Guiding a creative team to use AI

The Landscape

Asking a room of artists to adopt AI tooling can feel like herding cats, but It is a mistake to think it is a training problem, it’s a trust problem.

The objections I hear are not irrational. The models were trained on work that looks familar. The marketing around these tools spent two years openly promising to replace the people I lead. And a lot of the output was, frankly, outright insulting the teams who have spent their lives learning to make it good.

No argueing someone out of that. What you can do is be straight with them, and give them a reason to touch it that isn’t a threat.

Write the policy before you need it

In 2021 I wrote an AI policy for the studio. I’ve updated it every year since, and having group conversations with everyone, so people know they are heard. I wrote it early on, before clients asking for it and before deadlines forcing a decision. So the first conversation was controlled, safe, and most importantly had room to breathe.

The policy covers what we will and won’t put into which tools, how client IP is handled, what has to be disclosed, and what we do about provenance. It’s short, it’s specific, and it’s honest about the unsolved parts, and acknowledges it will be evolving.

It also allowed people could disagree with it in a specific way instead of feeling like the ground was moving underneath them every quarter.

One step, not the whole job

The failed pitch is “this will do your job faster.” People will only hear ‘faster’ and think, now I have to do more. Even the people who are curious hear it, and it puts them on the defensive.

The pitch that worked was personalized, what is an annoying step in your process ( which are frequently repetitive and dull ), and now you can do it in seconds instead of an afternoon. See my thoughts on doing repeat work from well over a decade ago.

People are less likely to feel replaced when the thing being replace is something they dont like. Once someone has had a genuinely good experience with a tool, they’ll go looking for the second use case on their own. Some examples people were shown how do do that started catching their attention were:

• Rotoscoping a plate you were going to hand-roto anyway.
• Generating twenty reference variations or explorations for a lighting direction.
• Cleaning up a background element nobody will ever look at closely.

Ambassadors carry it further than you do

When I would demo something to the whole studio it would not stick, but. A senior artist showing a peer the thing they used on last week’s job sticks immediately. So I stopped trying to be the source. I found the three or four people who were already curious, usually the more technical artists, not always the most senior, and I gave them time, early access, and my attention.

Then the adoption moves sideways, artist to artist, which is both faster and more credible than somthing from the mothership. It also means the use ideas spread that are the ones that survived contact with real work.

Teach the machine, not the button

I also sat with each of the artists working with them to not just develop skills but more importantly develop understanding, by giving them an appropriate level technical primer for their understanding and tasks, something I take seriously. When I explain roughly how these systems work, like What a diffusion model is actually doing, or why a language model produces halucinations, the attitude goes from “the AI is dumb” and “every success is luck” to, “this is a tool, that allows me to do things”.

When you dont simply train people on software, and build a mental model model accurate enough, they can reason about it when they reach a hurdle. That’s the same reason I’d rather explain why the pipeline is shaped the way it is than write a longer instruction manual.

Come back around

Adoption is not an event. I check in and stay in tune with process and execution. Some things stuck and became part of how people work. Those get elevated: documented, standardized, folded into the pipeline, presented back to the studio as the way we do this now. I have knowledge baske and workshop ticket forms for people to request more info or documentation for the rough edges as well.

Other things fail entirely. Someone was enthusiastic in March and hasn’t opened it since. The space is developing fast and things will not make it to the year end OKR before they should be let go, and finding them early, keeps people engaged and growing.

Either way, the reinforcement is what makes it real. One demo is a novelty. The fourth conversation about the same tool is a workflow, and that is where people become comfortable with the colorfull new thing.

Some people will say no

And they need to say it.

I’ve had people tell me directly that they aren’t going to use these tools, on ethical grounds or on craft grounds, and they’ve meant it. The worst thing I could do is treat that as a problem to be explained away or managed quietly. Ambiguity + New thing ≠ Good feelings.

Naming and discussing it clears the air. It lets us plan honestly around what work goes where, and what to do next.

Clear-eyed disagreement is a much healthier state for a studio than polite ambiguity. That’s true for AI and it’s true for every other contentious thing a team has to navigate together.

🐈