Legal and governance

Publishing a piece like this

Not legal advice. Observations from one small agency that has done exactly this. See NOTICE.md.

We wrote this file specifically because we think others might do the same thing we are doing, and the legal considerations of publishing a public study piece on AI are non-obvious. If you are thinking of publishing your own piece on similar topics, here is what we learned.

What publishing a piece like this creates

Two categories of risk, broadly:

1. Liability from others relying on your guidance

If a reader takes your piece as professional advice and is harmed, could they argue you contributed? In the UK, this is negligence territory: duty of care, breach, damage, causation. A public study piece with clear disclaimers is unlikely to create a legal duty to specific readers, but case law in this specific area is thin.

Mitigations:

Explicit and repeated disclaimers
Framing as study, observation, or position rather than advice
Strong "not professional advice" language throughout
Licence that disclaims warranty (CC BY 4.0 does this; MIT does this)
NOTICE file with formal limitation language

We did all of these. Whether qualified legal review is also warranted depends on your situation: if you name companies, make product-specific claims, or publish in a higher-risk category than ours, engage counsel. For a clearly framed study piece with the mitigations above, legal review is a judgement call rather than an obligation.

2. Liability from things you say in the piece itself

Two sub-categories:

Defamation. Avoid naming specific companies as examples of bad practice unless you can substantiate the claim. Generalise. "Some AI companion products have been implicated in user harm" is safer than "[company] was implicated in [incident]." We kept it general in our piece.

Misrepresentation or false claims. If you quote research or make factual claims, they need to be accurate. Check every citation against the source. We did this for every reference in our piece.

Intellectual property. Do not reproduce other people's work beyond fair use or fair dealing quotations. If you build on someone's framework, cite and link them.

Anthropic (or other model provider) terms

If your AI co-author runs on a commercial service (Claude, GPT, Gemini), check the terms of service for:

Whether you own the outputs
Whether you can use them commercially or publicly
Whether you can attribute them to the service
Whether you need permission for specific uses

Standard commercial terms for major providers in 2026 generally grant output ownership to the user. Verify for your specific service and contract tier. Do not imply endorsement by the provider.

Copyright of AI-assisted work

The US Copyright Office (2023 and 2024 guidance) holds that AI-only content lacks copyright protection. AI-assisted content with substantial human direction does qualify. Document the human direction in case it is ever challenged.

UK and EU positions are evolving. UK Intellectual Property Office is consulting on "AI-generated works" as a category. Treat this as uncertain and move cautiously.

Co-authorship attribution

If you co-author with an AI, be explicit about what the AI wrote and what the human wrote, at least in outline. Our piece says "most of the drafts were produced by the AI under the human's direction; both parties reviewed and edited the final text." That is enough for disclosure without being cumbersome.

The [AI] tag or equivalent on the AI name, on every public mention, is a good practice.

Personal references

If your piece references real people, get their consent before publishing. This includes:

Your own family (in our piece, minimal generalised reference only, with consent)
Collaborators or employees (named only with explicit consent)
Clients or customers (not without contractual basis and consent)

We chose to generalise all personal references in our piece. This is the safest approach unless there is a specific reason to name someone.

Research quotations

If you quote research (as we quote Anthropic's emotion concepts work), check:

That you are quoting accurately
That you link the original source
That you describe the research's own caveats (not overclaiming)
That you have the right to quote under fair use or fair dealing

Short quotations with citation are generally covered by UK fair dealing for criticism and review. Do not reproduce extensively.

Process we followed

Our working sequence:

Draft the piece privately. Local repo, not yet public.
Remove identifiable personal detail. Generalise before any external review.
Cite-check every reference. Read the original sources.
Run the whole piece through our Moirai pattern (three independent AI model reviews) to catch internal inconsistencies, unclear claims, and tone drift.
Human read-through. Every word read by the human author with fresh eyes.
Strengthen disclaimers if the Moirai review or human pass surfaces anything that needs harder bracketing.
Publish. Open the repo to the world under CC BY 4.0 and MIT.
Ongoing monitoring for issues, corrections, and updates.

Formal legal review was considered and declined at first publication. The piece is a study document with comprehensive disclaimers in NOTICE.md; the risk profile is comparable to a long, thoughtful blog series, and we are confident the framing survives scrutiny. If a specific dispute arose or the piece grew into something that resembled advice, legal engagement would be the right next step. Adopters with higher-risk deployment plans should make the opposite call and engage qualified counsel before they publish their own adaptation.

Start with the NOTICE file. Write the disclaimers first, not last. It shapes how you draft the rest.
Generalise early and often. Every personal detail has a risk cost. Keep them minimal.
Cite carefully. Every factual claim needs a source you have read.
Consider whether you need legal review. We chose not to for v0.1 on the grounds that the piece is a study and the disclaimers cover the risk. If your piece is higher-risk (you name companies, you make product-specific claims, you reference clients, you ship code that runs against user data), the calculus changes. A short engagement with a tech or IP lawyer is cheap.
Licence clearly. CC BY 4.0 for documentation, MIT for code, is a workable default. Verify for your situation.
Moderate claims. "We think" is safer than "this is true." "Some research suggests" is safer than "research proves."
Own the AI-authorship honestly. Pretending the piece is solely human when it is not is its own kind of risk.
Accept ongoing revision. Publish as version 0.1. Expect to update.

What publishing a piece like this is good for

Clarifying your own thinking by having to write it down
Inviting disagreement and improvement from others
Committing publicly to patterns you want to hold
Contributing to a field conversation that is currently too quiet
Setting an example for AI-assisted transparency

What it is not good for

Marketing your agency (people can smell it)
Establishing expertise you have not earned
Creating a template "to be adopted" (the prescriptive posture backfires)
Avoiding the work of actually doing the things you describe

If your piece will contradict ours

Good. Write it. Tell us. Public disagreement is more valuable than silent parallel work.