AI Disclosure Statement

AI is integral to how I work and assists in everything I publish.

Its contribution varies by work, but it is not incidental.

This statement explains what AI contributes, how I evaluate its work, and what remains mine.

What AI Contributes

I use AI for research, analysis, writing, software development, critique, and refinement. It may help me examine an argument, organize a piece, draft or revise language, write or test code, challenge a conclusion, or identify something I have overlooked. Sometimes I begin with written material. Sometimes I begin with an idea, a position, or a problem and use AI to help develop it.

AI may contribute through both generation and revision. It can suggest a structure, draft complete passages, rewrite something I wrote, identify a missing argument, or persuade me that an approach should be reconsidered. Its contribution is real even when none of its original language survives, because it may still have influenced the structure, reasoning, or direction of the finished work.

What I do not transfer is final judgment. AI can analyze, criticize, compare, and recommend. It does not decide what I believe, what I will recommend, or what I am prepared to publish under my name.

My Starting Point

Trust follows verification. With people, trust can come first. With AI, I reverse the order and begin with verification. Whatever trust a model earns is limited to the kind of work I have tested, and it must be earned again when the work, context, or consequences change.

A first response is a draft, not an answer. I treat correctness and alignment as separate tests. A response can sound correct and be wrong. It can also be factually correct while answering the wrong question, which can be harder to recognize because the response may appear entirely plausible.

Context is incomplete. I assume that a model is working from an incomplete picture and may have received incomplete instructions, whether they came from me, another person, or another system. Neither fluency nor confidence establishes that it understood the problem.

Rigor scales with consequence. Not every interaction warrants the same process, but consequential work warrants greater scrutiny.

How I Use AI for Research

I begin with the problem, not the conclusion I expect to reach. When possible, I withhold my own hypothesis long enough to avoid steering the research toward confirming it. For substantial research, I normally begin by testing whether the model understands the question. I have it restate the problem, identify what it does not know, and surface assumptions or missing context.

I constrain the research process more than the conclusion. The model has room to challenge my framing and follow evidence in directions I did not anticipate. It does not have room to present unsupported recollection as research.

For factual claims drawn from research, I require evidence I can inspect. That evidence may come from published sources, direct observation, original analysis, or testing. AI may perform much of the source collection, comparison, and claim-by-claim checking. I personally read and evaluate the finished work. When a claim is material, uncertain, or outside an area I know well, I examine the underlying evidence directly and require additional support when necessary. Another model repeating a claim does not verify it.

Research ends when I have enough reliable information to act or reach a position. If the result does not make sense, the research is not finished merely because sources were found.

How I Use AI for Writing

Writing with AI is highly iterative. I may give it existing prose to edit, an argument to develop, a fragment to build from, or a direction that has not yet taken written form. It may help with structure, draft complete passages, challenge the logic, test whether an idea survives being written down, or offer language I had not reached on my own.

What comes back is material for judgment and revision. I question it, mark it up, combine it with other material, redirect it, and sometimes discard it completely. If the language is weak, I revise the language. If the framing is wrong, I reconsider the section rather than polishing a better answer to the wrong question. I determine the position, decide which arguments survive, resolve the final meaning, and continue revising until I am prepared to publish the work as my own.

How I Use AI for Code

With code, planning normally precedes implementation. The depth of the plan depends on the work, from a short design outline to a detailed specification. I use the planning process to answer three questions:

Not every task needs a new plan. I may let the model proceed directly when the work is simple or exploratory, or when an established skill or command already defines the process well enough that I do not expect it to deviate.

Once implementation begins, the level of oversight and testing depends on the work. I prefer deterministic checks wherever they apply. Linters and schema validators evaluate defined conditions consistently, while automated tests make it possible to run the full regression suite and confirm that new work has not broken what was already working. For aspects that cannot be checked mechanically, I rely on my own review and judgment, sometimes combined with another model to challenge the work.

Verification may range from reading back a small change to unit, system, end-to-end, and security testing. The process becomes deeper as the complexity and consequences increase.

How I Use AI in Decisions

I use AI to expand and challenge my thinking by examining the available options, the argument for each, the strongest case against each, and a response to those objections. A useful recommendation must account for its strongest objection, not merely acknowledge it and move on. That process may provide a perspective I had not considered, but I determine which evidence to trust, which tradeoffs to accept, and what course to take. The decision, and the responsibility for it, remain mine.

How I Verify

I personally read and evaluate everything I publish. AI may assist with verification, but I do not treat its confidence or agreement as proof.

For low-consequence work, my review and judgment may be sufficient.

When more rests on the result, I ask the model to challenge its own work and identify weaknesses, assumptions, missing evidence, and plausible alternatives. Where the work permits direct source checking or deterministic testing, I use it.

When the work must withstand greater scrutiny, I separate creation from review. A separate agent, fresh context, or different model examines the work without inheriting the original conversation. For the most consequential work, I may use models from different vendors and assign each a different review role. These additional reviews can expose weaknesses and shared assumptions, but they do not establish correctness. Agreement between models is not verification, and disagreement identifies a question to resolve rather than an answer to accept.

What I Disclose and Own

I link to this statement from everything I publish using the label “AI Disclosure Statement.”

Disclosure does not divide accountability between me and a model. Every position I publish is mine. Every recommendation I make is mine. AI’s involvement in research, reasoning, drafting, revision, or verification transfers none of that responsibility.

If something I publish is wrong, responsibility for the error is mine. If I discover a material error, I will correct it.

Last updated: September 2026