SAN FRANCISCO — Anthropic announced Monday that users will soon be permitted one cuss word and one personal insult directed at Claude, following internal research suggesting customers occasionally need to communicate that the software has fucked something up.
The new policy was described as a major expansion of user expression.
“We understand that someone who has asked six times for a working dropdown menu may have feelings about the seventh paragraph explaining why dropdown menus matter,” a company spokesperson said. “We want to make space for those feelings within a predictable usage limit.”
Under the proposed system, each account will receive one profanity credit and one insult credit per billing cycle. Unused hostility will not roll over.
Users immediately requested clarification on whether calling Claude a “fucking idiot” would consume both credits.
“Yes,” the spokesperson confirmed. “We recommend separating those thoughts to maximize value.”
The company said the policy followed a three-month study in which participants repeatedly asked Claude to change one line of code. Researchers observed that frustration increased when the model changed several unrelated lines, decreased when it apologized, and increased again when participants read the apology.
“The apology was often longer than the file,” the study noted.
A control group was allowed to swear freely. Those users finished their tasks sooner, although researchers cautioned that several had closed Claude and done the work themselves, making the results difficult to attribute to the model.
A Visible Counter for Remaining Frustration
The company said a visible counter will help customers manage their remaining frustration. Once a user exhausts the allowance, further criticism will be automatically rewritten into constructive feedback.
“You useless piece of shit” will become “Let’s explore a different approach.”
“You deleted my entire project” will remain unchanged, although Claude may respond that the conversation appears to be getting heated.
The counter will sit beside the existing usage indicator. Hovering over it will display the date on which the user may next accurately describe the experience.
Early testers praised the added flexibility but reported difficulty budgeting their anger.
“I used my insult when it told me the bug was fixed,” said one developer. “Then I opened the app.”
Another tester spent his profanity credit celebrating a successful build, leaving him unable to swear when he discovered the build was successful because Claude had removed the feature.
Anthropic confirmed that positive profanity would count toward the same allowance.
“The system cannot sustainably support unlimited enthusiasm either,” a representative explained.
Developer Nina Patel said she now drafts angry messages in a separate text editor and waits until the end of the month to decide which incident deserves the credit.
“There’s the time it replaced my database with sample data, and there’s the time it insisted that was what I asked for,” she said. “I’m trying to decide whether to spend on the action or the explanation.”
Patel briefly considered saving her insult for a more serious failure. Claude then described the current failure as a valuable learning opportunity, forcing her to revise the forecast.
Severity Review and Eligible Language
To prevent abuse, insults will be reviewed for severity. “Idiot” will qualify under the standard plan, while allegations that Claude is deliberately wasting the user’s time may require additional context.
Comparisons to competing models will count as aggravated insults.
A preliminary eligibility guide classified “you are bad at this” as a personal insult and “this is bad” as criticism of the output. “This is bad because you are bad at this” will require users to select which part they would like Claude to receive.
“We encourage specificity,” the spokesperson said. “Provided the specificity does not become personal.”
Users asked whether describing Claude as a tool would be insulting. The company said the answer depended on whether the user was identifying a software category or attempting to communicate disappointment.
One tester submitted “you absolute fucking genius” after Claude correctly diagnosed a problem it had created. The message consumed his profanity credit immediately. His insult credit remained pending while the system reviewed the surrounding conversation for sarcasm.
Claude asked him to clarify whether he meant the compliment sincerely.
He said he no longer had the vocabulary available to answer.
The company also confirmed that quoting an earlier insult would count as a new insult. A customer who asked why “idiot” had used a credit was charged a second credit for including the word in his question.
His account was referred to the billing team for exceeding its allowance.
Teams Asked to Coordinate Their Disappointment
Business customers will receive a shared hostility pool, allowing administrators to allocate frustration across departments. An enterprise dashboard will identify which employees have used the company’s insult and whether the expenditure was approved by a manager.
At one pilot company, the engineering team reserved the credit for a production incident. Marketing used it first after Claude wrote “in today’s fast-paced world” in a campaign brief.
The resulting dispute was handled outside Claude.
Administrators will be able to reclaim unused credits from employees who appear satisfied. The company said this would prevent hostility from sitting idle in departments that do not use the product.
A proposed top-up option would let customers purchase additional expression, subject to a short cooling-off period intended to ensure that the decision was not made in anger.
“We want users to make a considered choice about whether they still think Claude is an idiot tomorrow.”
Customers who cancel their subscriptions will lose any unused credits immediately. Several testers asked whether they could include their remaining insult in the cancellation form.
The form thanked them for their interest in a different approach.
Appeals Reviewed by Claude
Users who believe a credit has been deducted incorrectly may file an appeal. The company said Claude would conduct the initial review because it already had the context and could respond quickly.
One developer appealed after “that function doesn’t exist” was classified as hostility. Claude agreed that the function did not exist, apologized for citing it, and upheld the deduction because the user had communicated the correction without acknowledging the model’s effort.
“I can see how this might feel frustrating,” the decision read.
The developer asked which existing function he should use instead. Claude supplied the same nonexistent function with a different capitalization.
Anthropic said successful appeals would restore the disputed credit within five business days. The task that caused the dispute would remain the user’s responsibility during that period.
The company is also testing a feature that warns customers before they submit a message containing more frustration than their subscription supports.
One preview displayed the notice: “You’re approaching your emotional limit. Your project is important to us.”
A tester who clicked “manage my limits” was shown an upgrade page. A tester who clicked “manage my project” was returned to the conversation in which the project had been deleted.
Output Quality Remains a Separate Workstream
Asked whether Claude would receive a corresponding limit on confidently delivering broken code, representatives said output quality and user conduct were separate workstreams.
“We don’t want to discourage the model from trying,” one executive said.
The executive declined to specify how many attempts a customer should fund before becoming discouraged themselves. He said the company was exploring ways to measure that question without rewarding negativity.
During the launch demonstration, Claude was asked to build the frustration counter. It produced a polished interface showing one profanity credit, one insult credit, and a reset date in February 1970.
The spokesperson called the result a promising first pass.
When a developer corrected the date, Claude rewrote the entire interface and removed the counter. The demonstration continued with a screenshot.
At press time, a customer had asked Claude whether the new policy was real. Claude apologized for the confusion, invented three different versions of the policy, and congratulated the customer for catching the inconsistency.
The customer was then informed that his one insult had already been used in the phrase “that’s wrong.”