charles · Mindful Wellness

Charles Mindful Wellness

AI Use & Disclosure

Copy on this site was drafted with AI assistance from the program's master specification, and reviewed and approved by Charles Muhammad, who retains full authorship and final judgment. The policy that governs every AI decision at CMW is published here in full.

This is CMW's umbrella AI policy. It covers how AI touches CMW's work as a whole. Individual programs may carry additional requirements layered on top of this document, not in place of it.

Preamble: What Governs This Policy

Charles Mindful Wellness (CMW) is a wellness practice serving Black queer communities. It uses AI tools in parts of its work. This policy sets the rules for how, and you don't need to know anything else about CMW to follow it.

One idea sits underneath every rule here. It's called AI Heterosis, and it borrows a word from plant breeding: heterosis is the extra vigor that shows up when two genuinely different lines cross. The claim is simple. Intelligence gets stronger when honest difference meets honest difference. It gets weaker when everything converges on one optimized answer. Much of the AI industry is chasing that second path: a single intelligence, refined toward a fixed endpoint, built by whoever controls the most computing power. Every AI decision CMW makes runs against that logic, not toward it.

This isn't abstract. AI touches actual bodies here: program participants, retreat guests, and the archive of Black queer knowledge CMW is building through its published work. So name what actually happens when CMW uses an AI tool. Two kinds of human knowing meet. The first is present and accountable: Charles, participants, people whose knowledge comes from lived consequence and who answer for it in real time. The second is distant: millions of people whose words and thinking were collected, mostly without credit or consent, and compressed into the tool. The machine is not a mind and not a partner. It is the filter those distant people get encountered through, fluent by design and accountable to nothing until a human steps in from outside it. This policy is how CMW keeps that meeting honest instead of extractive.

A crossing where everyone consented is exchange. A crossing without consent is extraction. Every section below exists to keep the difference between those two things visible and enforced.

1. Data Sovereignty and Consent

Your data belongs to you. CMW is a steward of it, not an owner.

  • Sexual orientation, gender identity, HIV or health status, race, and any nude or clothesfree imagery are never entered into a third-party AI tool without explicit, revocable, informed consent from the person the data belongs to.
  • Consent is asked for the specific use, not collected once and reused. If AI use expands beyond what you agreed to, you get asked again.
  • If your words, images, or stories are used in AI-assisted content, four commitments apply: your consent before use, compensation where the material generates revenue, control retained by you over your own material, and credit given by name unless you ask for anonymity.
  • Data use has to serve the community it came from, not just CMW, and CMW answers for how that data is handled.
  • Small groups carry a specific risk: a retreat cohort or a book club is small enough that even de-identified data can point back to someone. Any AI-assisted analysis of participant data accounts for that risk before it runs, not after.

2. Sensitive Data: Clothesfree and Body-Based Programming

Clothesfree, nude, and other body-based programming carries the highest-risk data CMW touches, in any program where it appears, present or future. The documented harm is real: platforms and AI systems have repeatedly flagged, removed, and de-monetized Black, queer, trans, and nude bodies at rates far above general content, and moderation systems routinely treat HIV and sexual-health material as inherently dangerous.

  • No participant image or video from clothesfree or nude programming goes through an AI image tool, a general cloud AI service, or any facial-analysis system, ever. No exceptions.
  • No biometric or facial-recognition processing of any kind touches CMW's community. Peer-reviewed research has documented facial-analysis error rates of 34.7% for darker-skinned women against 0.8% for lighter-skinned men, a 43x disparity built into the design. That is not a risk CMW is willing to run.
  • Participant names are never linked, in any AI prompt or document, to nudity, HIV status, or sexual orientation. If a name has to appear in AI-assisted work, the sensitive detail gets stripped from that instance first.
  • Anything tied to clothesfree or body-based programming is stored locally, not synced to a cloud AI service by default.
  • AI tools can be given aggregate patterns, general accessibility needs, and program logistics in service of program design. They are never given any specific participant's body, health status, or identity details, unless that participant has consented to that specific use.

3. Language Justice

AAVE and Black queer vernacular are not errors to be corrected. Peer-reviewed research has found AI systems flag African American English as toxic or offensive at roughly twice the rate of standard English. That bias runs directly at CMW's own writing voice and at the language of the communities CMW serves.

  • AI tools do not get final say over CMW's language. Anything an AI tool flags, corrects, or "cleans up" in Charles's words or in participant language gets human review before it's accepted, and the default is to keep the original.
  • Culturally specific language, reclaimed terms, and ballroom and Black queer vernacular are preserved deliberately. If an AI tool suggests replacing them with more "neutral" phrasing, that suggestion is rejected as a matter of policy.

4. Platform Strategy

The pattern is documented: major social platforms have removed or suppressed content from Black creators, queer creators, trans bodies, sex educators, and fat bodies at disproportionate rates. TikTok has publicly admitted to restricting LGBTQ+ hashtags, and an annual industry safety index has given failing grades to five of six major platforms for three years running.

  • Platforms are treated as untrusted distribution, not primary storage. Owned channels (the newsletter, the zine, charlesnarles.com) hold the real archive. Platforms are where CMW shows up, knowing the content could disappear without explanation.
  • Every takedown, shadowban, or unexplained reach drop gets documented: date, content, platform, what happened. CMW keeps its own evidence rather than relying on platform transparency that doesn't exist.
  • Where deliberate word substitution is needed to avoid moderation flags, it gets used, and the original uncensored version is always preserved in CMW's own archive.

5. Publication Standards

This covers newsletters, zines, long-form guides, and any AI-assisted content CMW publishes.

  • Disclosure: work that involved substantial AI drafting, editing, or research assistance says so. Charles retains full authorship and final judgment on all content regardless of AI involvement.
  • Bias review: AI-assisted content gets checked for stereotyping before publication. One major study found a leading generative model produced negative associations for queer people 70% of the time. CMW's content does not get to reproduce that pattern by accident.
  • Fact-checking: claims sourced through AI research get verified against the primary source before publication, not taken on the AI's word.
  • Credit: where a framework or methodology is original to Charles, that origin is stated plainly. Where it comes from someone else, they are credited by name, not absorbed as CMW's own.

6. Vendor and Tool Vetting

Not every AI tool is equally trustworthy with CMW's material. Every tool gets checked on two fronts: what it does, and where it runs.

What the tool does. Before a tool touches CMW work, four questions get asked:

  • Has this tool, or a close relative of it, been documented targeting or disproportionately burdening a marginalized group?
  • Do the tool's default outputs assume a demographic that isn't CMW's community, treating everyone else as an edge case?
  • Does the tool claim to protect privacy while actually increasing surveillance of the people using it?
  • Is the tool marketed as helping wellness or community organizations while its actual function is data extraction? This is the pitch CMW is most likely to hear directly, and the one easiest to miss when the sales copy sounds warm.

Alongside those: Does the vendor train on user input by default, and can that be turned off? What is the retention policy, and does CMW's data get deleted on request? Is there a stated way to report harm? Has the tool been documented, in peer-reviewed research or credible reporting, to carry racial, gender, or queer bias?

Where the tool runs. A tool can pass every question above and still be built on infrastructure that harms CMW's community. The compute behind these systems runs somewhere, on land, drawing power and water, and the documented pattern is that the burden lands on majority-Black neighborhoods and communities that look like the ones CMW serves. Before adopting a major vendor, CMW checks the public record for documented environmental or community harm from that vendor's infrastructure. Avoiding all cloud computing is not realistic for a solo practitioner, so this check is weighed rather than treated as automatically decisive. Weighed, not ignored.

Tools that fail more than one check get used cautiously, with more human review, or not at all for sensitive material.

7. Accountability When Harm Happens

CMW sorts AI-related harm along its Spectrum of Harm: inconvenience, exclusion, surveillance, misclassification, violence. The response scales with the tier, and this section is honest about what power a practitioner of CMW's size actually has.

When CMW's own process causes the harm. This is the one place full repair is entirely within CMW's control, and it applies at every tier. If you believe AI-assisted CMW content or process caused you harm, you have a direct way to say so: email, in person, or through program facilitators. CMW corrects the content or process, states plainly what changed, and you see the correction happen. This is not conditional on the harm being severe. A misapplied name, a mishandled detail, a language "correction" that erased something intentional: all get the same standard. Named, fixed, shown.

When the harm comes from outside CMW (a platform, a vendor, a system the community is forced to navigate):

  • Inconvenience: logged. No further action unless it becomes a pattern.
  • Exclusion: documented with specifics (what happened, who was affected, what it cost them), the immediate case corrected where possible, and the tool's continued use for that function put under real review. A repeated pattern from the same tool ends that tool's use for that function.
  • Surveillance: use stopped immediately on discovery. No review period. Anything that tracks, profiles, or logs behavior beyond the specific task it was asked to do does not get a second chance.
  • Misclassification: permanent disqualification, consistent with the standing ban on biometric and facial-analysis tools in Section 2.
  • Violence: categorical refusal of any tool or partnership that feeds into systems capable of this tier (policing, family separation, bail-setting), even indirectly. Where CMW's community is exposed to a violence-tier system through no choice of CMW's own, that gets documented and the community is warned directly.

What CMW does with the power it has. CMW cannot sue a platform or force a vendor to change its training data, and pretending otherwise would be dishonest. What CMW can do, at exclusion tier and above: feed documented incidents into existing coalition evidence pools rather than letting them sit in a file, file formal complaints with regulators when a pattern repeats, and name vendors publicly in CMW's published work when CMW declines to renew and is confident in the claim. None of this moves a company by itself. It is testimony as resistance, applied to a domain where legal and economic leverage isn't available at CMW's size.

8. Community Governance

This policy is not Charles's alone to hold. The operating principle is a simple one: nothing about us without us.

  • This policy is reviewed at least annually, and any time a new AI capability or tool enters CMW's practice.
  • A standing invitation exists for QTPOC community members, especially those in clothesfree, body-based, or identity-specific programs, to weigh in on how AI touches programs that involve their bodies and stories. This is not a survey CMW runs once. It's an open channel.
  • Charles holds final decision authority as founder, and community input on this policy is sought deliberately, not incidentally.

9. International Programming

Retreats and programs held outside the United States carry cross-border data risk that domestic programs don't.

  • Participant consent for AI-assisted data handling is confirmed under the laws of both the participant's home country and the host country, whichever is more protective.
  • Participants traveling from or through countries with anti-LGBTQ laws are flagged internally, with their knowledge and consent, so that no AI-assisted communication, marketing material, or data storage creates a paper trail that could expose them. This is not paranoia. AI-assisted surveillance has already been used to track LGBTQ+ people in some of the countries CMW's community travels through.
  • Local data protection law at the retreat site is researched and followed in addition to CMW's own standards, not instead of them.

10. What This Policy Does Not Resolve

Naming the edges honestly:

  • CMW can secure real, specific consent from its own participants under Sections 1 and 2. CMW cannot secure consent from the millions of uncredited people whose language and thought are already compressed into whatever tool CMW uses on a given day. This policy names that gap plainly instead of calling it settled, and refuses to pretend the consent CMW can get is the whole of the consent that's owed.
  • Material already used to train these systems before anyone was asked has no full remedy yet. This policy holds the line going forward. It does not claim to undo what's already been taken.
  • This policy will be wrong in places. It gets fixed in the open, through the governance process in Section 8, not quietly.

This policy operationalizes AI Heterosis (Charles Muhammad, 2026). The research findings referenced throughout draw on the work of Joy Buolamwini, Timnit Gebru, Émile Torres, Ruha Benjamin, Safiya Umoja Noble, Mutale Nkonde, Os Keyes, Sasha Costanza-Chock, Yeshimabeit Milner, Data for Black Lives, Hacking//Hustling, GLAAD, and UNESCO. Full citations available on request.