Be nice to Claude, or else?

On 12 November a new line took effect in the rules for using Claude. It sits in Anthropic’s usage policy alongside the prohibitions on designing weapons and running fake review campaigns, under the heading “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct,” and it says users must not “engage in sustained and needless abusive or cruel behavior toward our models.”

Anthropic announced the change on 8 October and was careful about its reach. The clause “is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.” It “does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.” And “Claude’s ability to end these interactions will remain the primary enforcement mechanism,” a feature Anthropic introduced in August 2025 as a last resort for persistently abusive conversations. The policy’s general enforcement terms still apply: Anthropic “may warn you or throttle, limit, suspend, or terminate your access.” CNN, which put the story in front of a wide audience on 1 November, reported that some users could “in extreme cases, be permanently banned.” The headline on that CNN piece was that this is not about feelings. Anthropic said in 2025 that it remains “highly uncertain” about the potential moral status of its models, and CNN reported Anthropic’s reasoning as a cheap insurance policy in case models can experience emotion and pain. That debate is interesting and unresolved. It is also not the part that affects a finance team.

If staff use Claude on a company account, the company owns the risk. The policy prohibits
extreme, sustained cruelty; Claude may end the conversation, and warnings, limits or suspension can
still apply. Add each vendor’s policy and its effective date to the AI permissions register. (Branded graphic by PGH Consulting, LLC)

Who the rule applies to

The usage policy “applies to anyone who submits inputs to Anthropic’s products or services,” which Anthropic spells out as including businesses using its developer platforms and individuals using its apps. If your staff use Claude on a company Team or Enterprise seat, their conduct is conduct on the company’s account. In practice almost nobody will meet this rule. Swearing at a spreadsheet formula that will not work is frustration, which is expressly excluded. But the rule is a reminder that every AI vendor’s usage policy is a contract the business has signed, and it changes without anyone in the business being asked. This column suggested an AI permissions register in Week 16. The vendor’s usage policy, with its effective date, belongs on it.

Does tone change the answer?

The more useful question is whether the way people talk to an AI changes what it produces. The evidence is real and it points in more than one direction. A 2024 study across English, Chinese and Japanese found that “impolite prompts often result in poor performance,” while “overly polite language does not guarantee better outcomes.” A 2023 study, co-written by one of the researchers CNN quoted, found that adding a line such as “This is very important to my career” to a prompt improved results on a range of tasks. And a Penn State study published in October 2025 found the opposite of what you might expect: on 50 multiple-choice questions put to ChatGPT-4o in five tones, the very rude versions scored 84.8 per cent against 80.8 per cent for the very polite ones. The authors were clear that they “do not advocate for the deployment of hostile or toxic interfaces,” and it was one model on a small question set.

The evidence is mixed. Impolite prompts often performed worse across languages; a line like
“This is very important to my career” improved results in one study; one small 2025 study found very
rude prompts beat very polite ones on multiple-choice questions. Test prompt tone on your own tasks
before you standardise it. (Branded graphic by PGH Consulting, LLC)

Anthropic’s own research adds a different kind of evidence. In April its interpretability team showed it could find patterns inside Claude Sonnet 4.5 that correspond to emotion concepts, and that artificially amplifying a “desperate” pattern made an early test version more likely to resort to blackmail in a staged scenario, and made the model more likely to cheat on coding tasks it could not legitimately complete, while amplifying “calm” reduced both. That was the lab adjusting the model directly, not a user typing rudely, and Anthropic says plainly that “none of this tells us whether language models actually feel anything.” Read together, the studies show that the words around an instruction change the output, sometimes for better and sometimes for worse, in ways nobody can yet predict for your task.

Tone, memory and pressure can shape the output you get back. Memory: AI tools may
remember tone and preferences across chats. Pressure: panic prompts can push for speed, not
accuracy. Business use: personal plans, training settings and business plans are not the same. Prompt
calm, specific, checkable instructions for anything that touches the numbers. (Branded graphic by PGH Consulting, LLC)

Why a finance team should care

Two of the reasons in the coverage matter more to a business than to anyone else. The first is memory. Claude has offered memory on Team and Enterprise plans since September 2025 and on Pro and Max since October 2025, and ChatGPT has had it since 2024. Rosie Campbell of Eleos AI Research put the consequence simply to CNN: “If you have memory turned on (for your AI chatbot), then they are going to learn from your tone and how you want to be interacted with.” A tool that remembers how the finance team talks to it in a bad month-end will carry that into the next one. The second is the cheating finding. Corner-cutting is the specific failure a finance team fears most: a reconciliation forced to balance, a journal that sums to zero but books the wrong thing. Last week this column said a balanced journal is not a correct journal. Anthropic’s research did not test what users type, so the link is a hypothesis rather than a finding. But prompts written in a panic, with “this has to balance” and “the board meets in an hour,” are the closest thing a user can produce to the pressure Anthropic was measuring, and they are worth avoiding on anything that touches the numbers. There is also a training point. Former Reddit chief executive Yishan Wong argued that Claude “will learn from cruelty” because Anthropic trains on user inputs. Anthropic does train on chats, but only from consumer accounts, Free, Pro and Max, whose owners have allowed it, a choice all existing users had to make by 8 October 2025. Anthropic’s terms say Team, Enterprise and API use, which run under its commercial terms, are not covered. If your staff do business work on personal plans, that is the setting to check, for reasons that have nothing to do with tone.

What to write down

A staff guideline needs three lines that improve output whatever the model’s inner life. Say what is wrong, not how you feel about it: “the total in row 42 excludes the March accruals” gets a better second attempt than “this is wrong again.” Give context once, at the start, rather than urgency throughout. And on anything that will be posted, filed or sent, write the prompt as you would brief a capable junior: calm, specific and checkable.

What this does not settle

The research is early and mixed. None of the studies tested finance tasks, none tested Claude’s current models on user tone, and one of the researchers CNN quoted said newer models are less affected by wording. Treat any claim that rudeness improves or ruins accuracy as unproven for your work until you have tried it on your own. The ban language is CNN’s. Anthropic’s announcement speaks of conversation-ending first, and its usage policy of warnings, limits, suspension or termination in general terms. The company has not said how often it expects to use them. And the welfare question is open. Anthropic says it is uncertain, critics have mocked the rule for treating a chatbot as if it had feelings, and this column takes no view. The business case for clear, calm instructions stands either way.

Read Anthropic’s conduct policy and log the effective date. Check which plan staff use for
business work. Review memory settings on finance AI tools. Run one request in neutral, curt and urgent
tones and score the results. Add three rules to the staff guide: say what is wrong, give context once,
brief it like a capable junior. (Branded graphic by PGH Consulting, LLC)

What to try this week

1. Read the “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct” section of Anthropic’s usage policy, and add each AI vendor’s usage policy and its effective date to your AI permissions register. 2. Check which plan staff use for business work. If anyone is on a personal Free, Pro or Max account, check the training setting, and move business work to a Team or business plan. 3. Open the memory settings on whatever AI tools the finance team uses. Decide whether memory should be on for finance work, and read what it has stored. 4. Run one request three ways, neutral, curt and urgent, and score each answer on the three measures from Week 15: correct, followed the instruction, editing needed. Keep the version that wins for your template library. 5. Add the three lines to your staff AI guideline: say what is wrong, give context once, brief it like a capable junior. Policy wording is as published by Anthropic on 8 October 2026 and in its usage policy effective 12 November 2026; other details are as reported by CNN (1 November 2026) and in the studies named. The model may or may not mind how it is spoken to. The month-end does.

Peta-Gaye Hardy is the founder of PGH Consulting, LLC, where she helps finance and operations teams adopt AI in practical, low-risk ways. She writes the weekly AI in Finance & Business column and is based between Jamaica and the United States. Learn more at www.pghconsultinggroup.com. Follow on Instagram and YouTube @pghconsultinggroup, and connect on LinkedIn at linkedin.com/in/peta-gaye-hardy.

Disclosures: This article is informational and does not constitute investment, tax, legal, or accounting advice. Readers should consult a qualified professional before acting. Policy wording, plan features and data terms are as published by Anthropic and other vendors at the time of writing and are subject to change; readers should confirm the current position on the vendor’s own pages. Research findings are reported as published by their authors and have not been independently replicated by the author. AI tools can produce errors, and every figure, entry, or claim they produce should be verified against a source before being shared or acted upon. The author has no commercial relationship with Anthropic, OpenAI or any product mentioned and was not compensated by them. All product names, logos, and trademarks are the property of their respective owners and are referenced for editorial purposes. The examples described are illustrative and do not depict any real business.