Case Study: The Policy Was Good, but it Didn't Cover This.
If you haven’t been following it, there’s been a great deal of hullaballoo over science communicator, novelist, and Internet personality Hank Green using AI. When Hank Green apologized for how he had been using AI, he linked to his own company's AI policy and told people to go read it.
He wrote that the policy is "public and well-considered and adhered to by the whole team." In the same post, described the practice that got him in trouble: leaning on AI to find papers and sources for videos on his personal channel (which Complexly does not produce).
I am not going to relitigate the controversy. I think the public conversation about it lost the plot early, and I don't want to add to it. What I want to do is read the document itself, because it is a rare artifact: a values-first AI policy, written by a content organization, published where anyone can check it.
The Policy
The policy has an interesting structure. It gives four reasons, then four rules, and every rule traces back to a stated reason.
The reasons:
Errors. Their content is distinguished by its integrity, these tools invent things confidently, and it is hard to catch errors at the rate the tools produce them.
Respect for other creators. The models were trained on people's work without permission.
Original thought. Their stated goal is to "imagine the world complexly," and a language model cannot generate original ideas or examine its own biases.
Ownership. Under US law, it is not settled who owns AI output.
The rules:
No videos written, edited, or fact-checked by AI including drafting or partial composition.
No public-facing illustrations or animation from models trained on non-consensual data sets. Models trained on images whose creators agreed to be in the data set may be used.
When using consensual-data tools for visuals, don't mislead viewers about what is real, and label where applicable.
AI may suggest ideas and do preliminary research, including identifying sources, provided the team verifies them.
Five Good Things
It gives reasons before it gives rules. In Amplify Good Work I argue that a policy people will actually follow has to justify itself, especially when it forbids something staff may already be doing. Rules without reasons produce secret use among staff, especially those who like AI and those who are overwhelmed by their workload. This document spends about as many words on “why” than on “what not to do.”
It's a management document. Often, AI policies are IT documents only, focused on security, licensing, and monitoring, but leaving aside questions of staffing and strategy. Every clause in the Complexly policy answers the question: who should do this task?
It refuses once and constrains three times. I describe six stances an organization can take when AI threatens a value: Refuse, Wait and See, Constrain, Compensate, Rethink the Work, and Shape the Ecosystem. Read quickly, this policy looks like a ban. But really, only the first rule is taking a Refuse stance: “We will not produce videos written, edited, or fact-checked by AI.”
The other three are Constrain: narrow boxes with guardrails: you can augment your work on this task with AI, but you need to use a particular type of tool, label your use, and keep AI use backstage, as it were. (You can learn more about the consensual data tools the policy refers to in the Ownership and Intellectual Property chapter of AGW)
It scopes by task, not by tool. This aligns with my recommendations: break the work into tasks, decide which ones need a human (writing, editing, and fact-checking), protect those, and offer guidance for doing the rest well (be cautious and clear about AI generated images, use consensual data models, verify sources it identifies, and support diversity of thought when generating ideas for videos.)
It rules by criteria and is not bound to the current state. Rule 2 does not simply ban DALL·E 2 and Midjourney. It bans models trained on non-consensual data, and offers those two as examples. This rule is therefore resilient to product changes: if a currently compliant product starts using non-consensual data, it’s off the list. (And if somehow one of the larger tools starts over with consensual data, well great!) They don’t have to go through a whole process to change the list of approved tools every time the available tools change. This may not be appropriate for companies with higher security concerns who want to manage all the licenses, but it is one way to keep your policy flexible and relevant.
The strongest single line in the document is the throughput claim in reason 1.
“We have fact checking processes in place, but it’s very hard to catch errors at the rate these tools produce them.”
It’s important that it can be wrong, certainly, but more importantly, errors arrive faster than their review process absorbs them. That is an realistic assessment of their own capacity, and the rule based on it is resilient to model improvements. It is also the same conclusion I come to in the Accuracy chapter: if you cannot afford to fact-check, you cannot afford to do the task.
Room for Clarification
The scope is public-facing content, and that's not everything. The policy doesn’t address more behind-the-scenes tasks, like internal or sponsor correspondence, community moderation, transcripts, metadata, or subtitles and localization. This could be an example of Drift— not having a policy and allowing a free-for-all— but it’s also possible that we got the public-facing policy for public-facing work because that’s what’s relevant to us… the public :)
There is no privacy or security clause. Rule 4 addresses who owns the output, but nothing addresses the characteristics of the input. This doesn’t surprise me all that much, though: the most sensitive content that Complexly is likely to have relevant to this public-facing policy is its own unpublished content. Many organizations feel that the risk to that kind of data is small and worth the benefits they get from using AI. That’s their call to make because the risk accrues only to them.
I would hope that any additional internal policy (for Complexly and any other organization that has staff and subscriber data) forbids uploading personally identifiable information into a chatbot that permits reuse of prompt data. (More on this in the Privacy and Security chapter, but check your settings (turn off “improve model for everyone” or similar) and your terms of service.)
The ban on AI fact-checking is wider than its own argument. The stated reason is that AI produces errors faster than humans catch them. That argues against letting AI write unfact-checked and against letting AI be the only factchecker. It does not argue against AI as an adversary, which can be a powerful tool in ensuring that scientific writing is not only technically accurate, but also clear, not misleading, and as complete as necessary. Using AI as an adversary could sound like: “What are the weakest parts of this script?” "What claims in this draft would a hostile viewer challenge?" “What could be clearer?” They could also (if they wanted!) create a skill in their chatbot of choice for fact-checking to add a layer of verification, rather than replacing the human doing the work.
The reasoning is relies on a changeable state, but there’s no mechanism to monitor it. The rationale hedges four times: these tools currently make mistakes, currently are not capable of this, legal questions remain. Those are Wait and See arguments, but used to support a Refuse stance. This could be stronger if the company either made a permanent Refuse decision (“We think these tools will never be trustworthy enough, therefore…”) or supported the Wait and See stance with a process (When, how, and among whom do we reevaluate the state of the models? What is the policy if the current state changes?)
It doesn’t address Job Security & Quality. The policy names harm to fellow creators and then declines to participate in it by choosing consensual data models instead (in fact, the exact way I suggest protecting Ownership and Intellectual Property values in the book!) These consensual data models protect artists’ ownership over their work by allowing them to opt in (and out) of their inclusion in the training data set, but as far as I am aware, they don’t pay artists for that work.
The policy could go one step further and protect Job Security and Quality for artists. All this would take is formalizing something that I know Complexly already does: commissioning artists in other areas of the business. (This is a Compensate stance: we’ll use the time and money we save using AI elsewhere to reinvest in a value, in this case, in paying artists.) Specifying under what conditions work should be paid or commissioned is certainly an optional constraint on their work, but could ensure that the practice of hiring human artists stays protected.
Questions Worth Asking About Your Own Policy
If your organization has an AI policy, or is drafting one:
Does your policy include reasoning for rules, and does that reasoning support the rules?
Do the rules leave room for flexibility as tools, norms, competition, and customers change, centered on the reasoning?
Who owns this document, when is it next reviewed, and how does someone request an exception?
Where does your policy identify a harm to someone outside the organization? Is compensating available?
Do you have a policy covering the less visible parts of your work: internal communication, backstage work, personal accounts? Who's watching that?
LLM Disclosure:
First, I read all the content myself. Then, I requested that Claude Opus 5 analyze the policy based on the frameworks in my book in two ways: 1) as a “full pass” and as a teaching example. I asked it to analyze instead of draft because I wanted to see whether and how it was different from my thinking and to get an analysis of it that was not informed by the context of the online drama (like I had been).
I read both, then I asked for a draft and added a bunch of additional information: my thoughts on the policy, the relationship between Hank Green and Complexly, and the text of the apology. I added “I'd like to use this conversation as a launching point for the analysis of the policy, but not spend a lot of time commenting on the actual controversy itself.”
Despite all that context, this still required quite a bit of editing. In addition to annoying AI voice stuff, the Good Things and Room For Clarification were thin (and they had annoying, AI-voice titles) and abstract in spots and clearly missing the experience of working in an actual organization with people. I removed some points, added a couple, and rewrote a few entirely. I did like the structure (not my strong suit) and it caught something I didn’t: the lack of a privacy and security clause. Classic human: it’s easy not to notice what isn’t there.
You’ll notice that I did not follow Complexly’s policies when writing this, because I had it write a draft. I try to compensate for that (IMO) by over-disclosing my use in this very note, so you can decide whether and how much to trust me :)
