AI resume screeners are now in the hiring loop at most large companies. The reaction to this has been a mix of panic ("I have to game the AI") and dismissal ("It's just a fancy keyword filter"). Both miss what these tools are actually doing.

This post is a working mental model: what AI screeners score, how they differ from old-school ATS keyword filters, and what to actually do about it.

What an AI screener actually is

Most large-company ATSs in 2026 include an AI-augmented ranking layer on top of the underlying keyword-match parsing. The flow is: parse the resume, score it against the JD using a semantic model, rank candidates by composite score, then surface the top N to a recruiter. The recruiter still makes the call on whom to interview.

The AI's job is to order the stack, not make hire/no-hire decisions. This matters because the right candidate response is different depending on which step you're being filtered at. Keyword tricks help against pure keyword filters; semantic resume writing helps against AI rankers; bullet quality helps against the recruiter who reads the top of the stack.

For the broader ATS model, see how-applicant-tracking-systems-work. For the keyword-stuffing question specifically, see keyword-stuffing-vs-keyword-fit.

What the screener is actually scoring

Based on disclosed vendor methodology (Greenhouse, Lever, Workday have all published methodology summaries), the rough weights of what AI screeners look at:

• Semantic match to JD (skills, responsibilities). The largest factor. The model compares your bullet content to the JD requirements using sentence-embedding similarity, not exact-string matching. Synonyms count. "Built a data pipeline in Spark" matches "ETL pipeline development" even though no words overlap.
• Years of relevant experience. Counted from your dates. Inflated dates are detectable when they overlap with other dated rows.
• Level / title progression. Whether you've operated at the level the JD requests, signaled by your title history.
• Quantifiable outcomes in bullets. Bullets with numbers consistently rank higher than bullets without — likely because the model has learned that quantified bullets correlate with strong candidates in training data.
• Recency of relevant work. Work in the last 2-3 years weights more than work from 8 years ago.

Myths vs. reality

A few persistent myths worth dismissing:

"You need to repeat keywords X times to score high." Modern semantic models care about concept presence, not repetition. Saying "Python" once is enough; saying "Python" eight times across bullets often triggers a stuffing signal that downranks you.

"White-text keyword stuffing tricks the AI." This worked occasionally in 2017. It does not work in 2025. The parsing layer reads the underlying text; hidden-text patterns are now actively flagged.

"AI screeners only look at the top of the resume." Modern screeners parse the whole document. The top is still weighted slightly higher (more recent work matters more), but the bottom of page 2 still gets parsed.

"You can fool AI by mirroring the JD verbatim." Verbatim mirroring is detectable as a stylometric anomaly — the resume reads as different "voice" in different sections, which is a stuffing signal. Mirror at the concept level, not the sentence level.

"AI replaces recruiters entirely." It doesn't. AI ranks; a human still makes the interview call on the top 20 candidates. Write for the recruiter who reads that top of the stack.

What to actually do

The practical implications:

Match concepts, not strings. If the JD asks for "experimentation" and you ran A/B tests, write the bullet in your own words but make sure the concept of experimentation is present. The semantic model handles the synonymy.

Quantify bullets where honest. Bullets with numbers consistently rank higher. Don't invent numbers; do surface the ones you actually have. For the case where you genuinely don't have metrics, see quantifying-resume-without-metrics.

Lead with recent, relevant work. Recency weights are real. If your most recent work is the least relevant, consider a summary section that emphasizes the relevant adjacent work. See resume-summary-section.

Skip the tricks. Hidden text, font-size manipulation, copy-paste of the JD into the resume in a "skills" block — these are increasingly detected and penalized. The downside risk has caught up with the upside.

Tailor for the recruiter at the top of the stack. The AI's job is to surface you. The recruiter's job is to decide to interview you. Optimizing only for the AI gets you to the recruiter; failing to optimize for the recruiter once you're there loses you the slot.

The semantic-vs-keyword shift in practice

A useful way to think about it: the old ATS was a strict string-match filter. If the JD said "Java" and your resume said "JVM-based development," the old filter scored you as a miss. The new AI screener understands "JVM-based development" implies Java. The implication for resume writing is that you should write in your own voice with the concepts present — not in the exact strings of the JD.

This is also why the old advice to "copy the JD's exact language into your resume" has stopped working. The model now reads JD-mirroring as a stylometric anomaly, not as a match signal. Mirror concepts, not phrases.

For the broader question of how recruiters read resumes after the AI has ranked them, see how-recruiters-use-ats-filters.

Where the AI is reliably bad

A few systematic weaknesses worth knowing:

• Non-linear careers. Career changers and people with unusual paths often rank lower than they should. The semantic model assumes a linear trajectory; pivots confuse it. The fix is a strong summary section that explicitly contextualizes the pivot.
• Roles with unusual titles. "Forward-deployed engineer" and "growth engineer" sometimes don't map to standard role categories. Spell out scope in the bullets so the model can map to the JD.
• Deeply technical depth. AI screeners are better at breadth than depth. If your value is "5 years of distributed systems expertise at a hard problem," surface it in clear language at the top.

What this isn't

A few clarifications:

• It's not a reason to panic. AI screeners are an additional layer, not a replacement for human judgment. Strong resumes still rank highly.
• It's not the same across companies. Vendor differences are real. Greenhouse's ranking behaves differently from Workday's. Don't optimize for a specific vendor's quirks.
• It's not a license to use AI writers blindly. Generic AI-generated resumes are increasingly detected as such by — fittingly — AI screening tools that have been trained to spot them. Use AI as a draft-and-edit tool, not a generator-of-record.

The short version: AI screeners rank, recruiters decide. Match concepts not strings. Quantify what you can. Skip tricks that worked five years ago. Tailoring for the human reader at the top of the ranked stack still matters most.