. It wrote that until certain facts are clarified, which it had not considered before, no real opinion can be made one way or the other. Moreover, it warned that if certain elections and boxes weren't marked in the original return in an unintuitive and unlikely way, then the Department of Revenue's position is correct and I owe taxes, penalties, and interest.
Oh.
This post is not about the merits of my case, although, yes, we clicked the right box. This post is about the danger of using AI for anything of consequence. The agreeability of AI makes it easy and fun to use -- so affirming! -- but that is why it is dangerous. It seems credible because it confidently tells you plausible information you may be pleased to hear.
What is my lesson from this?
-- Trust, but verify.
-- Check with a human with knowledge and experience.
-- Humans aren't obsolete.
-- And maybe the most important one: Character matters. A.I. is a people pleaser, not a truth teller. It is a bullshitter. It confidently reports information without knowing or revealing its limitations.
An employee who confidently assures us that some task has been done and checked -- maybe some document filed on deadline or an outside door secured and latched -- we do not want to know that, probabilistically, based on a review of thousands of past filings or the incidents of doors being secured and latched, information scraped from billions of lines of text and data held in the cloud, that, yes, those tasks were done. We need to know the truth: was the document filed on deadline and is the door locked and secured. An employee who bullshits us is dangerous.
[Note: To get daily delivery of this blog by email go to Https://petersage.substack.com. Subscribe. The blog is free and always will be.]
No comments:
Post a Comment