At the end of July, a debut crime novel became the most expensive casualty so far in the publishing industry’s war over AI. Fourteen publishing houses had bid for the manuscript, with the winning offer delivering a $2 million contract for the author. But when rumors began to circulate that the book had been written with the help of artificial intelligence, his own agents withdrew the book.
The author denies using AI and has pointed out that the manuscript had been read by editors and acquisitions teams across the industry. What they saw on the page excited them enough to open their wallets. But while the words and sentences remained the same, the emergence of concerns about their source was enough to kill the sale, and possibly the author’s future career.
It was at least the third such major scandal this year. The ferocity of the public response to these stories is understandable. Part of what a reader buys when they purchase a novel is the human being behind it. The experience of engaging with a work of art is often anchored in a meeting of minds, so the idea that AI contributed to the output can feel like a betrayal.
But these legitimate concerns risk fueling a more general, and far less well-grounded, presumption that any AI involvement in any output that humans once produced is a form of cheating. The result is a strange cultural moment: a tool we are constantly told we must learn to use is becoming something people fear being caught using.
{“blockType”:”mv-promo-block”,”data”:{“imageDesktopUrl”:”https://images.fastcompany.com/image/upload/f_webp,q_auto,c_fit/wp-cms-2/2025/10/creator-faisalhoque.png”,”imageMobileUrl”:”https://images.fastcompany.com/image/upload/f_webp,q_auto,c_fit/wp-cms-2/2025/10/faisal-hoque.png”,”eyebrow”:””,”headline”:”Ready to thrive at the intersection of business, technology, and humanity? “,”dek”:”Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.”,”subhed”:””,”description”:””,”ctaText”:”Learn More”,”ctaUrl”:”https://faisalhoque.com”,”theme”:{“bg”:”#02263c”,”text”:”#ffffff”,”eyebrow”:”#9aa2aa”,”subhed”:”#ffffff”,”buttonBg”:”#ffffff”,”buttonHoverBg”:”#3b3f46″,”buttonText”:”#000000″},”imageDesktopId”:91420512,”imageMobileId”:91420514,”shareable”:false,”slug”:””,”wpCssClasses”:””}}
When Anthropic announced last month that Claude will weave an invisible watermark into the text it processes, the public response reflected deep divisions. Some people celebrated: “The only reason you wouldn’t want this is to lie to people,” as one commenter put it. Others canceled their subscriptions, fearing that anyone who lets AI touch their writing can now be branded with a modern scarlet letter.
Employees are left in a double bind. They can avoid using valuable productivity tools and risk criticism from pro-AI bosses and colleagues. Or they can use them and worry that they will draw the ire of the crowd. Business leaders must steer a careful path through this difficult landscape. The first step is to set aside the question “Did you use AI?” Instead, we need to focus on a subtly but importantly different criterion: “What did you use AI for?”
Why care who produced the work?
There are three distinct things you might want to know about a given piece of work in a business context:
- Is the output good?
- What does the output tell us about the capability of the person who produced it?
- What does the process of producing it do for that person’s capabilities?
Historically, we often answered all three by looking at the same piece of work. If a junior analyst produced a brilliant report, we saw that the report was good, we could infer from its quality that the analyst was good, and it was reasonable to think that doing the work was helping the analyst develop their core skills.
Generative AI shows us that the connection between those three questions can break. A report can be excellent without the analyst having done anything apart from prompting a large language model (LLM). And in such a case, producing an excellent report does not help the analyst become a better analyst; at most, it helps them develop their ability to prompt LLMs.
Given this broken connection, it is a mistake to use provenance as a proxy for all three questions. Instead, leaders should follow a more nuanced and fine-grained approach, in which the purpose of the work plays a crucial role in deciding how AI should be involved in doing it.
Production: Judge the work
When the purpose is purely production, the question is simple: Is the work good? So, whatever the context, business leaders should define the criteria for what counts as good work and then judge the output against those standards.
A corollary of this is that the method of production is irrelevant. The employee could have used Claude or the ancient art of divination to produce the output—it doesn’t matter. The only thing that matters is whether the work meets the required standard.
But indifference to method cuts both ways: If using the tool earns no penalty, it also offers no shelter from responsibility. Once an employee has checked and submitted a piece of work, it is theirs. “The AI did it” does not excuse an error, and “I did it by hand” does not excuse mediocre work.
So the rule follows directly: For production work, judge the output and enforce ownership.
Assessment: Test the person
Sometimes, however, it is not the quality of the work that is ultimately important; rather, what matters is what the output says about the capabilities of the person producing it. Obvious examples of such moments are the job interview and the promotion discussion.
Here, provenance begins to become more relevant because for the output to serve as evidence of capability, there must be some meaningful connection between the person and the output. But instead of simply asking whether AI did it, it is much more informative to ask the person to talk you through the output. Why did they choose this structure? Why was this evidence important? What are the weak points of the argument?
This is particularly important given the current limitations of AI detection. Anthropic explicitly says that its watermark “can only determine that Claude was likely involved with the content at some point.” The extent and nature of the involvement is open—but it is precisely this that needs to be understood if we are using output to assess the capabilities of the person who produced it.
Of course, sometimes it may be important to know whether someone can perform a task without assistance. But then that should just be tested directly, instead of turning every ordinary deliverable into a purity test.
Development: Protect the learning
Learning by doing has always been a very important part of development, and efficiency is often legitimately sacrificed to the goal of skill development. Even if it costs the firm more, junior employees will be given some work precisely because doing it develops the expertise needed for more senior work later.
But if the junior employee is simply palming off the tasks to AI, the work loses its development function. For example, in a randomized study of developers learning an unfamiliar Python library, participants given AI assistance scored 17% lower on a subsequent test of conceptual understanding, code reading, and debugging. Companies may get better output today, but they will be weakening the supply of people capable of using that output tomorrow.
Further, the study found that participants who remained cognitively engaged—asking conceptual questions or requesting explanations of generated code—preserved much more of the learning. This means that it is not as simple as saying that using AI to produce output is bad for development—much depends on how the AI was used.
Therefore, leaders need to be intentional about how they develop people through work and with the use of AI. The only constant here is that development work should contain deliberate friction, because that is required for learning. Beyond that, there is no single right answer: Sometimes the employee should use AI freely. Sometimes AI should explain but not solve. Sometimes a first attempt should be unaided. The right constraint depends on the capability you are trying to build.
4 things to do instead of making AI use a purity test
Here are four things leaders can do right now to develop an AI policy that helps both the organization and its employees:
- Classify the work. Before setting an AI rule, ask whether the primary purpose is production, assessment, or development. Many jobs contain all three, but individual tasks usually lean toward one.
- Match the rule to the purpose. Production work should normally optimize for quality and accountability. Assessment should expose the capability being assessed. Development should preserve enough human practice to build the capability you will need later.
- Make people defend important work. Do this routinely, not as an AI interrogation. If employees regularly explain their reasoning, managers can observe judgment without turning provenance into a verdict.
- Treat detection as evidence, not judgment. A watermark may tell you that an AI system touched the work. It cannot tell you whether the work is good, whether the employee understands it, or whether using AI helped or harmed their development.
Focus on the things that matter
A business runs on quality, capability, and learning—and provenance is a weak proxy for all three. Ask what the AI was used for, and judge accordingly. Leave questions of authenticity to artists and their patrons.
{“blockType”:”mv-promo-block”,”data”:{“imageDesktopUrl”:”https://images.fastcompany.com/image/upload/f_webp,q_auto,c_fit/wp-cms-2/2025/10/creator-faisalhoque.png”,”imageMobileUrl”:”https://images.fastcompany.com/image/upload/f_webp,q_auto,c_fit/wp-cms-2/2025/10/faisal-hoque.png”,”eyebrow”:””,”headline”:”Ready to thrive at the intersection of business, technology, and humanity? “,”dek”:”Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.”,”subhed”:””,”description”:””,”ctaText”:”Learn More”,”ctaUrl”:”https://faisalhoque.com”,”theme”:{“bg”:”#02263c”,”text”:”#ffffff”,”eyebrow”:”#9aa2aa”,”subhed”:”#ffffff”,”buttonBg”:”#ffffff”,”buttonHoverBg”:”#3b3f46″,”buttonText”:”#000000″},”imageDesktopId”:91420512,”imageMobileId”:91420514,”shareable”:false,”slug”:””,”wpCssClasses”:””}}