Death by Dialogue: The Case for Killing the Legal Chatbot
Legal work needs density, not dialogue.
Ethan Mollick's recent piece on AI interfaces surfaces a finding that anyone who's used a chatbot for real work already suspects: the interface is working against you. Research on financial professionals using GPT-4o showed that the chatbot format itself created cognitive overload — walls of text, unsolicited tangents, sprawling discussions that compounded rather than clarified. The productivity gains from AI were partially eaten by the cost of parsing the AI's output. And the people hurt most were less experienced workers, exactly the ones who should benefit the most.
Mollick frames this as a general knowledge work problem. He's right. But legal has this problem in a uniquely compounded way, and nobody's talking about it.
Start with the stakes. When a financial analyst misses something buried in a chatbot's response, there's a correction cycle. When an attorney misses a clause conflict in a similar wall of text, there's exposure. The cognitive load tax in legal work goes beyond efficiency. It's a malpractice vector. Every moment an attorney spends parsing AI output format rather than evaluating AI output substance is a moment where something consequential can slip through.
Then add confidentiality. General-purpose AI tools can surface any relevant information. Legal AI has to surface the right information while proving it didn't cross a confidentiality boundary to get there. The interface needs to communicate both the answer and the provenance — where the reasoning came from, what data it didn't touch, why you can trust it. That's a design problem most legal AI vendors haven't even acknowledged, let alone solved.
Now consider how attorneys actually think. Redlines. Clause comparison tables. Risk matrices. Deal term sheets. These are cognitive frameworks built over years of practice. An attorney scanning a redline can absorb the delta between two contract versions in seconds because the format matches how their brain processes contractual change. Hand them the same information as chatbot prose and you've destroyed that efficiency. The format is doing work.
I've written before about how workflows beat autonomy in legal AI — that the best legal AI systems constrain AI behavior rather than maximize it, building governance and auditability into the workflow itself. The interface question is an extension of that argument. The interface IS the workflow. It determines what an attorney sees, in what order, with what context, and with what ability to act on it. Get the interface wrong and it doesn't matter how capable the model is underneath.
This connects to something I've been thinking about with decision traces. I've argued that the real institutional intelligence in legal organizations lives in the decisions those documents informed. How the partner structured that earnout. Why the analyst rejected that risk. What made someone deviate from the standard clause. Decision traces are where competitive advantage lives. But traces are only valuable if they're visible. If they exist as backend metadata that never surfaces to the person making the next decision, you've built an archive, not a tool. The interface has to make reasoning legible — showing which precedents informed a recommendation, which matters were considered, which confidentiality boundaries were maintained — in a format attorneys can evaluate in the flow of their actual work.
So what does a well-designed legal AI interface actually look like? Three principles:
First, familiar density. Attorneys don't need simpler interfaces. They need dense ones — the kind of information-rich surfaces they already know how to scan, powered by intelligence they couldn't generate manually. The mistake most legal AI tools make is either showing too much (chatbot output with every possible angle) or too little (single-answer oracles with no reasoning). The target is structured density: the right amount of information in a format attorneys already have muscle memory for. Think clause-level risk annotation overlaid on a document view, not a separate chat window telling you about risks in paragraph form.
Second, provenance as a visible element. Every AI-generated recommendation should carry evidence of what informed it and what didn't. Not buried in a tooltip or an audit log — visible in the primary interface. An attorney won't act on "similar deals suggest this clause structure" unless they can see which deals, confirm there's no conflict, and trace the reasoning. I've written about how decision traces need their own security model. They also need their own display model. The provenance layer is where trust gets built or lost, and right now most legal AI tools treat it as an afterthought.
Third, the interface as a governance surface. If you accept that workflows beat autonomy — that structured processes with human review points outperform unconstrained agents — then the interface is where governance becomes real. It should make clear what the AI accessed, what it didn't, where it's confident, where it's uncertain, and where human judgment is required. The UI itself becomes the control surface for legal AI governance. Not a separate dashboard. Not a compliance report generated after the fact. The governance information lives where the work happens, or it doesn't work at all.
Mollick points out that the only truly complete specialized AI interfaces exist for programmers — tools like Claude Code, Codex, Antigravity — because the AI labs are staffed by programmers building tools for themselves. Legal doesn't have that advantage. The people building legal AI tools are mostly engineers who've never sat in a deal room or negotiated an indemnification clause at 2 AM. The profession that invented the redline — arguably the most sophisticated document interface ever designed — is being asked to work through chat windows.
That gap is the opportunity. The companies that figure out what a redline-quality interface looks like for AI-powered legal reasoning will define the next phase of legal technology. The ones still shipping chatbots will keep wondering why their adoption numbers stall at the associates who are too junior to push back.
Source: Ethan Mollick, "Claude Dispatch and the Power of Interfaces" (One Useful Thing, March 31, 2026)
Ethan Mollick
