Your AI Read the Documents. Did You?

Stanley Druckenmiller says he writes everything with AI now. The debate that followed was about op-eds. The harder version of it is about analytical work - and it's the question we're answering in the design of the Covenant Workbench.

"Of course I used AI." That was Stanley Druckenmiller's answer when NOTUS asked about his Wall Street Journal op-ed "Let the Bond Market Speak," published August 24th, which questions Treasury Secretary Scott Bessent's interventions to hold down Treasury yields. "I'm not embarrassed by it," he added. "I write everything using AI now for the same reason I use a calculator when I do math problems." He denied the piece was AI-written from end to end, said he rejected many of the AI's suggestions and insisted the ideas and judgment were his own.

The Journal's editorial page editor, Paul Gigot, defended publication: AI is a "fact of modern life," and "nobody can doubt that his op-ed is his genuine opinion." Critics pushed back on disclosure, on authorship and on whether an unknown contributor would have been extended the same latitude.

To me the interesting question is the one underneath this debate, because it's the same question every credit professional now faces in their own work. When we use AI in our workflows, do we relinquish some of our humanity? Or are we simply adopting another tool, with efficiency gains akin to the calculator's?

AI isn't inherently dehumanizing, and the calculator comparison undersells what these tools do. A calculator answers a question you've already framed, by a method you chose, with a result you can verify by inspection. It never proposes the question, never argues and never sounds confident about something it hasn't checked. Generative AI does all three, which is why it's more useful than a calculator and why the analogy fails as a defense. The tool's effect on our thinking depends on the workflow it enters and the authority we give it there.

So maybe the better question is this: Are our brains "quietly atrophying" (as AI would write) while we bang out ever more work?

It's something I think about a lot in my own workflow, which tends to revolve around four general themes:

• There's analysis: establishing what the material says and permits.
• There's contemplation: weighing what it means, how provisions interact, what follows economically.
• There's articulation: expressing the conclusion clearly.
• There's judgment: deciding what to do and standing behind the decision.

AI can assist each of these, but not in the same way and not with the same risks. Druckenmiller, on his own account, used it mainly for articulation while keeping the framing and the judgment for himself. The trouble starts when we treat the four as interchangeable and hand over the whole shebang. The more important question is how we can benefit from AI without outsourcing the very thinking, judgment and engagement that make our contributions human.

Credit documentation is where I test that question: too much material to read closely and too much at stake to read carelessly. A high yield indenture or a private credit facility agreement runs to hundreds of pages, and the risk is rarely in one provision. Restricted Payments capacity compounds with reclassification rights; a definition amended in one section changes the effect of a covenant three sections away. The provisions that determine whether value can move away from creditors only reveal themselves when read together - and the people responsible for that reading are monitoring dozens or hundreds of positions at once.

That's a volume problem, which makes the field a natural application for AI, and a judgment problem, which makes uncritical AI output dangerous. A confident, polished summary of a covenant package reads the same whether it's right or wrong; if it's wrong, you find out at the worst possible moment - in a liability management exercise, when someone else read the documents more carefully than you did. What happens to our ability to think deeply if polished, confident answers are presented at the click of a button, before we've wrestled with the questions ourselves? A polished answer is not necessarily an understood conclusion.

That tension is the design problem at the center of the Covenant Workbench, the platform we're building at Fox Legal Training. The starting point is the part of the job AI genuinely should compress: finding, extracting and organizing provisions, so the analyst's time moves from locating language toward considering its interactions and economic consequences.

Two design choices follow. First, the review is intended to run on a structured covenant taxonomy rather than open-ended prompting alone. The taxonomy encodes what an experienced covenant reviewer looks for - the categories, the interactions, the places where flexibility tends to be added - so the machine's search follows expert framing instead of improvising its own. That's the difference between AI expressing expertise and AI simulating it. Second, we're designing a dedicated liability management exercise analysis capability, connected to the LME case studies I've built up over the past several years, so users can ask the two questions investors ask after every exercise: how did they do it, and could it happen to me? None of this is meant to tell you whether an LME will happen - that's a commercial decision made by people you haven't met. What it's designed to surface is the permissions, structures, precedents and interactions that would make one possible, for a human to weigh.

The tool is only half of it. Fox Legal Training exists to train credit professionals in covenant analysis, and the Workbench runs on the same framework we teach. That pairing is deliberate. A tool used expertly needs users trained in the analysis it accelerates - people who know what the taxonomy is looking for, why a provision matters and when an output deserves to be challenged. We're building the Workbench and training the people who will use it on the same analysis, so the expertise and the tool reinforce each other rather than the tool standing in for the human expertise.

There's a generational problem in all of this. Druckenmiller can credibly say the judgment was his because he formed his views over five decades without the tool. The analyst who is twenty-six today won't have that biography. If the machine answers before she has wrestled with the question, when does she develop the instincts to supervise the machine? The answer isn't withholding the tools - that produces slower juniors, not wiser ones. Firms must build foundational competence deliberately now that the workflow no longer builds it, which is exactly why the training and the tool must go together. A system that shows its sources and invites challenge makes the documents more legible to the person learning; one that hands down conclusions without transparency makes them invisible.

So where does assistance end and judgment begin? I've stopped looking for the line in the technology, because it isn't there. It's in the workflow, which means we get to draw it. Drawn badly, AI in credit analysis produces confident output, hollow understanding and accountability dissolved into the machine. Drawn well, it removes the part of the job that was drawing on the best of the human - hunting through page 214 for a definition - and returns that time to the part that was: analysis on what these provisions permit, how they interact and who is prepared to stand behind the answer. 

What stays irreducibly human isn't a task. It's ownership: framing the question, testing the output, resolving the uncertainty and making the final judgment. Druckenmiller says the ideas were his. The challenge for the rest of us is to build workflows - and train the people in them - so that claim stays true, and so we can prove it.