Stop the Copy-Paste: Manifesto

This could be a message in Slack (or Teams), a messenger, or an email. Maybe you were tagged in Notion or in an Office document.

You open it and see several paragraphs of text, or perhaps a list with all the signs of an AI-generated message: headings, lots of formatting, phrases like "this is not X, it's Y," heavily sprinkled with long dashes.

The person who sent you this message probably spent about ten seconds on it. They asked a chatbot to create the text and simply forwarded the response to you in full, without checking or analyzing it.

But now you have to read and check this wall of text, and then somehow respond to it. When you receive a text that even the sender isn't interested in, it forces you to do work they didn’t want to do. Such asymmetry in workload makes the situation offensive.

Several examples

Negative example❌

#strategy

The company is in the process of discussing the justification for expanding into the German market.

👤:

I asked Claude about this! Here’s what he replied:

Expanding into the German market provides new opportunities and presents significant challenges. Germany is the largest economy in the EU, with a solid consumer base. However, companies should consider the following key factors:

1. Regulatory Compliance

Germany has strict regulations, including the need to comply with GDPR, local labor laws...

[The next six sections follow. None of them mention your company, its product, budget, or references to the ongoing discussion.]

This kind of response overloads a live discussion, it was not requested and is too generalized.

Type: Overzealous

The participant wants to contribute to the discussion, so they asked the chatbot to respond, and then simply copied its answer. They did this with good intentions, sincerely trying to help, but the pile of AI-generated text disrupted the ongoing conversation. Now, other participants have to scroll to continue the conversation or stop to read and evaluate the answer.

The author thought that sending the message would help, but it requires effort on the recipient's part.

Negative example ❌

You:

Does anyone understand why the email open rate has decreased? Last month we changed the subject line.

👤:

ChatGPT said the following:

The decrease in email open rate may be related to a number of factors. This includes changes in the subject line strategy, a decrease in sender reputation, issues with list hygiene, and deliverability. Here are the most common reasons:

1. Subject line fatigue

If the subject lines have become too monotonous or predictable, they may stop attracting subscribers...

[Five more sections with standard email open rate diagnostics. They do not mention your subject line change, audience, or platform.]

Such a response is not relevant to the specific question and is too stereotypical.

Profile: AI Oracle

Someone asks a specific question. Another person sends it to a chatbot and pastes the generated response as the answer.

"ChatGPT said the following" — this is a degraded equivalent of LMGTFY (let me Google that for you) in the age of LLM. As a link or GIF, the LMGTFY phenomenon is easy to ignore, and it clearly conveys its meaning (a sarcastic comment). Copy-pasting is neither of these. The recipients have to figure out whether the response was generated by AI, whether it is correct, and which part of it answers the question (if it is even relevant). When you ask someone a question, you expect to hear their point of view and experience. In this sense, both LMGTFY and copy-pasting are etiquette violations, where the sender does not give the recipient the right to a simple human answer.

Negative example ❌

👤:

Hello team, this week I conducted research on our competitors. Here’s a summary:

Competitor Overview

The market is highly competitive, with many established players and new challenges emerging. Our key competitors are making the following offers in various price categories...

[The following is a five-page essay with careless assumptions and lack of concrete details. No dates, no sources, no current prices.]

The response is presented as personal work; hallucinatory details are possible.

Type: Literary Ghostwriter

The sender presents AI-generated results as their own work, without indicating that a chatbot generated them. Recipients have no reason to doubt this and may begin acting based on outdated, incomplete, or simply incorrect information.

Using AI as a literary ghostwriter puts trust in the sender at risk. If the content turns out to be incorrect, that trust decreases.

Why This Is Offensive

Sending AI-generated text without editing is like consuming junk food: it feels pleasant and easy, but it’s not in your true interest. This negatively affects your relationship with the recipient and does you a disservice by lowering your own level of understanding.

“Since ancient times, creating texts has been more costly than reading them. If a person saw a written text, they could be sure that, at least, another person had spent some time writing it down. The text itself was proof of thought, the basic unit of humanity.”

— Alex Marcinovich, It's rude to show AI output to people

Before LLM, writing required effort. Authors spent time and energy choosing words; their time and effort were balanced by the work their audience put into reading. Due to LLM, this balance has been disrupted; creating text now costs almost nothing for the "author," but the effort needed to read the text remains unchanged. The increase in verbosity from LLM further tips the scale. In some cases (for example, when inserting raw results from LLM into a chat), copy-pasting essentially becomes sabotage, clogging up a real conversation and blocking the chat window.

“Cognitive effort (and even the painful attempts to get out of a deadlock) is probably important for mastering a skill.”

— Anthropic, How AI assistance impacts the formation of coding skills

Writing is thinking. The process of writing forces the author to work through their thoughts, improving their understanding of the subject and concentration. Many studies have shown that delegating tasks to LLM increases cognitive debt. Replacing thinking with LLM ultimately reduces understanding and retention of the subject matter.

“The ideal AI response seems dismissive, even if its content is correct.”

— Blake Stockton, AI Writing Etiquette Manifesto

Before the emergence of LLMs, texts were trusted by default. Authors wrote them based on their own experience and perspective, and readers could judge the author's understanding of the subject based on the coherence of their writings. LLMs generate the next most likely token, with their main goal being to be useful, which explains their tendency to hallucinate (confabulations), as well as why many people consider LLMs to be nonsense generators. Modern LLMs usually have tools that allow them to find information supporting their answers, which reduces the likelihood of fabricating facts in responses (but does not eliminate it entirely). However, this still does not solve the problem of trust; the reader still has no way of knowing what the sender checked in the text, or whether they checked it at all. Therefore, LLM responses cannot be trusted by default; moreover, they increase the asymmetry of effort, requiring the reader to verify them.

Additionally, LLMs write texts with the confidence and authority of a specialist. This adds further uncertainty: the reader cannot assess the true level of the sender’s expertise on the topic. As a result, trust is further undermined, because the AI’s voice destroys the signal that recipients previously used to distinguish experience from confidently sounding nonsense.

“It seems to me offensive to publish texts that I haven’t even read myself. I will never publish a text that requires more time to read than it took me to write it.”

— Simon Willison, Personal AI Ethics

Previously, the principle was "trust, but verify." Readers trusted until their trust was undermined; the author was either worthy or unworthy of trust. Now, however, sending LLM outputs makes the chain of trust uncertain. Did the prompt author check the LLM’s answer diligently enough? Who is to blame when problems or errors are found: the prompt author or the AI? Was it an oversight, a missed verification step, or a complete lack of checks? Due to this uncertainty, the recipient does not know what they can trust, what has been checked, and what has not; they are forced to treat all texts with suspicion. As in the tale "The Boy Who Cried Wolf," once trust is lost, uncertainty spreads to all future messages from the sender.

The assumption of balanced effort and the presumption of trust are no longer guaranteed. Slopypasta creates a growing cycle of negative feedback, in which the sender deprives themselves of learning and authority, and the recipient expends effort and loses trust. Receiving raw AI output feels like something bad due to the cognitive dissonance caused by the violation of these assumptions.

Simple tips to improve the situation

Read.

Read the LLM output before forwarding it. If you don’t read it, you won’t know its accuracy, relevance, or timeliness.

Delegating work to AI creates cognitive debt. Working with the results helps evaluate your own understanding.

Verify.

Check facts before sending them. Everything you send carries your implicit approval — your reputation depends on managing the quality of your texts.

LLMs are trained to "be helpful," so in response to your queries, they produce outdated facts, incorrect numbers, and plausible-sounding nonsense. Moreover, LLMs are inherently not up-to-date: the cutoff of their knowledge at best corresponds to the information about the state of the world that existed at the start of training (several months ago).

Filter.

Limit responses to what is important. Filtering the generated output down to its most useful essence is your responsibility.

LLM has the motivation to use many words, even when it could be brief: models with API payment have the incentive to train chatty LLMs that consume many tokens; moreover, as the research shows, users often prefer longer posts with sophisticated formatting as more engaging.

Admit it.

Share how AI has helped you.

If you have read, checked, and edited the text, then submit it as your own, but preferably mention that AI helped you with the work. If you publish raw output, be upfront about it. In both cases, it may be useful to show your prompt and demonstrate how you worked with AI to get the final result.

Admitting this helps restore the trust signals that are undermined by sloppy copying and informs the recipient of what exactly you’ve checked, as well as warns them of potential issues.

Send AI-generated texts only on request.

Never dump unsolicited AI-generated results into a conversation.

Remember that AI-generated texts create an asymmetry of labor, so respect those you share them with. Sloppy copying shifts the entire workload of reading, verifying, and filtering onto the recipient, who didn’t ask for it and may not be aware of the effort required.

Share as a link.

Share AI texts as links or attached documents, rather than pasting all the text.

In messaging environments, large amounts of pasted text block the whole window and push out the main conversation. A link allows the recipient to choose when to engage in the discussion (and whether it's necessary).

AI’s capabilities are expanding, and they are becoming increasingly useful for drafting, brainstorming, or speeding up tasks. However, using AI shouldn’t turn your productivity into a burden for someone else. New tools require new etiquette.