Gartner predicts that by 2028, as many as 1 in 4 candidate profiles globally could be fake or materially misleading, driven largely by generative AI tools that make polished, keyword-perfect applications trivial to produce. The problem is already here: Willo's 2026 hiring research found that 76.6% of hiring teams currently encounter AI-generated or AI-assisted applications, and separate research found more than 80% of organisations have received portfolios containing inflated experience or unverifiable claims.
If you're looking to hire generative AI developers, evaluating portfolios beyond polished demos and well-written case studies is increasingly essential for distinguishing genuine technical expertise from AI-assisted presentations.
For generative AI portfolio red flags specifically, the stakes are higher than usual, because the exact skill being hired for, generating polished, confident-sounding content, is also the skill that makes a weak portfolio easy to fake. Here are nine patterns that separate a genuine GenAI builder from someone who's mostly demonstrated they can prompt an LLM to describe a project convincingly.
1. Every Project Reads Like the Same Template With the Names Swapped
A genuine builder's projects differ in the specific technical choices they made, chunking strategy, model selection, and evaluation approach, because different problems actually demand different solutions. A portfolio where every case study follows an identical structure and makes near-identical claims ("significantly improved efficiency," "streamlined the process") suggests the descriptions were generated after the fact from a template, not written from genuine, project-specific memory.
2. They Can't Explain Why They Chose One Model or Framework Over Another
Ask why they used GPT-4 instead of Claude, or LangChain instead of a direct API call, for a specific project. A real builder has an opinion tied to actual trade-offs they weighed, cost, latency, context window, specific capability gaps. A prompt tinkerer typically defaults to "it's what I know" or a generic answer about popularity, with no sign they seriously considered an alternative.
3. No Mention of Hallucinations or How They're Handled
Anyone who has genuinely deployed an LLM system has run into hallucinated output and had to build around it. A portfolio or interview answer with zero mention of hallucination mitigation, grounding, confidence thresholds, output validation against source documents, either means the person hasn't shipped anything real, or hasn't thought seriously about the biggest reliability problem in the field.
4. The Portfolio Only Shows What Worked
Real generative AI projects involve dead ends: a fine-tune that underperformed, a chunking strategy that hurt retrieval quality. This prompt worked in testing but broke on real user input. A portfolio that presents every project as a clean, linear success, with no "what didn't work" section anywhere, reads like it was written to look impressive after the fact rather than documented as it actually happened.
5. Impact Claims With No Baseline or Verification Path
"Improved response accuracy by 40 percent" means nothing without knowing 40 percent better than what, measured how. A genuine builder can name the baseline, the evaluation set, and the specific metric, and will often volunteer the limitation of their own measurement. A prompt tinkerer's claims tend to be precise-sounding numbers with no verifiable methodology behind them, which is a pattern serious enough that some hiring guides list it as the single most common inflation signal in AI portfolios.
6. They Can't Walk Through Their Own RAG Pipeline's Decisions
If a project claims to use retrieval-augmented generation, ask specifically how chunk size was chosen, what embedding model they used and why, and how they measured retrieval quality before generation even entered the picture. LLM integration developer guide covers exactly this kind of production-specific competency to screen for. A candidate who can only describe RAG at the conceptual level, retrieve then generate, without engaging with any of the actual engineering decisions, likely assembled the project from a tutorial rather than solving a real retrieval problem.
7. No Visible Awareness of Token Cost or Latency
Production LLM systems live under real constraints: unmanaged token spend can blow a budget fast, and latency directly affects whether users tolerate a feature. A genuine builder's case studies mention these constraints somewhere, prompt compression, caching, model routing to control cost, because they had to solve for them. A portfolio entirely silent on cost and latency, especially for a project claiming production deployment, suggests the deployment was smaller-scale or more theoretical than described.
8. Claiming Solo Credit on Work That Clearly Needed a Team
A claimed solo build of a large-scale, multi-component generative AI system, data pipeline, fine-tuning infrastructure, production serving, and monitoring, all built and owned by one person, is possible but uncommon, and worth probing directly. Ask them to walk through the infrastructure layer specifically. If they can't explain a part of the system they claimed credit for, the honest answer, "a platform team handled that piece, I owned the model and evaluation," is actually a stronger and more credible signal than an inflated solo claim.
9. GitHub Repos That Don't Actually Run
Clone the repository and try to set it up according to its own instructions. A surprising share of portfolio projects fail this basic test, missing dependencies, undocumented environment variables, or code that references a dataset that was never included. Beyond that, a single-commit history on a supposedly substantial project, no incremental development, no visible iteration, is a strong signal the code was generated or copied wholesale rather than built over time, the way real projects actually get built.
What Comes Next
As AI tools make it easier to produce a convincing-sounding portfolio and harder to distinguish genuine depth from polished description, the interview and trial-project stage will keep mattering more than the portfolio itself, not less. The nine patterns above aren't about catching someone in a lie for its own sake; they're about finding the layer where real experience and fabricated experience actually diverge, which is usually two or three follow-up questions past the initial pitch. GenAI skills for hiring managers is a useful companion read if you want the skills checklist to pair with this red-flags list. When you're ready to skip the guesswork, ai and ml freelance developers with a documented, verifiable track record can save you the vetting cycle entirely.
Frequently Asked Questions
Vague, unverifiable impact claims with no stated baseline are one of the most common and most damaging red flags, since a number like '40 percent better' is meaningless without knowing what it's being compared against. Combined with an inability to explain the actual technical decisions behind a claimed project when asked follow-up questions, this pattern reliably separates candidates who built something real from those presenting an inflated or AI-generated description of work they didn't fully do.
Clone their code and confirm it actually runs as documented, and ask specific follow-up questions about technical decisions, why a particular model, chunking strategy, or evaluation method was chosen, that a template-generated description wouldn't cover. A genuine builder can go several layers deeper than their written portfolio; someone presenting inflated or fabricated work typically runs out of specific detail within two or three follow-up questions.
Yes, using AI writing tools to polish portfolio descriptions is common and not inherently a red flag on its own. The concern is when the underlying project claims can't be substantiated with specifics when questioned directly, not the fact that the writing itself may have been AI-assisted. Focus vetting on verifying the technical substance rather than penalizing polished prose.
Hallucination is the single most common reliability problem in deployed LLM systems, and anyone who has genuinely shipped a production generative AI feature has had to build mitigation for it, whether through grounding in retrieved context, confidence thresholds, or output validation. A portfolio or interview answer with no engagement on this topic at all is a strong signal the person hasn't operated a real system past the demo stage, where hallucinations first become an operational problem rather than a curiosity.
Not automatically, but it's worth probing directly by asking them to walk through a specific infrastructure component of the claimed project. Large-scale generative AI systems typically involve multiple contributors across data engineering, model work, and deployment, so an honest breakdown of what they personally owned versus what the team built is a stronger signal of credibility than an unexamined solo claim, and doesn't necessarily indicate dishonesty if they can speak to it clearly.
A scoped, paid trial project, typically $500 to $2,000 on a real subset of your actual use case, is generally more reliable than a portfolio review alone, since it's much harder to fake real-time problem-solving on your specific data than to write a polished retrospective case study. Use portfolio review to shortlist candidates worth a trial, then let the trial itself be the deciding evidence before committing to a longer engagement.
Hire Generative AI Experts
Accelerate your project development with experienced generative AI developers who can deliver production-ready solutions, bypassing the hassle and cost of in-house hiring, schedule a consultation to learn more
Hire AI Developers