Artificial Intelligence Standards and Use

Law firms tend to handle this subject by saying nothing at all or by offering a paragraph of reassurance that commits to nothing in particular, and neither approach gives you what you need in order to decide whether to send me your file. What follows is a description of the tools I actually use, what happens to your matter when I use them, and the limits I have set on myself. I have written it so that you can check it against what the vendors publish, because a policy you cannot verify is a slogan.

The knowledge bank

The firm keeps a database of its files offline, converted into markdown, which is a plain text format that computer programs read cleanly. That database serves as the single source of truth for client matters, so that when I need to know what happened in your case I can query the bank instead of re-reading the entire file, which is faster and does not depend on my recollection of a matter I last touched in March.

What that database does is retrieval. It is searched, the way you would search a filing cabinet that answers questions. No model is taught anything from it, and nothing about your matter is absorbed into any model's weights, which is a distinction that gets lost constantly and is worth stating plainly.

What happens when a commercial model is queried

Queries against the knowledge bank run through Anthropic's Claude on the commercial API using API authentication rather than through a consumer account, and that distinction carries real consequences for your file. Under Anthropic's commercial terms, inputs and outputs sent through the API are not used to train models by default, so there is no opt-out to perform because training was never the arrangement. Anthropic deletes API inputs and outputs from its systems within thirty days of receipt, which is the published retention period for commercial traffic.

There are two exceptions you should know about rather than discover later. Content that Anthropic's automated trust and safety systems flag as violating its usage policy may be retained for up to two years, with the classification scores retained for up to seven, and Anthropic may retain anything it is required by law to keep. I clear the conversation history after each query and keep that history on machines the firm controls, which removes it from the interface, though it does not accelerate deletion on Anthropic's side, which runs on the schedule described above.

Sensitive matters never leave the office

Some matters involve personally identifying information, trade secrets, or facts that I am simply unwilling to route through anyone else's servers regardless of what their terms say. For those, the firm runs open weight models on its own hardware, meaning the model file sits on a machine in this office and the analysis happens there. No API call is made, no vendor is involved, and nothing crosses the network, so your case data never leaves the lawyer's laptop and is adjusted entirely offline.

At the time of publication the firm runs Qwen 3.6-37B and Gemma 4.1 for this purpose, on separate machines. When those models change, this page changes with them.

Where models come from

No model that faces a client originates outside the United States, and the fact that a system could be isolated from the network does not change that rule, because a rule that bends whenever the architecture makes it convenient is not a rule. That standard governs the answering service described below and anything else a client interacts with directly.

The REAL rating and how this firm applies it

Every page on this site publishes a REAL rating, which stands for Reported Engagement with AI Level. It is a self-reported disclosure standard maintained by Real Good AI, a nonprofit research organization, and it exists because asking whether someone used artificial intelligence produces a far less useful answer than asking how they used it. The firm endorses the standard and applies it to its own work as follows.

Mark	Level	What the standard means by it
	REAL 0 No AI	No intentional use of artificial intelligence of any kind, which reaches further than most people expect, since automatic transcription and automated background removal both count.
	REAL 1 Automation	Specialized tools that generate nothing, such as spell check, filters, captions, transcripts, and traditional machine learning models.
	REAL 2 Brainstorming	Artificial intelligence used for ideas, structure, or outlining, where none of its output is copied into the finished work.
	REAL 3 Moderate AI	Generated building blocks that reached the final work without changing its overall meaning. Any generated material that survives into the product places the work at this level or higher.
	REAL 4 Major AI	Generated major components or ideas, such that the work would have been different without the model, and a human making the same contribution would be credited as an author.
	REAL 5 Full AI	Generated the whole or the large part with minimal human intervention, or any product where output reaches the reader without human oversight.

Applying that scale to this website, the firm reports a 1 where artificial intelligence writes the site's functionality and security code, since that work produces machinery rather than anything you read. Design work sits at a 2. Where a model edits the firm's own writing for readability, sharpening sentences I drafted, the firm reports a 3. Where a model drafts the content and the firm edits it, the honest number is a 4, because at that point the model is the writer and I am the editor, and describing that as anything lower would defeat the purpose of publishing a rating in the first place.

Nothing on this site is a 5. Every page is read by an attorney before it publishes, so no generated text reaches you without a lawyer having gone through it. The firm also publishes no artificially generated illustrations, with the single exception of flow charts and diagrams, which are colored blocks and arrows rather than depictions of anything real.

Vera, the answering service

Calling outside business hours may put you through to Vera, an artificial voice that identifies itself as an answering machine at the start of the call. Vera exists to take appointment requests and to answer questions whose answers are already published on this website, which means she is deliberately built without the ability to take your confidences, and you should not offer her any. Anything that belongs in a privileged conversation belongs in a conversation with me.

Call transcripts live on a machine the firm controls and are cleared every thirty days, which I do myself rather than leaving to a scheduled job that nobody checks.

Who trains on your data

No one does. No vendor trains a model on your matter, and the firm does not train one either. What the firm does is index your file into a searchable local database so that it can be queried quickly, which is a genuinely different operation from training and which lives entirely on the firm's own computers. If that ever changes, you will read about it here first.

Questions

This page will be wrong eventually, because the tools move faster than anyone's policy documents. When it is wrong I would rather hear about it than have it sit here, so if something here does not match your experience of working with this firm, tell me and I will either fix the page or fix the practice.

Last reviewed August 2026.