When AI Refuses to Help With Sensitive Legal Records
A legal professional we interviewed described a straightforward research task: upload a set of judicial opinions, then ask AI to identify the convictions and sentences discussed in them.
The questions concerned facts already recorded in those decisions. But according to the interviewee, Claude declined to answer on content-policy grounds.
The interviewee understood why the subject matter might raise concerns. That did not remove the practical problem: they still needed to review the material.
Later in the conversation, they described a different experience with Scoopie.
“I haven't run into that problem.”
That observation helps explain one aspect of Scoopie’s value in their work. It also needs context. The interviewee was using Scoopie to review witness transcripts, not the judicial opinions involved in the reported refusal. They had not tested the same documents and questions in both tools.
This is an account of one professional’s experience, not a controlled comparison. Names, affiliations, locations, and identifying case details have been omitted.
The material was sensitive. The task was factual review.
The interviewee was not asking AI to invent a scenario or produce harmful content. They were trying to extract information from existing legal documents.
That distinction matters in legal document review. Sensitive material can be part of the record a professional needs to understand. The subject matter does not disappear because an AI assistant declines to discuss it.
In this instance, the requested assistance stopped before the user could assess an answer, check its sources, or decide whether it was useful.
Unnecessary refusals are a challenge AI developers have acknowledged. In its 2024 Claude 3 announcement, Anthropic described efforts to reduce refusals of harmless requests through better contextual understanding. That historical announcement does not explain this particular incident, but it shows that distinguishing a legitimate request from a harmful one is a recognized problem. Source: Anthropic
This does not mean Claude cannot support legal work. Anthropic also documents its use in legal applications such as Harvey. The narrower point is that this interviewee encountered a refusal during a particular research task. Source: Harvey customer story
For someone evaluating an AI assistant, that introduces a practical question alongside accuracy and speed: will it help with the material the job actually requires?
In Scoopie, the interviewee was working through witness testimony
The interviewee’s Scoopie workflow focused on understanding testimony.
They prepared text versions of transcript PDFs, checked the converted text against the originals, and connected those files to Scoopie. From there, they asked questions about individual witnesses and events described in the record.
The questions were closely tied to the documents:
Who was this witness, and how were they involved?
What did this witness say?
What did several witnesses say about a particular event?
These were not requests for a general explanation of the law. They were questions about what specific people had said in specific records.
The interviewee explained that a summary helped them get oriented before reading the transcript closely:
“Having that summary before I read the transcript makes it a lot easier to know what I'm reading.”
That is a modest but useful role for AI. The summary did not replace reading. It gave the reader an initial understanding of who was involved and what to look for.
When discussing content refusals, the interviewee explicitly said they had not encountered that problem in this Scoopie workflow. Being able to carry out those questions was therefore a meaningful part of their experience.
It does not establish that Scoopie will answer every request involving sensitive material. It establishes what this user reported being able to do.
The benefit was being able to continue the review
For this professional, a useful answer was the start of further work.
After getting an overview, they could ask about a particular event or examine how a statement lined up with the testimony. They described asking Scoopie to check a statement against selected witnesses’ accounts, including whether it was supported, missing context, or inconsistent with the testimony.
The value was in having assistance with that first pass through the material.
A refusal provides no summary to examine and no supporting passage to check. In the Scoopie workflow the interviewee described, they could move from a question to an answer, then back to the record.
Scoopie’s source references supported that process. The interviewee described clicking a citation to see where the text supported a claim. Scoopie’s official product description likewise explains that users can follow answer sources into a document viewer to inspect highlighted passages. Source: Scoopie
The central benefit here was not a promise that the AI’s interpretation was correct. It was that the interviewee could get assistance with the material and then inspect the basis for that assistance.
An answer still needed to be checked
The interviewee was clear about where their own responsibility began.
“Before I file anything, I'm going to manually verify it against the PDF.”
That detail is important. Their workflow used text files prepared from transcript PDFs. They checked the conversion before using the files and returned to the original PDF when verifying material for a filing.
An AI answer was not treated as the official record. Nor was a source reference treated as proof that every interpretation was sound.
The professional still needed to assess what the witness actually said, whether surrounding testimony changed the meaning, and whether the answer accurately represented the record.
This is what made the reported experience useful: AI assistance was available for the preliminary review, while the final assessment remained with the person doing the work.
A practical reason to keep using Scoopie
This interview does not establish a universal winner among legal AI tools. It reveals a concrete reason one professional found Scoopie useful.
They had experienced a content-policy refusal while researching judicial opinions in another tool. In their separate Scoopie workflow, they reported being able to ask questions about sensitive witness transcripts without encountering that problem.
That allowed them to get oriented, investigate particular statements, and check the supporting text before returning to the original record.
If you are evaluating Scoopie for document review, start with material you are authorized to use and a question whose answer you can check. Assess whether the response addresses the task, represents the document fairly, and gives you useful evidence to inspect. Review your organization’s data-handling requirements before connecting confidential records.
The aim is not simply to find an AI that will answer. It is to find assistance that helps you do the work, with enough visibility to decide whether the answer deserves to be used.
Sources
Anonymized product interview and companion transcript supplied for this article. Identifying details have been withheld. The reported refusal was not independently reproduced, and the model version and settings were not established.
Introducing the next generation of Claude, Anthropic.
Harvey transforms legal work with Claude, Anthropic.