One night I pasted my entire CV into GPT-4 and asked it to tell me what was wrong with me.
I want to be honest about what this post is before you read it as advice, because it is not advice. It is a transcript with commentary. Everything quoted below is the model’s opinion about one specific resume — mine — generated in a chat window, unverified, by a system that has never met me, never seen me work, and cannot tell the difference between a strong CV and a CV that is merely well-formatted. I read the output. I am the data point. Take the generalizations accordingly.
I did it because asking a person to critique your career honestly is expensive and awkward, and asking a language model is free and happens at 1 a.m. That trade has a cost. The cost is that the free thing tells you what you want to hear with a confidence it has not earned. This is a field note about noticing that, while it was happening, to me.
The first answer was a hug
The opening verdict read like a letter of recommendation I wrote about myself and forgot writing:
Your CV is comprehensive and showcases a strong background in finance, IT consulting, and enterprise systems… Overall, your CV is strong and presents a compelling picture of your skills and experiences.
This is the part where it told me I was great. And maybe I am! But notice what the model is actually doing: it is summarizing my CV back to me in a slightly warmer tone and calling the summary an analysis. “Comprehensive,” “robust educational foundation,” “impressive” — these are the adjectives of a system optimizing for me not closing the tab.
It did land three concrete suggestions, and these I’ll defend as genuinely useful because they’re true of almost everyone’s resume, including yours:
- Be more concise. Detailed is not the same as readable.
- Quantify achievements. “Led an ERP migration” is a sentence; “led an ERP migration that cut close time from 12 days to 5” is evidence.
- Make LinkedIn match the CV. Two documents that disagree about your career make a recruiter trust neither.
None of that required a multi-billion-parameter model. A good mentor says it in one breath. But the model said it for free, and it was right, so I’ll take the win and the caveat together.
So I asked it to stop being nice
The flattery was the tell. A critique that only finds strengths is not a critique; it’s a mirror with good lighting. So I changed the prompt to remove the model’s escape hatch. I asked it to analyze my CV again and only discuss weaknesses, and to do it by comparing me to a hypothetical peer with three times my experience.
That framing matters, and it’s the one genuinely clever thing I did in the whole conversation. “What’s wrong with my CV?” lets the model dodge — there’s always something nice to say instead. “Where do you fall short of someone with 3x your experience?” gives it a fixed reference point it can’t flatter its way around. You can’t tell someone they’re doing great relative to a person who is, by construction, doing better.
The answer got sharper immediately. The model stopped describing my resume and started describing a gap:
With significantly more experience, a peer might have taken on broader leadership roles, such as CTO, CIO, or consulting firm partner… A more experienced individual may have had the opportunity to lead global IT strategies, including digital transformation initiatives.
It listed things the more-senior version of me would have that I don’t: publications and speaking engagements. Patents. Board memberships. A track record of mentoring people up. Bigger budgets, higher-stakes programs, mergers and acquisitions. Relationships with C-level executives held over a longer arc.
Reading it back, I noticed two things at once. The first: this is good. It’s the most useful output of the entire session, because it reframed “improve my resume” as “here is the shape of the next ten years.” The second, which took longer to admit: most of this is a description of being more senior. Of course someone with triple the experience has more publications and bigger budgets. The model dressed a tautology in career-coaching language and I almost wrote it down as insight.
Where it started making things up
Here is the part I want flagged loudest, because it’s the failure mode that matters when you take advice from a confident text generator.
Among the development recommendations, the model suggested I:
Engage with cutting-edge technology projects… deeper expertise in emerging technologies like AI, blockchain, or quantum computing, applied to finance or enterprise resource planning.
Blockchain and quantum computing applied to my actual ERP work. Read that again. The model does not know what my job is. It knows the words near “finance” and “enterprise systems” in its training data, and “blockchain” and “quantum” are words that hang around those neighborhoods sounding impressive. This is not a recommendation. It is autocomplete wearing a blazer.
That’s the whole risk of this exercise in one bullet point. The model is equally fluent when it’s right (“quantify your achievements”) and when it’s pattern-matching nonsense (“have you considered quantum computing for your accounting close”). The prose is identical. The confidence is identical. The only thing that tells them apart is a reader who already knows enough to catch it — which is exactly the reader who needed the advice least.
To its credit, the model put a disclaimer on the comparison itself:
Remember, these considerations are speculative based on the premise of comparing to someone with triple your experience.
It told me it was speculating. I appreciate the honesty, even as I note that “here is some speculation, delivered in the same authoritative voice as the real parts” is precisely the thing that makes it dangerous. A disclaimer at the bottom does not change the font of the sentence above it.
What I actually kept
So I ran the most automatable version of career reflection there is — outsourced the introspection to a robot — and here is the honest yield, separated into the two piles it belongs in.
Kept, because it’s true and I can act on it: quantify everything, cut the length, make the documents agree, and treat “what would the 3x-experience version of me have” as a roadmap rather than a grade. That last one is worth the price of admission, and the price was zero.
Discarded, because it’s flattery or filler: every adjective in the first answer, the quantum computing, and the general lesson that a model comparing you to a more experienced peer will mostly describe more experience and call it a diagnosis.
The uncomfortable symmetry is that I’m a person who automates his problems for a living, and I tried to automate the one task that resists it: looking honestly at your own work. The model couldn’t do that for me. What it could do was hold up a structured mirror and occasionally fog it with things that sounded like advice. The reflection was useful. Telling the reflection from the fog was the actual work, and that part is still mine.
If you do this — and it’s a fine thing to do at 1 a.m. for free — change the prompt to deny the model its exit. Make it compare you to someone better, not to nothing. Then read the output as one opinion from a system that cannot tell when it’s guessing, and keep only the parts you could have defended to a human anyway.
It told me my CV was strong. It might be right. I don’t think it knows.