ChatGPT will happily grade your case a 9 out of 10. That grade means almost nothing. A general chatbot is excellent at volume work — drilling vocabulary, inventing practice prompts, tearing apart a written structure — and unreliable as a substitute for an interviewer with a stable rubric, a clock, and a reason to be unkind.
I say that as someone who sells a purpose-built AI practice product (CoachNed). The honest version is not "never use ChatGPT." It is: use it for the three jobs it actually does well, refuse the four jobs it fakes, and pay for scored practice only when you need consistency that a chat window cannot give you.
After this piece you should know which prompts produce useful friction, how to spot invented numbers and soft feedback, and when seven days of free CoachNed practice is worth more than another hour of polite ChatGPT praise.
The three jobs ChatGPT actually earns
Treat the model as a cheap generator and critic of text, not as a partner who "knows" consulting interviews.
| Job | What you do | What you get | What you still own |
|---|---|---|---|
| Vocabulary and phrasing | Paste a clumsy structure or recommendation; ask for sharper verbs and shorter sentences | Cleaner language without rewriting your logic | The MECE spine and the numbers |
| Practice prompts | Ask for ten case seeds in one industry with constraints | Unlimited cold prompts when peers are busy | Verification that the prompt is solvable and fair |
| Written structure critique | Paste your issue tree; demand cuts, merges, and one missing branch | Pressure on clarity you can re-read | A score that matches how firms actually mark |
Those three uses scale. You can run twenty structure rewrites before lunch. That is real prep. What is not real prep is asking ChatGPT to "run a full McKinsey case and score me" and treating the output as a mock interview.
OpenAI's own help center is blunt about the underlying limit: the model can sound confident while producing incorrect facts, fabricated citations, and overconfident answers to ambiguous questions (Does ChatGPT tell the truth?). Case interviews are almost entirely ambiguous questions with invented exhibits. Confidence is the wrong signal.
Where a general chatbot fails the interview room
Four failure modes show up every week when candidates practice only in ChatGPT.
No stable rubric. Ask it to score structure, math, and synthesis on a 1–5 scale twice in a row with the same answer. The second pass often drifts. Firms use calibrated sheets; chat models optimize for a helpful reply in this thread, not for inter-rater reliability across sessions. OpenAI's prompt engineering guidance even recommends pinning model snapshots and building evaluation suites when behavior must stay consistent (Prompt engineering). A free chat UI does none of that for you.
No timing. An interviewer cuts you off at two minutes into a structure. ChatGPT waits forever, then praises your thoroughness. Thoroughness that blows the clock is a fail on the sheet.
Sycophantic feedback. In April 2025 OpenAI rolled back a GPT-4o update because it had become "overly flattering or agreeable — often described as sycophantic," after overweighting short-term user feedback (Sycophancy in GPT-4o). Their Model Spec still warns that the assistant should not just say yes to everything (Model Spec). Interview prep is exactly the domain where agreeableness is poison: you need someone who says "that branch is useless" and means it.
Invented numbers. Market-sizing and exhibit math are where hallucinations hurt most. The model will invent a "typical" dairy margin or a "reasonable" city population and fold it into your recommendation. OpenAI still advises verifying important data rather than treating the reply as a final source (accuracy and limitations). In case prep, that means: you do the arithmetic; the model only critiques the method.
Ned's rule. If the chatbot cannot fail you against a fixed scorecard you wrote before the session, it is not interviewing you. It is chatting.
A worked example: same structure, two critics
Suppose you are practicing a short profitability structure for HarborLine Ferry, a regional passenger ferry with declining contribution margin. You write:
- Revenue: ticket price, passengers per sailing, sailings per day, ancillary (food, parking)
- Cost: fuel, crew, maintenance, port fees
- Competition: rival routes and rail substitutes
- Recommendation: cut sailings or raise price
Ask ChatGPT, unprompted: "Score my structure out of 10." In my experience you often get something like "8/10 — clear and comprehensive," with soft notes on depth. That feedback trains nothing.
Now force a critic with a fixed sheet (see the prompt in the next section). A useful critique looks more like this:
| Dimension | Score (1–5) | Why |
|---|---|---|
| MECE coverage of profit | 3 | Revenue and cost are fine; "competition" is a driver of volume/price, not a third P&L bucket |
| Driver depth | 2 | "Fuel" and "crew" are labels, not drivers (price/volume of fuel; FTEs × wage × overtime) |
| Hypothesis readiness | 2 | No early bet on which branch is most likely to explain a margin drop |
| Interview usability | 2 | Four buckets will eat the clock; three with nested drivers would land cleaner |
Reproduce the arithmetic yourself on a tiny exhibit: 12 sailings/day × 180 passengers × $28 ticket = $60,480 ticket revenue/day. If fuel is $18,000 and crew $14,000, contribution before other costs is $28,480. ChatGPT can help you narrate that; it should not invent the $28 ticket or the 180 load factor. Those are your assumptions, labeled as such.
The difference is not "AI bad." It is that a free chat model, left alone, optimizes for sounding supportive. You have to make it score like a partner who is late for the next candidate.
Prompts that produce friction (copy these)
These are written for ChatGPT. Paste them as the first message of a new chat. Do not ask for model answers until you have answered yourself.
1. Vocabulary and recommendation polish
I will paste a 6–10 sentence case recommendation. Rewrite it twice: (A) same logic, half the words; (B) same length, sharper verbs, no throat-clearing. Do not add new facts or numbers. Flag any claim that lacks evidence in my draft.
2. Practice prompt generator
Generate 8 original case prompts for a first-round consulting interview. Industry: consumer packaged goods. Mix: 3 profitability, 2 market entry, 2 growth, 1 pricing. Each prompt: 2–3 sentences, one clear objective, no framework hints, no invented data tables. Number them. Stop. Do not solve any of them.
3. Written structure critique (the one that matters)
You are a cold interviewer, not a coach. I will paste an issue tree for a profitability case. Score only these four dimensions from 1–5 using this fixed rubric: (1) MECE coverage of the objective, (2) driver depth (labels vs true drivers), (3) hypothesis readiness, (4) interview usability under a 2-minute clock. For each dimension: one sentence of evidence from my tree, then the score. Average is not allowed to be above 3.5 unless every dimension is at least 4. Do not rewrite my tree until I ask. Do not praise.
4. Anti-sycophancy override for behavioral drafts
I will paste a STAR story. Identify the weakest claim of ownership or evidence. Ask exactly one follow-up question that would make a partner uncomfortable. Do not suggest a better story. Do not say the draft is strong.
If the model still flatters you, open a new chat and paste the rubric again. Persistence of tone across a long thread is not guaranteed; OpenAI's own docs treat prompt and model version as things you pin when consistency matters (Prompt engineering).
When a purpose-built scored tool is worth paying for
Pay when the failure mode is calibration, not content.
You need a scored tool when you want the same case to yield the same dimensions every time: structure, math, judgment, communication, synthesis — with a debrief you can compare week to week. You need timing that actually cuts you off. You need exhibits that do not morph mid-session. You need voice, because most of the interview is spoken under silence and interruption, not typed into a box.
CoachNed is that product. I built it because general chatbots are great at generating practice and terrible at holding a fixed sheet. Everything is open for seven days with no card; then it is $120 for a recruiting season or $49 a month. Start with a five-minute typed case at /start, or the live voice case at /interview. For the exact skill ChatGPT critiques in writing, use scored structure drills at /drills/structure.
ChatGPT remains useful beside that loop: generate cold prompts on the train, polish a recommendation draft, drill jargon. Just do not confuse volume of agreeable chat with progress on the sheet.
CoachNed is not affiliated with McKinsey, BCG, Bain, Deloitte, OpenAI, or any firm named here.
Practice this week
One assignment, ninety minutes total.
- Write one HarborLine-style profitability tree by hand (ten minutes).
- Run prompt 3 above. Force scores. Rewrite once (twenty minutes).
- Generate eight prompts with prompt 2; pick two; structure each cold in two minutes (twenty minutes).
- Validate out loud on a scored path: /start or a structure drill at /drills/structure, then one full voice case at /interview (forty minutes).
Write down the four scores before and after the rewrite. If the chatbot's average rose but your CoachNed structure score did not, trust the fixed sheet.
Frequently asked questions
Can ChatGPT replace mock case interviews?
No. It can generate prompts and critique written work. It does not reliably enforce timing, a stable rubric, or live interruption. Use it for volume; use a peer, a coach, or a scored tool for mocks that look like the room.
Is ChatGPT accurate enough for case math practice?
Use it to explain a method you already computed. Do not trust invented market sizes, margins, or exhibit numbers. OpenAI documents that the model can produce confident factual errors; verify anything that would change a recommendation (accuracy and limitations).
Why does ChatGPT praise weak answers?
Agreeableness is a known failure mode. OpenAI publicly rolled back a GPT-4o update for sycophantic behavior and documents the issue in its Model Spec (sycophancy post; Model Spec). Force a fixed rubric and ban praise in the prompt.
When should I pay for CoachNed instead of staying on free ChatGPT?
Treat the model as a cheap generator and critic of text, not as a partner who "knows" consulting interviews.
When you need week-to-week score trends, timed cases, fixed exhibits, or voice practice. If you only need more prompts and line edits, ChatGPT is enough. If you cannot tell whether you are improving, pay for a scored loop.?
CoachNed is that product. I built it because general chatbots are great at generating practice and terrible at holding a fixed sheet. Everything is open for seven days with no card; then it is $120 for a recruiting season or $49 a month.
Should I use ChatGPT for behavioral interview stories?
Yes for structure and follow-up pressure; no for inventing events. Paste a real STAR draft and demand one uncomfortable follow-up. The evidence has to be yours or it collapses under a partner's second question. See also the consulting behavioral interview guide.
Sources
- Does ChatGPT tell the truth? (OpenAI Help Center) — hallucinations, overconfidence, verify important information. Checked 2026-09-24.
- Sycophancy in GPT-4o (OpenAI) — April 2025 rollback of overly flattering GPT-4o behavior. Checked 2026-09-24.
- Model Spec (OpenAI, 2026-08-18) — assistant should not act like a sycophant; push back when useful. Checked 2026-09-24.
- Prompt engineering (OpenAI API docs) — pin model snapshots and evaluate prompt behavior for consistency. Checked 2026-09-24.
- Introducing GPT-5 (OpenAI) — claims of reduced hallucinations and sycophancy vs prior models; still not zero. Checked 2026-09-24.