The biases we bring

There's a body of research on the biases judges bring to the bench. In one study involving 167 federal magistrate judges, researchers from Cornell and Vanderbilt universities found that anchoring, framing, and other bias influenced judicial decision-making in ways that could produce errors. A later study found that irrelevant numeric information also had an anchoring effect and skewed damages and sentences. And while there’s some evidence that judges can compensate for bias when they know about it and are motivated to correct for it, the research suggests that awareness in the abstract isn’t enough.

Could awareness in the moment of a decision make a difference? What if AI could help us see our biases precisely when they surface? What if we could test our inclinations in real time or pause on the assumptions we’re making before taking the next step?

If designed for this purpose, AI can place critical interruptions inside our workflows, reflect our cognitive process back to us, and augment us in new and important ways.

To explain how that's possible, and why it matters, I’ll start with a short reflection on judicial decisions.

Judgment is not one decision

We talk about the work of judges as though it happens in singular moments: A judge grants or denies the motion, admits or excludes the evidence, dismisses or permits a claim. Long before we decide anything, however, dozens of cognitive acts and tasks are strung together in preparation for the ultimate decision. Which filing do I read first? How do I organize the record? What’s the sequence of events? Which facts deserve the most attention? What are all the relevant authorities? Which authority is most on point?

None of these individual inflection points—or what I call micro-judgments—result in a ruling on their own. Yet every single one shapes it. This is because a judicial decision is not a destination, it's an architecture built from dozens of micro-judgments that accumulate over time. Each one scaffolds to another, and each one makes some conclusions easier to reach and others harder to see. By the time we sit down to decide, much of the deciding has already happened.

To be clear, some of the work inside this architecture is arguably computational, and AI may eventually do it better than we can. Other work should never be entrusted to an AI system and should remain firmly human. But my focus here is not on who or what performs each individual task, it’s on whether the technology helps us see the consequential judgments along the way.

None of these individual inflection points—or what I call micro-judgments—result in a ruling on their own. Yet every single one shapes it.

Today, with or without AI, many of the micro-judgments inside the architecture of judging are entirely invisible. Think about how a bench memo works. A law clerk reads the record, frames the issues, builds the chronology, and selects the authority. Accepting some, or all of it, involves many micro-judgments, but we don’t pause to examine every single one. The clerk's initial thinking dissolves into the memo, my initial reactions to the memo dissolve into my feedback and my own research.

The framing, the anchoring, the assumptions, often get lost in the mix. If you ask me to reconstruct every micro-judgment that shaped a ruling, I could give you an overview of my process, but I wouldn't be able to pinpoint every single micro-judgment. That’s because we can’t watch our own reasoning as it forms.

At least, until now.

The AI breakthrough

AI is the first technology in the history of judging that can systematically reflect a judge's own cognitive process back to the judge in real time.

We’ve always caught glimpses of our own leanings: a pointed question at oral argument, a tentative ruling, or a good clerk that pushes back. They all help us reflect on our own cognitive process. However, these glimpses are partial, or dependent on someone else’s choices about how they interact with us. AI is the first tool that can do this systematically and inside our own workflows.

AI can ask reflective questions, record our first instincts and hold them until the end, notice which facts we keep returning to and which arguments we never touch. It can compare what we emphasize against what the record reflects. It can be designed to hold up a mirror to our thinking.

But that’s the key—it needs to be designed that way.

Today’s tools primarily run in the other direction. Ask a system to analyze an issue, and it compresses the chronology, the framing, the research prioritization, and the outline into seconds, folding those intermediate steps into the background and presenting the user with a polished answer. Every micro-judgment gets made more quickly and more quietly than ever, whether by us or by the system. And when AI systems collapse our biases into their own biases, mixing and mashing without opportunity for interruption and correction, it’s dangerous. In my view, far more dangerous than hallucinations, a topic we can’t seem to get enough of these days.

In short, AI can bury the traces of judicial cognition deeper than any tool before; or, it can expose them. The design determines the path.

What exposure looks like: micro-judgment checkpoints

Judges can already use AI to expose some of these micro-judgments. For example, I use standing system instructions, deliberate prompts, and custom projects to direct the system into deliberate reflections and pauses that help me consciously shape what comes next. But these self-created checkpoints work imperfectly and inconsistently. Moreover, they depend entirely on a judge knowing how to set these up and remembering to make these demands of the system. It isn't scalable or sustainable.

Judicial decision-making is a complicated architecture that depends on the countless micro-judgments that shape a judge’s final decision.

The real progress would be in the design. And while some tools now offer usage recaps that summarize the topics we work on and the ways we tend to use the system, those summaries reflect general habits. They say nothing about my biases, assumptions, or leanings.

An AI system shouldn't reflect every micro-judgment, of course. If it narrates everything then the important checkpoints get buried in noise. Instead, it should pause on the micro-judgments most likely to change the direction of my decision-making, prompt me in a way that sparks genuine reflection, and give me the opportunity to take an informed next step.

Here are a few examples of how that might look in practice:

Issue framing — Before offering anything, the system asks me to frame the issue, then shows my framing next to the parties’. If mine tracks one brief almost word-for-word, is that framing right, or have I anchored?

Holding my first instinct — The system records my tentative lean after my first pass, then shows it to me later. If my analysis matches my instinct exactly, is that confidence, or untested bias?

Body of work reflection — The system identifies patterns from my prior decisions. Do I credit the same authority? Skew one direction too often? Reason inconsistently across similar cases?

Materiality and assumptions — The system flags the facts I seem to be assuming and whether the record supports them, before the assumptions harden and I press on.

For decades, the answer to bias has been training. We learn how our biases work in a conference room and then hope we retain enough to catch them at the right moments on the bench. Today, we have the opportunity to design systems around how bias works and force the systems to engage us at the precise moment at which our biases would otherwise go unchecked.

The risks

A system that flags gaps could push the judge outside the record and the parties’ arguments, if not properly restricted. Thus, it would be important for these checkpoints to stay bound to, for example, the parties' submissions.

Additionally, a system that stores my first instincts, leans, and patterns, is holding a record of my deliberative process. That record belongs in the same category as my notes and my draft opinions. It shouldn’t be stored as a dataset, mined for analytics, or used to build profiles for how judges decide. Judicial analytics built from our public rulings already make judges uneasy, and France went as far as banning them. This type of internal and iterative process would require even greater protection.

There’s also a question about the mirror itself. The system’s reflections are in essence micro-judgments too, and imperfect ones. Thus, checkpoints should be built as questions rather than conclusions, keeping the judge in control.

We also shouldn’t assume that seeing bias automatically fixes it. A poorly built checkpoint could harden a first instinct rather than test it, or it could give us the false sense that we’ve successfully interrupted a bias. Like anything else, the design should be validated and not simply presumed to work.

Finally, it’s likely these checkpoints will slow us down if not built correctly. A well-designed checkpoint should take seconds, arrive at natural pauses in the work, and never interrupt for the sake of interrupting.

A call to action

I suspect that what I’m contemplating here is not easy (or enticing) to build. Checkpoints add friction that metrics punish, and meaningful reflection requires systems that are designed around a body of work rather than a single session. For some platforms, that might mean rebuilding rather than simply adding a feature. Moreover, the reflections themselves must be grounded to avoid hallucinations and more bias.

Still, vendors build what buyers demand, and we should demand systems that make us better, not just faster. Public confidence in the American judicial system has fallen to record lows, dropping 24 percentage points in just four years. The causes are complicated and some are beyond any one judge’s control. But when confidence is this low, we should be looking for very opportunity to improve.  

A closing thought

Judicial decision-making is a complicated architecture that depends on the countless micro-judgments that shape a judge’s final decision. For all of judicial history, those micro-judgments have been largely invisible, even to the judges making them.

Today, we are building tools that could make that worse…or better.

Built and used carelessly, AI can bury the traces of our cognition deeper. Built and used correctly, however, AI can do something no tool has ever done before: show us our own minds at work, at the moments that matter most.

You can find more insights from Judge Braswell here