On AI-Assisted Editorial Systems
Organizations today face a growing problem: presently, Large Language Models can produce prodigious amounts of content, but not at the level of expert specialists or expert writers. The result can too often become what we call “slop”: AI-generated content that is transparently inferior to human content, which readers resent as deceptive, insulting, and a time-waster.
But it would be folly to try and run a modern editorial department without LLMs assisting in the research of topics, the production and improvement of human-overseen drafts, the checks for grammatical and legal compliance, or the use of agentic AI to coordinate between humans and the tools we use to manage projects and communicate with each other, tools such as Slack for communication and Jira for project management.
The basic idea is to automate processes that AI can handle well, using it as a complement, not a substitute, for human judgment. This is all the more important when an organization is trying to demonstrate and publicize the work of technical domain experts, who themselves are almost definitionally not also writing experts.
In the process I use, I poll employees for topics and then assign a topic to an LLM (in my case, ChatGPT). Chat then conducts research, checks its work to eliminate hallucinations, and creates a draft, either using a human-generated first draft or as a preliminary draft from which the expert constructs their own version. I use n8n’s agentic AI capabilities and integrations with Slack and Jira to track progress through each stage, sending drafts back and forth, and triggering tasks on Jira’s scoreboard (the Kanban board) to move along a pre-programmed sequence.
The first draft then goes through a series of LLM-based processes, such as checking it against the in-house style guide and flagging spelling and grammatical errors. Here, the LLM needs to be trained and fine-tuned. ChatGPT doesn’t natively check for the kind of issues that Grammarly finds, but could be trained to do so. Even so, I find that I reject half of Grammarly’s suggestions and make ten times as many changes myself when I copyedit. So even seemingly mechanical tasks are anything but.
Importantly, the author must sign off on the draft before it progresses, since it is vital that any content appearing under someone’s byline has their full buy-in before it goes any further.
Next, the content enters a legal compliance loop, first passing through another LLM sieve trained on the law and standards relevant to the given industry. Again, there are outside vendors who offer such services, which is, in many cases, more cost-effective than producing one’s own, but that decision depends entirely on the organization, its resources, its particular industrial context, and its organizational philosophy. Increasingly, companies are moving towards owning their own AI processes rather than contracting with outside vendors.
But, just as in the case of content and style, a human needs to oversee the legal compliance AI, and the legal draft must be either approved or rewritten by an attorney before going to the next stage, which is another check with myself and perhaps a technical editor if I have such a colleague. This stage is not perfunctory, but is certainly less intensive than previous passes since we are now close to the final draft.
Our AI agentic automater then shuttles the draft via Slack to the management representative, who will either approve it or request edits, before it is sent to marketing for graphic design and scheduling in a Content Management System (CMS).
After publication, we use automated processes to collect metrics and then an LLM to produce a post-publication analysis, which is routed to all relevant stakeholders who were involved in the creation of the piece or are directly affected by it.
When I first started creating content for a student newspaper, we used iMacs that could run only one program at a time, and we still cut-and-pasted photographs into layouts with scissors and glue. Those days are long gone—and as with any technological change, there exists a good deal of nostalgia and even sadness about the passing of the newsroom with its clattering typewriters and ink-stained scribners. But organizations today are tasked with charting a path into the future, not waxing on about the past. And LLMs, modern project management and communication software, and easy-to-use, AI-assisted agentic workflows can not only make institutional content faster but also better. This is a classic example of AI creating jobs rather than destroying them; organizations that previously made do without an editor are finding that AI requires one, and that AI-assisted editorial operations place humans in the loop rather than replacing them altogether.
