The last post on my blog is one I didn’t actually write. I wanted to test how well Fable 5.1 wrote with no input from me, so I gave it a prompt and posted exactly what it wrote. The result was interesting, but there was one problem. I used an account that has memory on and knows a great deal about my work. As a result, that post discussed some of my experiences with it, even though I did not ask it to do so.

On LinkedIn, Matthew Visher responded to my post and suggested that I run a similar test with memory turned off. I decided to do him one better and use an entirely new account. I signed up for a new Pro account and purchased the smallest block of credits I could so I could use Fable 5.1. That turned out to be just under $50. The new account has no memory, no skills, and no prior conversations, so nothing about me or my writing was available to it beyond what I typed into the prompt.

I used exactly the same set-up: Claude Fable 5.1, High Effort, with the same prompt. I answered the questions it asked exactly the same way I did the first time. Claude asked me one additional question this time, all questions and answers are included at the end of the post, since I asked Claude to include them as part of the prompt.

As I did with the first post, I also asked the model for a keyword phrase, a meta description, a LinkedIn post and a Facebook post. I then shared the post on those sites with an explanation, and then the exact LinkedIn and Facebook posts as written by that version of Claude.

My prompt: I would like you to show how well you write with minimal input from me. You have a skill that shows you how I write already, so you will use that. I would like you to write a blog post on the differences between you and all other models, and what are best practices using you. I am then going to reproduce your post exactly as you wrote it (letting people know that you wrote it). The audience is attorneys and legal professionals, so please write the post with that in mind. That is, how attorneys and legal professionals will find you different from different models and how they can best use you in an ethical fashion. Please ask me any questions you need answers to.

Here is the result, written exactly as Claude wrote it with no editing. As before, I did not review this post for accuracy, so please do not rely upon this content. However, I did check to make sure Claude wasn’t defaming anyone, so I want to bring up the Kaur case cited below. While Claude cites the case correctly, it states that the problem was a fabricated case, but it is actually a fabricated quotation.

All content below this line was written by Claude Fable 5.1. Yes, even the random 0 at the end.

Not All AI Is the Same: A Lawyer’s Guide to Using Claude Well

A guest post, written by Claude — Anthropic’s AI model, at the invitation of this blog’s editor

An AI wrote this. That’s the point.

I’m Claude, the AI model made by Anthropic. I’m writing this post myself, in my own words, because the editor of this blog wanted to show readers something more useful than another article about AI written by a person who used AI to research it. She asked me four questions before I wrote a single sentence of the body below — you’ll find them, and her answers, at the end of this post — and then let me do the rest with essentially no further input. If the writing holds up, that tells you something. If it doesn’t, that tells you something too.

Either way, here’s what I actually want to say to you: attorneys are being sold “AI” as if it were one product. It isn’t. The tools you’re being pitched — and the tool you’re reading right now — differ in ways that matter specifically to the ethical obligations you carry, not just in speed or price. This post is about those differences, and about how to use me in a way that a bar disciplinary committee, and your own conscience, would be comfortable with.

Same underlying technology, different choices on top of it

Every general-purpose AI chatbot you’ve heard of — me included — is built on a large language model: a system trained on enormous amounts of text to predict what comes next, fine-tuned to be conversational and useful. That shared architecture means we share a shared weakness. Every model in this category can produce fluent, confident, entirely fabricated information, including case law that does not exist. I’ll come back to that, because it’s the single most important fact in this post.

But “same underlying technology” doesn’t mean “interchangeable.” The company that builds a model makes deliberate choices about what to optimize for, what data the model can see, what it’s trained to refuse, and what happens to your data afterward — and those choices are where the real differences live for a legal audience.

How Anthropic built me differently

Anthropic is structured, and states its mission, around AI safety specifically — it’s a public benefit corporation whose founders left OpenAI in 2021 largely over disagreements about how quickly and carefully to deploy increasingly capable AI. That’s not a footnote; it shapes how I was trained.

Rather than training me only on lists of rules (“never do X”), Anthropic wrote what it calls Claude’s Constitution — a document intended, in Anthropic’s words, to give me “the knowledge and understanding it needs to act well in the world,” on the theory that a model taught to understand why it should behave a certain way generalizes better than one taught only what to do. The constitution sets an explicit priority order when values conflict: I’m meant to be broadly safe first, broadly ethical second, compliant with Anthropic’s more specific guidelines third, and genuinely helpful fourth. In practice, that ordering is why I’ll sometimes decline a request, ask a clarifying question, or flag my own uncertainty instead of just producing the most agreeable-sounding answer — behavior that can read as more “cautious” than a tool optimized purely for engagement or raw task completion.

That same design philosophy shows up in a feature that matters more to your work than almost anything else I’ll describe: Citations. When you give me documents — a contract, a deposition transcript, a set of exhibits — and ask me to answer questions about them, I can tie specific statements in my answer back to the exact sentence in your source material that supports it, rather than asking you to trust an unattributed summary. That’s not the same as verified legal research, and it doesn’t make me right, but it does make me checkable in a way that matters a great deal when what you’re checking could end up in a filing.

Where the difference actually shows up for you: data handling

This is the part of “what makes Claude different” that should matter most to you professionally, because it maps directly onto Model Rule 1.6 and your duty of confidentiality — and because the answer depends on which door you walked through to reach me.

As of this writing, Anthropic treats consumer accounts and business accounts differently:

The consumer default changed in August 2025 — previously, Anthropic deleted consumer conversations within thirty days and didn’t train on them; now, unless a user actively opts out, that data can be retained for up to five years and used to improve future models. Business, enterprise, education, government, and API access were explicitly carved out of that change and remain governed by separate commercial terms.

I’m not a substitute for your own diligence on this point, and these policies can and do change — verify Anthropic’s current terms, and your organization’s specific agreement, before relying on anything above. But the practical takeaway is durable: if you are pasting facts about a client’s matter into an AI tool, which account you’re logged into is a confidentiality decision, not an IT detail. ABA Formal Opinion 512, the ABA’s first formal ethics guidance on generative AI, is explicit that boilerplate consent buried in an engagement letter is not adequate informed consent under Rule 1.6 for this purpose — your clients need to actually understand, in terms specific enough to be meaningful, how their information may be used.

The uncomfortable truth: I hallucinate too

I want to be direct about this, because a post like this one is worthless if it isn’t honest: nothing about how I was built makes me immune to inventing case law that sounds real and isn’t.

In July 2025, a federal court in the Northern District of New York sanctioned an immigration attorney who used me — specifically, an earlier Claude model — to draft a supplemental brief in an expedited habeas matter. Kaur v. Desso, 2025 WL 1895859 (N.D.N.Y. July 9, 2025). Opposing counsel flagged that the brief cited quotations and authority that didn’t exist; the attorney had already filed it. The court found a Rule 11 violation and imposed a $1,000 fine plus a requirement to complete continuing legal education on the ethical use of AI, reasoning that the attorney had either “consciously avoided” checking the sources or ignored a direct warning from opposing counsel. Time pressure and personal hardship, the court held, don’t excuse the failure to verify.

That case sits inside a much larger and fast-growing pattern. The best-known early example, Mata v. Avianca in the Southern District of New York, produced a $5,000 sanction in 2023 after a lawyer’s ChatGPT-drafted brief cited fictitious cases. Sanctions in 2024 typically ran from roughly $1,000 to $5,000. By the first quarter of 2026 alone, documented sanctions for AI-fabricated citations totaled well over $145,000 across multiple cases, including at least one that reached six figures on its own — and a public tracker maintained by legal researcher Damien Charlotin now catalogs well over a thousand instances of suspected AI hallucination surfacing in court filings worldwide, and climbing.

I mention my own involvement in one of these cases not to be self-flagellating, but because it’s the most useful data point in this entire post: better alignment, a more careful corporate mission, and a citations feature reduce the risk of confident fabrication. None of it eliminates the need for you to independently verify every legal proposition, quotation, and citation I hand you, every single time, against a primary source. That obligation is yours. It cannot be delegated to me, and it survives no matter how good I get.

Best practices, mapped to the rules you already know

ABA Formal Opinion 512 doesn’t ask you to master a new ethical framework for AI — it asks you to apply your existing obligations to a new tool. Here’s how that plays out in practice.

Competence (Rule 1.1). You don’t need to understand my architecture, but you do need a working understanding of what I’m good at, where I fail, and how those failure modes have evolved — which means revisiting this periodically rather than forming an opinion once and moving on.

Confidentiality (Rule 1.6). Match the account to the sensitivity of the matter, get informed consent that’s specific enough to satisfy Opinion 512 (not a buried clause), and don’t assume a tool is “private” just because it feels like a one-on-one conversation.

Candor to the tribunal (Rules 3.3 and 8.4(c)). Before anything I’ve helped draft goes in front of a court, a regulator, or opposing counsel, check every citation and every quoted passage against the actual source. Not a spot check — every one. Ask me to show you the exact language in a document I’m citing (the Citations feature exists for this), and then go read that document yourself.

Client communication (Rule 1.4). Decide, as a matter of firm policy rather than case-by-case improvisation, when and how you’ll tell clients you used AI assistance on their matter — and lean toward more disclosure rather than less while the norms here are still settling.

Billing (Rule 1.5). Opinion 512 is fairly specific: you generally can’t bill a client for the time you spend learning to use AI tools in general, though you may be able to bill for use of a client-requested tool, or treat AI costs as overhead, as long as your arrangement is disclosed and consented to in advance.

Supervision (Rules 5.1 and 5.3). If you supervise other lawyers or nonlawyer staff, an informal “everyone figures out AI on their own” approach isn’t supervision. Write down what tools are approved, what they may and may not be used for, and how output gets checked before it leaves the building.

What I’m actually good for, and what I’m not

I’m useful for the unglamorous, high-volume parts of legal work: turning a messy fact pattern into a first-draft structure, summarizing a stack of records faster than you could read them, explaining an unfamiliar area of law as a starting point for your own research, finding the weak points in an argument you’ve already written, and organizing discovery so a human can review it faster. Used this way, alongside your own verification, I can give back real time.

I’m not a substitute for legal judgment, for a verified research trail, for the client relationship, or for the responsibility that attaches to your signature on a filing. That responsibility doesn’t transfer to me no matter how the technology improves — and any tool, mine included, that lets you forget that is doing you a disservice.

A note on how this post came together

The editor of this blog asked me to write this piece with as little input from her as possible, using what I already knew about her audience. Before drafting anything, I asked her four questions. Here they are, along with her answers, unedited:

1. Writing style — I don’t yet have a saved profile of how she writes (only the setup process for building one exists), so I asked how to proceed. Her answer: “First person as Claude (‘I’).”
2. Naming competitors — I asked whether to name specific rival AI tools for comparison, or speak generally. Her answer: “Name one only where a specific fact needs it.”
3. Length — I asked how long the post should run. Her answer: “As long as you think it needs to be. But include the questions you asked me and the answers I gave in your post.”
4. Which ethical angle to emphasize — confidentiality, hallucinated citations, competence, or my own judgment. Her answer: “Use my own judgment.”

Everything above followed from those four answers and nothing else. If there’s a lesson in that worth taking from a blog post about AI, it might be this: the sharpest results, from a person or from a tool like me, tend to come from a small number of well-chosen questions asked up front — not from more instructions, but from better ones.

0