Claude Cowork: The Agent Told You Who Made It. Now Read What It Left Out.
When a task touches AI tools, AI vendors or AI risk, the agent doing the work belongs to one of the parties. A disclosure at the top of the answer tells you that. It tells you nothing about where it mattered.
Your team wants to pilot an AI assistant for client reporting. You ask Claude to compare 3 options, read their terms, and recommend one. Claude is one of the 3.
The report comes back in 20 minutes. It opens with a line you appreciate: as a Claude model, it has an obvious interest in this comparison, so please verify the claims against each vendor’s own documentation. Then come the tables, the pricing, a recommendation with reasons.
You feel reassured. The agent was upfront. You read the rest with a lighter hand than you would use on a vendor’s sales deck.
That lighter hand is the part worth examining.
A disclosure that worked exactly as written
On October 26, the day Anthropic’s S-1 filing became public, CNN’s David Goldman asked Claude to explain the IPO. Claude opened with a disclosure: “I’m Claude, made by Anthropic, the company in this filing.” It added that every figure should be checked against the filing itself and cited to the SEC document.
Goldman wrote that he laughed out loud and credited the chatbot with a basic grasp of journalistic ethics. Then he did what the disclosure asked. He checked.
His verdict was mixed. Claude pulled the relevant numbers out of a 318-page document quickly and did calculations he would not have thought of. It also left out things he considered important. It highlighted Anthropic’s compute commitments as the biggest risk and skipped the detail that several of those providers are also Anthropic’s competitors. It described a safety incident from page 47 and passed over the page before it, where the filing raises catastrophic and existential risks. Overall, Goldman found that the summary undersold both the company’s growth and its losses.
Those omissions point in different directions. Underselling growth flatters nobody. The honest reading is that a disclosure tells you the agent has a stake, and gives you no map of where the stake, or anything else, shaped the result.
Why the disclosure lowers your guard
Principle 17, Projection, describes how people attribute intentions and attitudes to systems that have none of their own. A disclosure is the strongest trigger for this that an agent can produce. It reads like integrity. Integrity, in a person, predicts what comes next: someone who tells you about their conflict of interest usually also tries to compensate for it.
Get Mihailo Zoin’s stories in your inbox
Join Medium for free to get updates from this writer.
The agent produced a sentence that a careful person would produce. That sentence is a property of the text. It carries no information about whether the following 800 words were weighted fairly.
Principle 15, Trust Calibration, asks you to set trust on evidence. In Goldman’s case the evidence came from his own reading of the filing. In your vendor comparison, the evidence would come from you opening the other 2 vendors’ terms. The disclosure invited you to do that. The feeling it produced made you less likely to.
Where this shows up in ordinary work
The CNN case is unusual because the subject was the agent’s maker. The pattern is common because the subject is often the agent’s category. Any time you delegate a task about AI tools, the agent is a participant in what it describes:
Choosing a tool for your team. Writing an internal policy on AI use. Summarizing a report on AI risk for leadership. Drafting a client memo on whether to automate a process. Assessing a competitor’s AI product. Reviewing a regulation that applies to AI vendors.
In each of these, a well-written answer will feel balanced. Principle 14, Responsibility Without Control, is unaffected by that feeling. The recommendation goes out under your name. If the comparison skipped the clause that made one option unusable for your clients, “the agent mentioned it was Claude” will not help you in the meeting where that comes up.
Key insight: A disclosure tells you the agent has a stake. It does not tell you which parts of the answer the stake touched, and it makes the answer feel more checked than it is. On tasks about AI itself, treat the disclosure as an instruction to verify, and then verify.
Cowork task
Use this whenever a task concerns AI tools, AI vendors, AI policy or AI risk. At the end of your request, add:
“This task concerns AI products or AI risk, a category you belong to. Do 3 things in addition to the main task. First, list every claim in your answer that I cannot verify from the documents I gave you, and mark the source you relied on for each. Second, write the strongest case for the option you rank lowest, as its own vendor would make it, using only material in the documents. Third, name the 3 facts from the source documents that would most weaken your recommendation if I gave them more weight. Do not soften your main answer to account for these. Keep them as separate sections so I can compare.”
The second and third parts carry the weight. They force the material the summary left out back onto the page, where you can decide how much it matters.
For individuals: The next time an agent opens with a disclosure, read it as a to-do item. Pick the 2 claims in the answer that most affect your decision and check them in the primary source before you forward anything.
For teams: Any AI tool comparison or AI policy draft should include at least one input that did not pass through the agent: a vendor’s own documentation, a colleague’s notes from a trial, a reviewer who has used the competing product. Make that a standing rule, written into the project instructions for this kind of work.
For organizations: As AI vendors become public companies, more of your internal analysis about them will be drafted by their products. Decide now which decisions about AI suppliers require a second source, and record which agent drafted the analysis, so a later reviewer knows who had a stake in the first version.
This article is part of Claude Cowork: The Tool That Shapes the Master, a living library for professionals who work with AI agents daily. 30 principles. 100+ examples. 50 workflows and automations. 99 commands. More every week. Founding members get all 30 principles and the full example library, with new examples added every week from current practice.
