AI in resumes: how ChatGPT made the fake-candidate market 5× bigger

A year ago I could tell a strong CV from a weak one in five minutes. Now — no. ChatGPT made every CV equally polished, with STAR structure, impact metrics and tidy formulations. Strong senior and bootcamp graduate look the same. AI didn't kill CVs — it made them a much less reliable source of truth.

A year ago I could tell a strong CV from a weak one in five minutes. Now — no. ChatGPT and its clones have made all CVs equally smooth, with "STAR structure", impact metrics and tidy formulations. Strong candidate and weak candidate, three-year product engineer and someone who finished a bootcamp last month — they all look approximately the same. This doesn't mean AI killed CVs — it means CVs have become a much less reliable source of truth about a person.

What changed in 18 months

Before early 2025 I'd look at a CV and see variation. Telegraph style and dry facts in one. Long stories in another. "Responsibilities and results" structure in a third. Portfolio of cases in a fourth. From style you could read character, presentation experience, sometimes even level of autonomy.

Now almost every CV is "polished" through ChatGPT or specialised services like ResumAI, Easy.cv, Teal. Structure is the same. Phrasing is similar. Everyone has "increased DAU by X%", "optimised performance by Y%". The person who actually did it and the person who never worked with metrics look the same.

In parallel a whole "experience inflation" market grew up: for €300-800 they sell "commercial experience" with pre-prepared references at fictitious agencies. ChatGPT is just the final polishing layer in that loop.

Markers of a fabricated profile

Disclaimer: no single marker proves anything. But in combination they signal it's time to dig deeper. My current red-flag set:

• Pristine metrics without context. "Increased conversion by 47%" — from what to what? At what scale? What experiment? If the person actually did it, they'll explain within a minute. If not, they'll generalise.
• Too-even career graph. 12 months here, 18 months there, 24 months in the next role. Real careers are almost always "ragged" — sometimes overstayed, sometimes left in six months, sometimes stuck for four years. Suspicious smoothness is worth checking.
• Stack too broad for the level. A junior with three languages, five frameworks, Kubernetes and Kafka. Real experience reads as "familiar with", "worked alongside", "wrote in a team". Fabricated reads as "used in production".
• Match-to-JD phrasing word-for-word. If your job posting says "experience with PostgreSQL, Redis, RabbitMQ" and the CV has exactly those three words in the same order — it's either a meticulous person or a CV tuned to the posting.
• Companies hard to find. Generic-sounding LLC names. Some are real small studios. Some are "shell companies" from the experience-inflation market. Check them — a minute on a corporate registry usually clarifies.
• GitHub without contribution history. Person claims three years of commercial Python, but GitHub shows two forks of tutorial projects with one "initial commit". Not a verdict, but a reason to ask.

Changing the interview for the ChatGPT era

The shift isn't radical, but it's systematic. Main thing: you're assessing not "does the person know the technologies" but "can they describe the context of specific decisions." ChatGPT knows technologies cold. It doesn't know context.

Situational questions instead of knowledge

Bad: "Tell me about indexes in PostgreSQL."

Good: "Tell me about the last time you added an index in production. Which table? How many rows? What was before and after? Was there downtime? Who made the call?"

The first question passes a person who read a chapter. The second only passes someone who actually did it. ChatGPT can fabricate a plausible answer to the second, but it can't answer five follow-up "why" probes in a row.

Paired "why" questions

On any technical choice, ask two "why" in sequence. "You used Kafka — why Kafka and not RabbitMQ? — Why was that argument the dominant one for your specific task?" One level of abstraction ChatGPT handles. Two levels and it usually starts drifting.

Reverse engineering

Show the candidate a small piece of someone else's code (not yours, not theirs) — something resembling a typical backend service. Ask: "walk me through what's happening, and what would you change." Can't be prepared in advance, and can't be live-piped through ChatGPT without being obvious.

Pet-project walkthrough

If the candidate has a pet-project, ask them to screen-share and do a mini-walkthrough: "show me where users are stored", "how do you handle errors", "how do you test it." Ten minutes of this beats an hour of abstract Q&A.

Don't turn hiring into a search for the guilty

The temptation is obvious: scammed once today, polygraph tomorrow. This works badly for two reasons.

First. Most candidates "polishing" CVs through ChatGPT are normal people with real experience who simply want to look professional. If you treat them as suspects, they walk to competitors who behave like humans.

Second. Hard verification reads as low trust. On the senior market this is read immediately: "they're unsure of themselves and unsure of candidates." Strong candidates prefer employers who can evaluate without a polygraph.

Healthy frame: not "how do I catch a fraud" but "how do I make real skills visible." Then the person who's trying to deceive falls out on their own — they simply have nothing to show at the level of concrete cases.

Three minimal changes you can make tomorrow

If you have a current interview format and want to make it resistant to AI-CVs — three minimal moves:

• Remove knowledge questions. "Tell me about SOLID", "what's the difference between a hash table and a tree". Replace with situational.
• Add 15 minutes of reverse engineering or live coding. Not leetcode on linked lists — a small case from your domain.
• Two minutes on "tell me about the team at your previous job." Size, roles, how decisions were made, who you paired with. Can't be invented. Can't be memorised.

The experience-inflation market specifically

A separate category — people literally buying CV lines. Not just AI polishing, outright fabrication. Markers: shell companies, date inconsistencies, inability to provide a reference with a real former manager.

In the junior and middle market these CVs appear regularly. At senior level they're rarer because the fraud cracks at the first deep interview. Defence for juniors and middles: real technical discussion of a task plus a request to put us in touch with "one former colleague". If there isn't a single contactable name, ask why.

When we get involved

We redesign technical interviews for teams that hit the fake-CV wave. One-off engagement, 1-2 weeks: review your current questions, replace "knowledge" with situational, train tech leads to filter AI polish.

If you've closed a "middle" who turned out to be a junior in skill — email [email protected]. We'll review your current interview format and propose specific changes.