What AI Can and Cannot Do for Cultural Infrastructure
Research on generative AI shows task-specific gains and real limits. The important question is what a tool can help with, and what still needs human judgment.
I am interested in tools that make useful work easier to do. I am not interested in pretending that technical speed removes the need for cultural knowledge, consent, or a person who is accountable for the result.
The results depend on the work
A field study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond followed 5,179 customer-support agents and about three million chat conversations. It found that access to a generative-AI assistant was associated with a 14% average increase in issues resolved per hour. The estimated gain was larger for novice and lower-skilled agents, while experienced agents saw little effect. The study concerns one company and one support workflow, so it is evidence about that setting, not a general productivity rate for creative work. Read the NBER working paper, Generative AI at Work ↗
A different result came from an experiment with 16 experienced open-source developers working through 246 tasks in repositories they knew well. In that study, access to early-2025 AI tools increased task completion time by 19%. The sample is small and specific, but it is a useful counterexample to the idea that adding a model always makes skilled work faster. Read the METR study of experienced developers ↗
The studies do not cancel each other out. They show why claims about AI need to name the task, the people, the tool, and the measure. Faster output is not automatically better work, lower total cost, or less coordination.
Exposure is not the same as impact
The International Labour Organization estimates that one in four workers globally are in occupations with some exposure to generative AI, while 3.3% of global employment falls into the highest exposure category in its index. The ILO is estimating modeled task exposure. It is not saying that one in four jobs will disappear, or that employers have already seen a matching productivity gain. The report expects transformation of tasks to be more common than wholesale replacement. Read the ILO working paper, Generative AI and Jobs ↗
Some decisions should not be delegated
AI is a tool for infrastructure, not for creating this Journal writing or artists' work. People still have to decide whether a result is accurate, whether it respects the source, whether a person agreed to be included, and whether the work is useful to the community it describes.
For cultural infrastructure, that distinction is practical. A model cannot infer that a listing is current, that an artist welcomes a specific approach, or that a location is appropriate to publish. It cannot turn an unagreed relationship into consent. It also cannot decide what a client should pay for local knowledge or who should deliver the work.
Lean does not mean less human
I do not think the right response to new tools is to treat every human task as overhead. If some parts of a workflow become quicker, attention moves to the brief, context, review, credit, rights, production, and follow-through. Those are not decorative additions to the work. They are part of what makes the work usable and accountable.
That matters when people talk about building a creative network with AI in its infrastructure. A useful network is not just a database plus a model. It needs a public record people can understand, clear boundaries around private information, maintained relationships, and a route from a specific brief to paid local work.
How this applies to Countercult
Countercult's public Network, protected member context, Studios delivery, and Fund have different roles. The public layer helps people discover artists, spaces, and places. Studios can deliver a defined cultural brief. The Fund is a business-owned reserve for selected local creative work, not an automatic payout loop. I have set out those boundaries in What Countercult Publishes, Protects, and Builds.
I want technology to help the work become more precise and easier to maintain. It should not make our decisions harder to inspect or our relationships harder to understand. If a tool reduces friction while people retain judgment and responsibility, it can be useful. Speed by itself is not the outcome.
An ethical data commons matters because artists' unique IP and community knowledge need protection. It must include artist consent, clear purposes, and shared governance, rather than treating creative work as raw material.
The question is not whether AI can generate more. It is whether it helps us find each other and claim our seat in the circular economy.
Have a cultural or place-based brief?
Countercult Studios works with organizations on research, cultural strategy, curation, production, and place partnerships.
The productivity results cited here are specific to their samples and tasks. The ILO exposure estimate describes modeled technical exposure, not job loss, adoption, or realized productivity.
This piece was informed by Shona Neary’s writing, conversations, and thinking around independent culture, artists, and place.
editorial process: Our Journals are pulled from our writings over the years to get information out as quickly and as widely as possible, and to get to what matters: finding each other and claiming our seat in the circular economy. Some ideas overlap between Shona Neary's work on cultural infrastructure and their artist practice as Sketchy Shona. AI is a tool for infrastructure, not for creating Journal writing or artists' work. We welcome original reporting and pitches from journalists and independent publishers.