Using AI to Widen Perspective Without Borrowing Judgment
BY NICHOLAS PROY, ESQ.
A DEMAND LETTER can sound firm when you draft it, but unnecessarily heated when someone else reads it. A large language model (LLM) can help catch that, sort of. The danger is that this same tool can go one step further and offer a polished interpretation of what the letter means and what response makes sense.
Much of the lawyer-facing discussion about artificial intelligence (AI) still focuses on visible failures: fake citations (aka: citation laundering), inaccurate summaries, invented quotes, and confidentiality problems. These risks are real. But there’s a quieter risk that doesn’t always announce itself as an error. Sometimes the problem isn’t that the system gives a false fact. Sometimes the problem is
that the LLM offers a polished interpretation that is really a professional judgment call.
That distinction matters. While many articles on AI focus on accuracy, namely checking whether the LLM got the facts right, and how to avoid citation laundering. This one is about agency: making sure the lawyer remains the one exercising judgment, especially when the LLM sounds calm, balanced, and credible enough to make borrowed judgment feel like the lawyer’s own.
AI can help broaden perspective by testing tone, offering alternative readings, identifying missing questions, or flagging language that may land harder than intended. That’s very different from letting AI decide what a communication means, how serious a
AI can help broaden perspective by testing tone,
offering alternative readings, identifying missing questions, or flagging language that may land harder than intended.
That’s very different from letting AI decide what a communication means, how serious a threat
is, or what strategy a matter requires.
threat is, or what strategy a matter requires. Those are matters of professional judgment, and the lawyer remains responsible for them.
The line is simple: when AI gives you facts, verification of those facts is the work. When AI gives you perspective, judgment is the work. The practical question is how to let AI widen your perspective without quietly or unknowingly outsourcing your professional judgment.
Why This Matters In Ordinary Practice
In my own solo practice in Carroll County, I use AI for drafting, issue spotting, summarizing, tone checks, and second looks at communications before they go out. Those uses can feel low-risk because they are often framed as light editing or brainstorming rather than research or formal legal analysis.
But low-risk isn’t the same as no judgment involved. An LLM may avoid hallucinating cases and still frame how someone reads a letter, which interpretation feels natural, or which response posture seems sensible.
We routinely operate in settings where tone, motive, seriousness, and response posture matter. A demand letter may be posturing, or it may be the final step before suit. A client email may be ordinary frustration, or it may reveal the start of a larger problem. An opposing counsel communication may be routine positioning or preparation of a record for later use. In those situations, our work isn’t simply to process words, but instead to interpret them in context and decide what response best serves the client’s interests.
The risk starts when a tone check becomes a judgment substitute.
In one anonymized, non-confidential exercise, I used two different AI systems to review the same demand letter text from a recipient’s perspective. I was simply looking for perspective: how the letter might read to another person, whether it sounded unnecessarily heated, and whether certain phrasing was likely to escalate the situation. The point was not to compare products or conduct an empirical test. The point was that the same text produced materially different readings.
One LLM’s output framed the litigation threat as mere posturing, with no serious intent to sue. The other treated the same letter as a dramatic warning sign that a suit was imminent. It seems my simple tone check now produced readings at opposite ends of the same spectrum.
Then I realized that the LLMs were no longer merely helping me see tone from a recipient’s perspective. Their outputs were inferences about motive, seriousness, and strategic posture, all from the exact same text. In other words, they were doing something that looked increasingly like professional judgment: assessing motive, seriousness, and likely next steps. And because the outputs were confident and fluent, a recipient who uploads a communication to an LLM with a prompt like “analyze this and let me know your thoughts” may receive a credible-sounding interpretation generated from incomplete context.
That’s the line to watch. The safer question is, “How might this read?” The riskier question is, “What does this mean, and what should I do?” For my test, I did what many recipients
It’s tempting to use AI as a sounding board.
Asking whether language may read as more accusatory, sarcastic, or escalatory than intended can be a reasonable use of the tool, depending on the tool, the information entered, and the lawyer’s review of the output.
A Practical Example: When Tone-Checking Turns Into Judgment
The distinction became clearer to me in a simple use case: testing the tone of a demand letter before sending it.
Many lawyers have drafted a letter or email while irritated, or while trying to strike the difficult balance between firmness and restraint. Sometimes the draft says exactly what needs to be said. Sometimes it goes further than it should. It’s tempting to use AI as a sounding board. Asking whether language may read as more accusatory, sarcastic, or escalatory than intended can be a reasonable use of the tool, depending on the tool, the information entered, and the lawyer’s review of the output.
might realistically do: upload the letter to ChatGPT with an “I received this, what does it mean?” prompt. And then the two inconsistent responses presented themselves.
Perspective Is Useful
An LLM can surface alternate readings of a communication, especially when a draft has been read so many times that the lawyer no longer hears it like a fresh reader. It can flag language that sounds more emotional, accusatory, dismissive, or escalatory than intended. It can also point out missing questions when a draft is too focused on rebuttal and not focused enough on what still needs to be learned.
In a profession where wording matters, a quick second look at how a message might land may save time and reduce avoidable friction.
In a profession where wording matters, a quick second look at how a message might land may save time and reduce avoidable friction.
When AI Crosses The Line
The crossing point isn’t always obvious, because it often happens gradually. It starts by asking whether a draft sounds too harsh. The AI responds not only with a tone assessment but also with interpretations of motive, likely next steps, hidden leverage, or the credibility of a threatened action. We then begin to absorb not just language feedback but also strategic framing.
The danger begins when the tool starts:
telling the user whether a threat is serious or merely posturing; characterizing a sender as bluffing, desperate, angry, or likely to escalate;
implying what strategic response is most appropriate; making one posture feel obviously wiser than another; presenting a cleaned-up, persuasive interpretation that pushes aside ambiguity; or validating the user’s emotional reaction rather than testing it.
The issue here isn’t that the tool is necessarily wrong. The issue is that even when the output sounds credible, the lawyer may start to borrow the system’s judgment rather than exercise their own independent professional judgment.
That’s a problem because professional judgment isn’t just pattern recognition, because it depends on context that the model may not have, or may not reliably weigh. Consider such factors as: the history between the parties, the client’s goals, the cost of escalation, the
surrounding record, the forum(s), the assertions and levels of damages, and the personalities of those involved. It also depends on what has already been said off the page and how the communication may look later if it’s attached to a motion, affidavit, complaint, or grievance.
A tool that sounds calm and reasonable can still be wrong about what matters.
Why This Is Easy For Lawyers
To Miss We’re trained to analyze language, assess risk, and identify how communications may be perceived later. That’s precisely why an LLM can be deceptively persuasive in legal work. It produces the things that can look like judgment on the page: organized reasoning, balanced phrasing, clear conclusions, and apparent confidence.
The risk may be easiest to miss when lawyers use AI in seemingly modest moments: “How does this sound?” “Am I overreacting?” “Is this too aggressive?” “What might the other side do with this?” Those questions feel safer than asking for legal research or a final filing. But they can still implicate core professional judgment. If the system makes a communication sound harmless, the lawyer may become less careful. If it makes a threat sound imminent, the lawyer may become more defensive than the situation requires. If it validates the lawyer’s tone, the lawyer may stop asking whether the draft is strategically wise.
The danger isn’t only hallucination. I think of it as cognitive steering: the output can shape what the lawyer treats as important before the lawyer has made an independent judgment.
This is a Bias Problem
This is a bias problem, but not in the usual sense. The practical concern is that the LLM can shape the lawyer’s starting viewpoint, emphasis, and posture before the lawyer has even finished thinking.
When lawyers hear bias, they often think of political, demographic, or dataset bias. Those are real concerns, but they are not the only types of bias that matter here. The more immediate problem for us is judgment bias: the risk that a polished output will anchor the lawyer’s interpretation before the lawyer has independently assessed the communication in context.
I think about that risk in several practical ways:
Anchoring. The first interpretation offered by the system may become the reference point against which later thinking is measured.
Confidence transfer. Fluent prose can make the conclusion feel stronger than the underlying support warrants.
Selection effects. The system may foreground certain considerations and omit others, leading the lawyer to treat what is mentioned as what matters.
False reassurance. If a response sounds balanced and professional, the lawyer may mistake tone for reliability.
Borrowed posture. The system may subtly nudge the lawyer toward a more aggressive, passive, dismissive, or fearful stance than the lawyer would otherwise adopt.
These aren’t exotic problems. They are familiar human problems sped up by a machine that’s very good at generating credible-sounding language.
A Practical Workflow For Preserving Judgment
The answer isn’t avoidance, it’s disciplined use. For me, that means several practical habits.
If multiple AI tools are used, their disagreements should not be treated as a puzzle to be solved by taking a vote. Their agreement should not be treated as objective confirmation either. Multiple outputs may widen perspective, but they still leave the judgment with the lawyer.
How This Connects To Ordinary Law Practice
This comes up in ordinary law practice more often than it first appears: tone-checking a difficult email to opposing counsel; reviewing a client’s accusatory message and deciding whether it reflects panic or a genuine issue; reading a settlement communication and deciding whether it signals openness or hardening; or deciding whether a follow-up should be firm, quiet, immediate, delayed, or not sent at all.
First, define the task before using the tool. Am I asking for perspective, or am I asking for a decision? If the real question is “What does this mean and what should I do?” AI may still help surface considerations, but the answer has to come from the lawyer’s judgment.
Second, use AI to surface possibilities, not conclusions. A helpful use is to ask how a communication might be perceived, which parts sound escalatory, or what alternative readings exist. A riskier use is asking whether the sender is bluffing, whether the threat is credible, or what strategic posture I should adopt.
Third, treat the output as a mirror, not a decider. A mirror can show you something about how your language may land. It cannot tell you what professional response is best.
Fourth, return to the actual sources of legal judgment: the file, the facts, the law, the history of the matter, the client’s goals, and the consequences of being wrong.
Fifth, be especially careful when the tool appears to validate your first reaction. That’s often the moment when the need for independent judgment is greatest, not smallest.
If multiple AI tools are used, their disagreements should not be treated as a puzzle to be solved by taking a vote. Their agreement should not be treated as objective confirmation either. Multiple outputs may widen perspective, but they still leave the judgment with the lawyer.
It appears anywhere the lawyer is tempted to let the tool collapse ambiguity into a clean, confident story. A quick answer feels efficient. But efficiency isn’t the same thing as sound judgment, and a polished interpretation isn’t the same thing as a strategically wise one.
The Professional Responsibility Piece
The broader, general professional-responsibility point isn’t complicated: a lawyer’s duties do not disappear because the tool sounds thoughtful. For Maryland lawyers, the relevant starting point remains our existing professional responsibility obligations. Depending on the tool, the task, the information shared, and the governing rules, the use of AI may implicate duties of competence, diligence, confidentiality, communication, candor, fees, and supervision.
That’s why this problem deserves separate attention from the now-familiar stories about hallucinated citations. False cases are easier to identify as errors once discovered. Distorted judgment is harder because it often arrives in the form of something polished and measured.
The risk isn’t only that AI may be wrong. The risk is that AI may influence what the lawyer treats as important, credible, threatening, or worth verifying in the first place.
Bottom Line: Perspective Is Useful, Judgment Is The Work We can use AI responsibly by testing tone, exploring alternative readings, organizing possibilities, and identifying blind spots. These are real uses, and they are valuable. But we should be careful not to let AI perform the harder work of judgment. The system shouldn’t be treated as the
decision-maker on what a threat means, whether a sender is bluffing, how seriously to take a communication, or what strategic response best serves the client. Those are not merely drafting tasks. They are professional judgments. The same lesson applies beyond verifying facts. Used carefully, AI may help a lawyer see more clearly by testing tone, surfacing alternate readings, and identifying blind spots. Used carelessly, it can make borrowed judgment feel like the lawyer’s own. That is the line lawyers have to guard.
We
can use AI responsibly by testing tone, exploring alternative readings, organizing possibilities, and identifying blind spots. These are real uses, and they are valuable.
Nicholas B. Proy earned his B.A. in Intelligence Studies from Mercyhurst University and his J.D. from the University of Maryland Francis King Carey School of Law. He also holds CompTIA A+, Network+, Security+, and Server+ certifications. His writing has appeared in the Maryland Bar Journal and 2600: The Hacker Quarterly, with a focus on practical technology, privacy, and AI-verification issues affecting lawyers and other professional users.
