Almost every manager believes they give good feedback. Almost every employee can recall a piece of feedback that landed like a slap and changed nothing. Both can be true at once, and the gap between them is where careers stall and good people quietly leave. The skill isn't being "honest" or being "kind." It's knowing what kind of feedback you're actually giving, and earning the right to be heard.

The quick version

  • Feedback is not automatically helpful. A landmark meta-analysis found that in more than a third of cases, feedback made performance worse, usually when it aimed at the person rather than the task.
  • Most confusion comes from mixing up three different things people call "feedback": appreciation (you matter), coaching (here's how to improve), and evaluation (here's where you stand). Naming which one you're giving prevents most fights.
  • The best givers do two things at once, care personally and challenge directly. Soften too much and the message never arrives; challenge without warmth and it gets rejected.
  • Receiving well is a separate skill. The instinct to defend or dismiss has predictable triggers, and you can learn to spot them in the moment rather than be run by them.

The idea in depth

The uncomfortable starting point is that feedback doesn't reliably work. In 1996, Avraham Kluger and Angelo DeNisi published a meta-analysis in Psychological Bulletin, "The effects of feedback interventions on performance", pooling 607 effect sizes across more than 23,000 observations. On average, feedback did improve performance. But the headline that should keep every manager honest is the variance underneath the average: in over a third of cases, the feedback intervention actually reduced performance. Their explanation has held up well, when feedback pulls a person's attention onto themselves (their ego, their standing, their worth) rather than onto the task, the energy that should go into the work goes into self-defence instead.

Which points to one rule: keep feedback aimed at behaviour and its effects, not at character. "The report went out with three errors and the client noticed" is a task. "You're careless" is a verdict on the person, and a verdict is exactly what flips attention inward. The reframe is simple to say and hard to hold under stress: describe what happened and what it caused, then talk about what to do differently. You are giving them somewhere to put the energy.

An honest limitation. Kluger and DeNisi's work is about feedback interventions in mostly experimental and task settings; it doesn't model the messiness of a manager–report relationship built over years, where trust and timing change everything. It tells you feedback can backfire and broadly why, not the exact recipe for your next 1:1. Treat it as a warning label, not a script.

Three things we all call "feedback"

A great deal of feedback pain is a category error. In Thanks for the Feedback (2014), Harvard's Douglas Stone and Sheila Heen draw a line through three different things that all travel under the word: appreciation (recognition, "I see you, this mattered"), coaching (help to improve or grow), and evaluation (a rating against a standard, where you stand, and against whom). They're all legitimate. The trouble starts when the giver means one and the receiver hears another.

The classic collision: a manager offers coaching ("here's how you could sharpen the next deck"), and the report hears evaluation ("I'm being told I'm not good enough"). Once evaluation is heard, the coaching can't get in, the person is too busy calculating what the rating means for their job. The fix is almost embarrassingly small: label the type out loud. "This is just an idea to try, not a mark against you" sounds almost too simple, but it tells the receiver's brain which channel to open. And separate the streams in time: evaluation (the formal "here's where you stand") deserves its own conversation, because when it rides alongside coaching, it drowns it out.

flowchart TD
  A(["You're about to
give feedback"]) --> B{"Which kind
is this, really?"}
  B -->|"Recognition"| C(["Appreciation
'I see you, this mattered'"])
  B -->|"Help to improve"| D(["Coaching
'here's how to grow'"])
  B -->|"Rating vs a standard"| E(["Evaluation
'here's where you stand'"])
  C --> F(["Say which one
you mean, out loud"])
  D --> F
  E --> F
Stone & Heen's three types, most disputes come from the giver meaning one and the receiver hearing another. Leaders Loop

An honest limitation. The three-type distinction is a practitioner framework, not a peer-reviewed finding, its authority comes from decades of negotiation and conflict work at the Harvard Negotiation Project, not from a controlled study. It's a clarifying lens, and a good one, but don't dress it up as proven science.

Care personally and challenge directly

If Kluger and DeNisi tell you what to aim at, Kim Scott tells you how to hold your nerve. A former leader at Google and Apple, Scott built Radical Candor (2017) around two axes that most people treat as a trade-off but are really independent dials: care personally and challenge directly. Do both and you get candour. Drop the challenge and you get what Scott calls ruinous empathy, the well-meaning silence where you protect someone's feelings today and sabotage their growth for years. Drop the care and you get obnoxious aggression; drop both and you get manipulative insincerity, the worst quadrant of all.

Most managers don't fail at feedback by being too harsh. They fail by being too kind to say the hard thing.

Her diagnosis matters because it names the common failure correctly. The stereotype of the brutal boss is real but rare; far more managers live in ruinous empathy, softening the message until nothing survives the journey. In practice that means working the two dials on purpose: open by showing you care (specific, not flattering, "I want you to do well here, which is why I'm telling you this"), then say the hard part plainly and early, not buried in the ninth minute of a vague chat. The "care" earns you permission; the "challenge" is the thing they actually needed.

quadrantChart
  title Care personally vs challenge directly
  x-axis "Don't challenge" --> "Challenge directly"
  y-axis "Don't care" --> "Care personally"
  quadrant-1 "Radical candour"
  quadrant-2 "Ruinous empathy"
  quadrant-3 "Manipulative insincerity"
  quadrant-4 "Obnoxious aggression"
Scott's two dials are independent, not a trade-off, the goal is the top-right, and the common failure is top-left. Leaders Loop

An honest limitation. Radical Candor is a model from practice, easily misused as a licence to be blunt while skipping the "care" half, the part that does the real work. And what reads as "direct" varies sharply across cultures and power gaps; the same words from a senior leader to a junior new joiner land very differently than between peers. Use the framework, but read the room.

A worked example

Take a team lead, call her Priya, whose strongest engineer, Sam, keeps shipping clever code that nobody else can maintain. (Illustrative scenario; not a real team.) For two review cycles she's said nothing, telling herself he's brilliant and she doesn't want to dent his confidence. That's ruinous empathy with a polite name, and the cost is mounting: two other engineers have privately complained, and a release slipped because only Sam could touch the module.

Watch the difference the frameworks make. The version that backfires: in a tense standup, Priya says "Sam, you're a bit of a lone wolf, it's becoming a problem." That's a verdict on the person (attention flips inward, per Kluger and DeNisi), delivered as ambush, with no signal of which feedback type it is. Sam hears evaluation, I'm being marked down, and spends the conversation defending himself.

The version that works: in their 1:1, Priya names the type and the care first. "This is coaching, not a mark against you, you're one of the best engineers I've got, which is why I want to fix this." Then the challenge, pointed at task and effect: "Last week's auth module is elegant, but Maya and Tom couldn't extend it without you, and that's why the release slipped. I need code the team can own, not just code that works." Now there's somewhere to put the energy, a concrete behaviour, a visible consequence, a clear ask, and the relationship is a reason to listen, not a thing under threat.

flowchart LR
  A(["Sam ships code
only he can maintain"]) --> B{"How does Priya
raise it?"}
  B -->|"Verdict, ambush,
no label"| C(["'You're a lone wolf'
→ Sam defends himself,
nothing changes"])
  B -->|"Name the type,
care then challenge"| D(["'Coaching, not a mark.
The release slipped
because only you
could touch it.'"])
  D --> E(["Sam has somewhere
to put the energy
→ behaviour shifts"])
Same facts, two deliveries, one points at the person, one points at the work. Leaders Loop

Frequently asked questions

Doesn't the "feedback sandwich" solve all this?

Rarely. Wrapping criticism between two compliments tends to do one of two things: the receiver discounts the praise as mere packaging, or they cling to the praise and miss the point entirely. It also muddles Stone and Heen's three types, appreciation and coaching mashed into one confusing parcel. Clearer to name the type, lead with genuine care for the person, and make the one thing you need to change unmissable.

How do I receive feedback without getting defensive?

Start by knowing what's about to hijack you. Stone and Heen describe three triggers: the truth trigger (the feedback seems wrong), the relationship trigger (you don't rate the person giving it), and the identity trigger (it threatens how you see yourself). Naming the trigger as it fires, "this stings because it's hitting my sense of being competent", buys you the half-second to get curious instead of combative. A useful holding line: "Tell me more, what did you see?" It costs nothing and turns a verdict back into information.

What if the feedback is genuinely unfair or wrong?

You can find the one true thing inside a clumsy delivery without accepting the whole package. Even badly-aimed feedback usually contains a grain of how you're being perceived, and perception is data even when it's unfair. Separate two questions: "Is this accurate?" and "Is this how I'm landing on people?" The second can be worth acting on even when the answer to the first is no.

How often should I give feedback?

Little and often beats the annual ambush. The yearly review is where feedback goes to be too late, by the time it's delivered, the behaviour is months old and the moment to change it has passed. Small, specific, in-the-moment notes (both appreciative and corrective) keep the stakes low and the signal fresh, so no single conversation has to carry a year of unsaid things.

Isn't all this just being nice about telling people off?

No, the opposite. The point isn't to soften hard messages until they vanish; that's the ruinous-empathy trap. It's to deliver the hard message in a way that actually changes behaviour rather than triggering defence. Caring about someone is the reason to be honest with them, not an excuse to dodge it.

Related in the Toolkit

Feedback is an emotional-intelligence skill before it's a communication one, it rests on noticing your own reactions (self-awareness & emotional self-regulation) and reading the other person's state in the moment (empathy & social awareness). Get those wrong and the cleverest framework still lands badly.

Where to go next