Ask any senior engineer or recruiter. Candidates who have applied for more than five jobs know how to stuff keywords into their resumes. Hidden white text at the bottom. Repeated bullet points. Fake skills sections with every variation of every technology. If you’ve been screening resumes for a while, you’ve seen all of it.
This is not fringe behavior. It’s the rational response to a broken screening system. If the filter is Boolean, “does Python appear 3 or more times, does AWS appear 2 or more times, does Kubernetes appear 1 or more times,” then stuffing keywords is the correct strategy to pass the filter. The ATS taught candidates to do this.
Your ATS rewards candidates who learned how to game it. The candidates who just wrote honest descriptions of their real work get filtered out in favor of the ones who figured out the rules.
Why legacy ATS keeps falling for it
Every legacy applicant tracking system uses some form of keyword matching under the hood. The details vary. Boolean extraction, TF-IDF scoring, bag-of-words, semantic similarity with embeddings. The mechanism is the same: match tokens from the resume against tokens from the job description.
This has a fundamental vulnerability. It’s deterministic. A candidate who knows the rules can reliably produce the output they want, regardless of actual fit. Even the “AI-powered” versions (semantic embeddings, BERT-based matching) are vulnerable. A candidate who copies job description language into their resume summary scores high on semantic similarity without any actual experience.
What actually works
The fix isn’t a smarter keyword model. The fix is to stop ranking by keywords altogether. CurriculoATS reads resumes for measurable outcomes instead of token presence.
Specifically, the AI looks for four categories of signal:
A candidate who writes “managed Python” five times gets no credit. A candidate who writes “built a real-time fraud detection pipeline that processed 2M events/sec” gets recognized for the actual outcome, even though “Python” might appear zero times in that sentence.
Why resume poisoning stops working
Hidden white-text keyword dumps add nothing to an outcome-based score. The model isn’t counting token occurrences. It’s reading descriptions of real work. An unused skill list doesn’t look like outcome-based work, so it contributes zero.
What does work, for candidates, is writing specific descriptions of what they actually built. That’s a good thing. It rewards clarity and effort over gaming the system. It’s also structurally impossible to game, because there’s no hidden ranking rule to exploit.
If the ATS reads outcomes, honest detailed descriptions win. Keyword stuffing becomes pointless. Candidates who learned to game the system suddenly don’t have an advantage over candidates who wrote the truth.
Written reasoning changes the game for recruiters
CurriculoATS writes a full reasoning paragraph for every 0-100 score. This matters in one specific way. If a candidate lands at 78, you can see exactly which outcomes matched the job. If a candidate lands at 52, you can see what was missing. No more guessing at the black box.
This also protects strong candidates from unfair low scores. If the AI misses a key outcome, the reasoning paragraph makes it visible. A recruiter can flag the score, override it, and the model learns from corrections over time.
Real-world impact
Curriculo’s Impact Scoring engine reads outcomes and writes reasoning for every candidate, every time. Combined with signal-based screening that evaluates work instead of tokens, adversarial resume poisoning stops working.
Here’s the best part: the full outcome-based scoring engine is on the free plan. Not a trial. Not stripped-down. Every feature including 0-100 AI scoring with written reasoning, unlimited team members, no credit card. Pro at $100/month (currently $50 early bird) just adds unlimited active jobs and priority support.
Related reading: What is outcome-based hiring? and Signal-based hiring methodology.