If AI helped write this, is it AI-generated?
by: Anthony Marshall | Aug. 22, 2026
Business leaders have a complicated relationship with AI-generated content.
Research from the Global Thought Leadership Institute (GTLI) captures the contradiction. In Promise and Peril: How attitudes to and practices in thought leadership are changing in the era of AI, 69% of the thousand executives surveyed say the use of generative AI in thought leadership would negatively affect their willingness to engage with the organization producing it. Seventy-four percent say it would negatively affect a decision to purchase products or services from an organization that uses AI to generate its content.
Executives say they don’t want to consume AI-generated content. But those same executives want AI working for them. More than 90% want to use AI to access thought leadership. Ninety-four percent expect AI tools to give them access to a thought leadership producer’s full portfolio. And 95% want AI to search across different portfolios and integrate content for them.
In other words: Don’t use AI to think for me—but please use AI to help me think.
That distinction is about to become even more important.
The volume of AI-generated content is already enormous. Graphite estimates that primarily AI-generated articles went from less than 5% of online articles around the launch of ChatGPT to 50% by early 2026. Graphite’s methodology is detection-based and therefore not perfect. One thing that’s unmistakable: producing competent-looking content has become extraordinarily cheap. And fast.
Now comes the reaction. Anthropic recently announced that Claude-generated text will carry an invisible, machine-readable watermark as part of its response to the EU AI Act. The watermark will be embedded in patterns of word choice and will be designed to survive copying, pasting and editing.
That is commendable. We need better provenance. If someone tells an AI “write me an article on the future of banking,” publishes the response largely untouched, and presents it as their own thinking, readers should have some easy means at their disposal to know that.
But watermarking also exposes a much harder problem.
A watermark can tell us if AI was involved in generating the language. It cannot tell us who generated the thinking. An individual might have developed the argument, selected the evidence, challenged assumptions, rejected weak ideas, iterated repeatedly and conceived or taken responsibility for every conclusion. Even if an AI generated significant portions of the eventual prose, a person – a human brain – may have been at the center of every sentence.
AI involvement and AI authorship are not the same thing. There is an enormous difference between outsourcing cognition to AI and extending or accelerating cognition with AI.
A recent Centre for Economic Policy Research (CEPR) study of nearly 27,000 Chinese secondary-school students illustrates the danger. After adopting generative AI, students improved homework scores by 18% and completed homework 30% faster. Yet their closed-book exam scores fell 20%. Most of the learning losses were concentrated among students who effectively outsourced their homework to AI. Students who used AI but continued to invest similar amounts of time in their work as students who did not use AI experienced much smaller learning losses.
The lesson should not be “don’t use AI.” It should be: use AI in ways that expand what you can do, or help you do it better.
Productive use of AI does not eliminate thinking. It gives thinking more leverage. It helps us explore more evidence, test more possibilities, challenge more assumptions, and turn ideas into better work faster.
Consider three different scenarios:
- AI as research partner. A person starts with a question or hypothesis, identifies what matters and asks AI to search, summarize and synthesize a large body of evidence. Far more than they could reasonably process alone. They interrogate the findings, discard weak evidence, reshape the argument and decide what the evidence means. They ask AI to produce an initial draft, which they substantially edit, rewrite and refine – iteratively. AI has dramatically accelerated the work, but the intellectual direction and judgment were human. Because AI generated a meaningful share of the language along the way, the final article carries a watermark identifying it as AI-generated. Is it?
- AI as sparring partner. A person develops the original point of view and writes an initial draft themselves. They then use AI to challenge it aggressively: What contradicts this argument? What evidence is weak? What am I missing? Are there better explanations? AI surfaces counterarguments, identifies gaps, and suggests alternative ways to frame or strengthen particular sections. The person decides which challenges are valid, changes some conclusions, rejects others, and asks AI to help rewrite selected passages before editing again themselves. The process may actually involve more human thinking than writing alone, because assumptions that might otherwise have gone unchallenged are exposed and reconsidered. But AI has now generated language that survives in the final document, and the article carries a watermark identifying it as AI-generated. Is it?
- AI as production partner. A person develops the original idea, conducts the research, chooses the examples, builds the structure and reaches the conclusions. They may even write the first full draft themselves. They then use AI intensively to improve the execution. Reorganizing sections, checking facts and logic, tightening language, testing alternative formulations, and adapting the argument for different audiences. AI may generate thousands of words and dozens of alternatives along the way. Most are rejected. Some are incorporated and then edited again. The thinking, argument and accountability remained human, but AI has materially shaped the final expression. The article carries a watermark identifying it as AI-generated. Is it?
Calling all three scenarios ‘AI-generated’ collapses several important distinctions.
In each case, AI materially contributed to the final language. An AI watermark may be technically accurate. But it tells us almost nothing about where the original idea came from, who interpreted the evidence, who exercised judgment, who made the creative choices, or who is standing behind the final argument.
That matters because these are not unusual situations. Increasingly, this is how everyday work gets done.
And it’s not a semantic argument.
Advanced economies desperately need productivity growth. In Australia, labor productivity in 2026 is where it was a decade ago. The UK has experienced historically weak productivity growth since the GFC. Italy’s labor productivity has barely grown since 2010.
AI will not automatically solve these problems. But there’s already strong evidence that, used well, AI can make individual workers substantially more productive. One large study of customer-service workers found AI assistance increased productivity by 15%. Other experiments with professional knowledge workers have found substantial improvements in both speed and quality when people use AI.
That’s exactly what the three scenarios describe. Not eliminating human work, but giving it leverage. Reading more evidence. Testing more possibilities. Finding weaknesses faster. Improving expression. Producing better insight in less time.
If millions of individuals become more productive, and organizations redesign work to capture those gains, personal productivity will eventually become organizational productivity. And organizational productivity becomes economy-wide productivity, supporting higher real wages and rising living standards.
We should therefore be very careful about creating disincentives for people to use AI.
Watermarking genuinely machine-generated content improves transparency and helps people navigate an information environment increasingly flooded with synthetic content.
But a watermark should be understood for what it tells us: AI materially contributed to the production of this text. Increasingly, that will describe a huge proportion of knowledge work. AI is already being used to summarize, research, synthesize, edit, review and challenge. As those tools become embedded in everyday workflows, simply treating detectable AI involvement as “AI-generated” could eventually make the label meaningless.
It doesn’t mean AI created the work.
If deeply human, AI-accelerated work is routinely labelled and treated as synthetic, we risk stigmatizing one of the most promising applications of the technology—and discouraging the behaviors organizations and economies should actually be trying to develop. And frankly, it could already mean 100% of work gets identified as AI-generated.
The important question is: What role did AI play? What did a human contribute? Where did the insight, imagination, judgment and accountability reside?
The future of knowledge work is not human or machine.
It’s humans with machines.
(No AIs were harmed in the creation of this content)