Clay, a software company, made a rule official across its whole organization, and the rule is that you have to stand behind every sentence you share. If a reviewer questions one of your lines, “the AI wrote that part” is not an answer you are allowed to give. What makes it worth reading is that it governs the writer rather than the tool, which is the opposite of how most companies have tried to write this policy.
What the policy says
Sophie Alpert, an engineer at Clay, wrote it for her own team and published it. Enough other groups inside the company asked to be included that co-founder and chief operating officer Varun Anand made it official company-wide.
It permits more than the phrase “AI policy” suggests. You can brainstorm with a model, draft with one, and proofread with one, and you can paste raw model output into a shared document as long as you label it as raw model output. Nothing about the tooling is forbidden.
What it requires is ownership at the level of the sentence. Alpert’s four principles are that you stand behind every idea and sentence you communicate, that you treat writing as a way of thinking rather than a way of avoiding it, that you spend more time writing than your readers will spend reading, and that you favor being concise over being long.
Her argument for the whole thing sits in one line. “There are no lossless transformations of natural-language text — every rewrite and rephrase changes the meaning of your writing.”
Why a rule about sentences works better than a rule about tools
Most AI policies written so far try to draw a line around the software, which produces rules like “no generative AI in customer-facing copy” or “AI permitted for internal drafts only.” Those rules share a weakness, since they are unenforceable and everyone knows it. Nobody can tell from a finished paragraph which tool touched it, so a tool-based rule becomes an honor system that punishes the people who disclose and rewards the ones who do not.
A rule about accountability sidesteps the detection problem entirely. It does not matter how a sentence was produced, because the test is applied afterward, in a room, by a person asking about a specific line. Either the writer can defend that line or the writer cannot, and that is observable in a way that provenance is not.
The second thing the rule does is quieter. When you know you will have to defend any sentence in the document, you read your own draft differently before you send it, and the reading is where the sloppy claim gets caught. The policy changes behavior at the point of writing rather than at the point of review.
What this changes if you hire writers rather than employ them
Clay wrote this for employees, and the interesting part is what happens when you apply the same test to a vendor, because the outsourced version of the problem is worse in a specific way. An employee who cannot defend a sentence is in the room and has to sit with it. A contractor who cannot defend a sentence has already invoiced.
So the question to put to any writing service is whether a person will defend a specific line you point at, and the follow-up question, which is the one that separates vendors, is who that person is. A service that rotates writers by availability cannot answer it, because the writer who produced the line has moved on to another account by the time you have questions about it. A service that assigns writers and keeps them there can name somebody.
Ask a third question about where the model sits in the process. There is a real difference between a service that uses AI for research, fact-checking, and scoring, and a service that uses AI to produce the draft and puts a human on review. Both will tell you humans are involved. Only one of them has a person who made the argument to begin with, and reviewing a paragraph is a much shallower kind of attention than writing one.
We land on the same side of that line, which is why our own writers draft everything and the models do research, fact-checking, and SEO scoring instead.
Writing the version of this policy your team can use
You do not need Clay’s document. The useful parts compress into four questions any team can sit down and answer.
Can everyone here defend every sentence they send? If the answer is no anywhere, that is the gap the policy closes, and it does not require you to ban anything.
Does your review process ask about specific lines? A policy about defending sentences only works when somebody occasionally asks. Reviews that approve documents wholesale never test it.
Are your documents getting longer? Length is the visible symptom of generated drafting, since the cost of producing another paragraph fell to nothing while the cost of reading one did not. A rule that you spend more time writing than your readers spend reading pushes back on that directly.
What happens when somebody says the AI wrote it? In your company, is that an acceptable answer? Whatever the answer is in practice, that is your working policy, whether or not you have written one down.
Questions teams ask about AI writing policies
Does an AI writing policy have to ban anything? No, and Clay’s does not. It permits brainstorming, drafting, proofreading, and even pasting raw model output when that output is labeled. The obligation it adds is that a person owns the result.
How do you enforce a policy you cannot detect? By moving the test off the artifact and onto the author, since asking a writer to defend a line works regardless of how the line was produced. Detection tools try to answer a question the policy does not need to ask.
Is AI-assisted writing a problem for search or AI visibility? The measurable problem is quality and sameness rather than provenance, because a page assembled from the same sources as everyone else’s page has nothing distinguishing for a search engine or an answer engine to prefer. A person with a position produces something quotable, and that is a content problem rather than a compliance one.
Should a policy like this cover outside vendors? That is where it earns the most, since your own team shares an office and a reputation while a vendor shares neither. Put the standard in the brief and ask the question during the engagement rather than after a campaign underperforms.
Who should write the policy? Somebody who writes. Alpert wrote Clay’s for her own engineering team before it spread, which is a better path than a policy handed down by people who will never be subject to it.
Want writing somebody will stand behind?
Every piece we deliver is drafted by a human writer and read by an editor before it reaches you, so there is always a person who can answer for any line you question. Have a look at what we do or book a demo and meet the writer before you commit to anything.
Explore our writing services
Content Services →
