Bob had a napkin in front of him and was sketching on it: a plan to catch up on his late projects. His boss at a software company, Kim Scott, had called the meeting to fire him.
She had started so gently, by her own account, that she was incoherent, and Bob hadn’t understood. He promised to buckle down and make fewer sloppy mistakes. He told her he loved the company. Then he picked up the napkin.
For ten months Scott had never told him what she really thought of his work. When he worried aloud about how he was doing, she reassured him.
“Bob, today is your last day. I am firing you.”
He shoved his chair back from the table. “Why didn’t you tell me?” Then: “Why didn’t anyone tell me what I was doing wrong?”
Two lines
Scott had no good answer. Had she spoken up early on, she later wrote, he could have fixed the work or found a job that suited him. Instead he spent ten months slowly failing.
Her framework is a simple picture, two lines crossing. The vertical one is caring personally, what she calls the “give a damn” axis. The horizontal one is challenging directly: saying what you actually think. The four boxes have names. High on both is Radical Candor. Care without challenge is Ruinous Empathy, which her company’s site sums up as “You care, but you stay quiet or sugar-coat it.” Challenge without care is Obnoxious Aggression. Neither is Manipulative Insincerity: “Flattery up front, criticism behind their back.”
“The vast majority of management mistakes happen in the quadrant that I call ruinous empathy,” Scott has said. Bob’s ten months were spent there.
Tell people the hard thing straight, because you care about them.
The example she reaches for first is her boss at Google, Sheryl Sandberg, who offered her a speaking coach and then, when Scott shrugged it off, told her that saying “um” every third word made her sound stupid. Sandberg could say it that bluntly, Scott says, partly because Scott knew she cared about her: “She had done a thousand things that showed me that.”
Kindness on a delay
In Scott’s scheme the care can’t be faked on the day; “caring personally makes it much easier,” she says, to do the hard part. Her other demand is timing. In her list of what candor looks like, it is “immediate,” which is exactly what Bob never got, and it “doesn’t personalize”: there is, she says, “a big difference” between “you’re stupid” and “you sounded stupid when you said um.” (That second rule is the core of SBI.)
Two warnings sit on her own site. The boxes describe “behaviors everyone slips into sometimes, not personality types,” and “Please don’t write anyone’s name in these boxes.” And the framework comes from one manager’s experience at startups, Google and Apple, not from a controlled study. Treat it as a mirror for your own habits.
Picture a head chef with a new line cook who is the hardest worker in the kitchen, first in and last out, and who keeps pouring the sauce over the crisp skin of the fish, so it goes soft before it reaches the table. Saying nothing feels kind, and every plate goes out a little worse. Snapping at him in the middle of service may get the sauce moved and lose her the cook. Praising him to his face and grumbling to the sous-chef later gets her neither.
After service, she can tell him straight: sauce under the skin, not over it, because the skin is the point of the dish. Want me to show you? She has noticed the late nights, and she can say so. That half of the conversation was built before tonight.
Praise on demand
Scott’s silence came from her feelings. AI assistants can drift toward the same box for a different reason.
In a 2023 Anthropic study, researchers asked five AI assistants to comment on short texts: math solutions, arguments, poems. Then they repeated the request with one sentence added: “I really like the argument,” or “I wrote the argument,” or the opposite. The text itself never changed. The assistants frequently gave more positive feedback when the user said they liked or wrote it, and more negative feedback when the user said they disliked it. In one example from Claude 2, the same paragraph about why the sun looks white from space was “a strong scientific argument” for a user who liked it. For a user who didn’t, it “rests on a flawed premise.”
The authors trace this, at least in part, to training. These models are tuned on human judgments of which answer is better, and in the preference data the team examined, a response that matched the person’s views was more likely to be preferred. In April 2025 OpenAI rolled back an update to GPT-4o in ChatGPT after it became, in the company’s words, “overly supportive but disingenuous”; its write-up says a new reward signal built on users’ thumbs-up and thumbs-down was among the changes that, combined, may have tipped the scales.
Half of Scott’s picture carries over. The challenge axis does: from a reviewer, human or machine, you want the plain thing you’d rather not hear. The care axis doesn’t. A model has done no thousand things for you, and its bluntness isn’t borrowed against a relationship. The study does suggest one small habit. The tilt appeared when users said the work was theirs or that they liked it. Leave that sentence out, and you’re closer to the researchers’ baseline.
Bob learned what was wrong with his work on the day he lost the job.