AI-assisted vs AI-generated content: where to draw the line

The conversation about AI in content has settled into two opposing camps. One camp says AI generation is going to replace human writing. The other says AI cannot produce anything of value and the marketing team should pretend it does not exist. Both camps are wrong, in opposite directions, and the practical question for B2B tech marketing teams is more useful: where exactly do we draw the line between AI-assisted work that improves the output and AI-generated work that hollows it out.

We have spent enough time using ChatGPT and Claude in production content workflows to have a clear view on this. Here is where we draw the line.

What “assisted” actually means

AI-assisted content, in the way we use the term, is content where a human is doing the thinking and the AI is doing the typing-adjacent work. The human decides the angle, conducts the interviews, structures the piece, makes the editorial calls and signs off the final draft. The AI helps with the parts of the work that are mechanical: sorting an interview transcript, drafting a first-pass outline, generating alternative phrasings for a specific paragraph, summarising a long source document.

AI-generated content, by contrast, is content where the AI is doing the thinking and a human is doing a light edit. The prompt was something like “write a 1,200-word post about managed IT services for the UK market.” The output is a generic, plausibly-worded essay with no specific examples, no genuine point of view and no obvious value to a buyer.

The line between the two is not where the AI is involved. It is where the thinking happens. We are entirely comfortable with AI in the workflow. We are not comfortable with AI doing the editorial work.

Why generated content fails in B2B tech

A generated post on a B2B tech topic fails on three specific counts. None of them are about the writing being detectable as AI. The detectability is a red herring. The failures are about value.

The first is specificity. B2B tech buyers are reading content because they have a specific question. A generated post answers the question abstractly because the model has averaged across a lot of similar content. The generated post on “how to choose an MSP” reads like every other generated post on the same topic, because they are all averaging the same source material. The buyer reads it, learns nothing they did not already know and bounces.

The second is point of view. The buyer is also reading to understand how the vendor thinks. A generated post has no point of view because the model has no stake in the answer. The result is content that is correct, balanced, complete and useless for a buyer trying to differentiate vendors.

The third is evidence. A generated post cannot draw on the customer interviews, internal data and SME conversations that make B2B tech content credible. It can mention them, in a vague way, but it cannot ground itself in them. Without that grounding, the credibility is not there.

These three failures are why generated content does not rank, does not get cited in AI search and does not close deals. Our note on writing content LLMs cite covers the citation angle, and our content KPIs in the AI search era post covers the measurement side.

What AI is genuinely useful for

The places where AI saves us real time, without hollowing out the work:

Interview transcript sorting. A 60-minute interview produces around 8,000 words of transcript. ChatGPT or Claude can extract the strongest quotes, group answers by theme and surface contradictions. The writer is still doing the editorial work, but they start from a sorted base instead of a wall of text.

First-pass outlines for known topics. For a topic the writer has briefed thoroughly, an AI outline is a useful sanity check. It rarely has the right structure on its own, but it surfaces sections the writer might have missed. We treat the AI outline as a critique of our own draft outline, not as the outline itself.

Alternate phrasings for stuck paragraphs. When a paragraph is right in substance but wrong in cadence, asking an AI for three alternative phrasings is faster than rewriting from scratch. The writer picks none of them in full, but pulls a phrase from one and a structure from another. The output is the writer’s, with assistance.

Meta descriptions, alt text and short-form variants. These are the genuinely mechanical writing tasks where AI is reliably useful. A 150-character meta description that captures the substance of a 1,400-word post is the kind of work that benefits from a few AI-generated options to choose between.

Research summarisation. When the writer needs to absorb a long whitepaper, regulatory document or technical reference quickly, AI summarisation is a working tool. The writer still reads the source, but they read it with a map.

In all of these cases, the AI is making the human writer faster at the work the human writer should be doing anyway. That is the line we hold.

What AI should not be doing

The places we keep AI out of, by policy:

Writing the published draft. No published post on a client site is written by AI. The draft is written by a human, even if AI helped with parts of the workflow.

Writing customer quotes or case study narratives. These have to be sourced from a real interview. Generating a plausible-sounding customer quote is fabrication, regardless of whether the customer would technically have agreed with it.

Producing the strategic point of view. Opinion pieces, predictions and frameworks have to come from a person inside the business with a stake in being right. AI cannot have a stake.

Generating long-form pillar pages from scratch. A pillar page is the page that has to demonstrate topical depth. Generated pillar pages are the most obviously hollow form of AI content and they fail to rank for that reason. Our pillar page structure post covers what a pillar page actually needs to do.

Approval of factual claims. Anything that says “according to our experience” or cites a number has to be checked by a human against a real source.

How this affects search visibility

The search engines and AI search engines are increasingly good at identifying generated content, but the more important point is that they do not need to. Generated content fails on its own merits. It does not earn the engagement signals, the backlinks, the citations or the brand searches that build authority. The site that publishes generated content fades from rankings naturally, because the content is not earning attention.

Our notes on AI search optimisation and how LLMs cite sources cover the mechanics of how AI search picks content. The short version is that AI search rewards specific, evidence-grounded, clearly-attributed content. Generated content is none of those things by default.

The exception worth flagging is structured-data content, where the underlying data is real and the AI is helping to format it. A programmatic SEO page generated from a database of real product comparisons is fine. A “blog post” generated from nothing is not. Our programmatic SEO for tech post covers where the boundary sits.

What our editorial policy looks like in practice

For the content programmes we run, we have a short, clearly written policy. It is the kind of document that fits on one page in Notion.

1. AI may be used in research, transcript handling, outline review, phrasing alternatives, meta descriptions and alt text.
2. The published draft is written by a named human writer.
3. Quotes from customers and SMEs come from real interviews, no exceptions.
4. Any AI-assisted output is reviewed by a human editor before publication.
5. The byline is the human writer, not “the team” or an AI tool.

We share this policy with clients and we expect them to share it with us. The teams that have an explicit policy ship better content than the teams that have an informal “we use AI sometimes” approach.

Where the line will move

The line will move. Models are improving and the parts of the production process safe to delegate will widen over the next two years. What is unlikely to move is the requirement that the thinking, the customer relationships and the editorial judgement stay human. Those are the parts of the work that earn the audience’s attention.

If you are working out an AI policy for your content team, drop us a line. Our content marketing service and AI SEO service both cover where the line sits in practice.