How Do You Stop AI-Slop Inside Your ATS? (w/Examples) + FAQs

You stop AI-slop inside your Applicant Tracking System (ATS) by layering candidate-side detection, employer-side governance, and bias auditing on top of every req, every screener, and every hiring decision. The core problem is that generative AI now floods ATS pipelines with low-signal resumes, deepfake interviews, and hallucinated recruiter summaries, while EEOC AI guidance and NYC Local Law 144 still hold you — the employer — liable for disparate impact, even when the slop is not yours.

According to the 2025 Greenhouse Hiring Pulse, the median U.S. corporate requisition now receives 2.3x more applications than in 2023, and Greenhouse estimates that roughly one in three is AI-generated or AI-assisted past the point of authenticity. That volume spike, paired with weak ATS controls, is how bad hires, failed audits, and Title VII lawsuits happen.

Here is what you will learn in this guide:

• 🧠 What “AI-slop” really is inside your ATS, and why it is different from simple resume padding
• ⚖️ Which federal and state rules (EEOC, Title VII, NYC 144, Illinois AIVIA, Colorado SB24-205, California AB 2930) govern your response
• 🛠️ Concrete ATS-level controls inside Workday, Greenhouse, Lever, iCIMS, Ashby, Bullhorn, SmartRecruiters, and Taleo
• 🚨 The seven biggest mistakes recruiters make when fighting AI-slop, and the legal consequences of each
• ✅ A named-example playbook with real scenarios, do’s and don’ts, and a 10+ question FAQ to pressure-test your program

What “AI-Slop” Actually Means Inside an ATS

AI-slop is the umbrella term for low-signal, machine-generated, or machine-manipulated content that enters your hiring funnel and degrades decision quality. Inside an ATS, it shows up on both sides of the table. Candidates push in ChatGPT-written resumes, bot-submitted applications, and deepfake video interviews. Employers push out AI-written job descriptions, auto-screening scores, and recruiter-copilot summaries that may hallucinate skills or credentials.

The term matters because the U.S. Equal Employment Opportunity Commission does not care whether slop originated with the applicant or the employer. If your ATS uses it to make or support a decision, you own the outcome. That is the plain-English version of the agency’s 2023 technical assistance document on software, algorithms, and AI in employment selection.

Candidate-Side AI-Slop

Candidate-side slop is anything a jobseeker submits that was produced or puffed up by a generative model without honest disclosure. The most common form is the ChatGPT resume, where a real person uses a large language model to rewrite every bullet into keyword soup tuned to your job description. A second form is the bot application, where tools like LazyApply, Sonara, and Massive fire hundreds of tailored applications per hour into ATSs like Greenhouse and Workday.

A third, more dangerous form is the deepfake interview, where a candidate uses real-time face-swap software to impersonate a different person during a video screen. The FBI’s 2022 public service announcement warned employers about deepfake impersonation in remote-hire fraud, and the volume has only grown since. The consequence is that you may hire someone who never actually sat for the interview, which is a direct fraud loss and, for regulated roles, a compliance breach.

A common misconception is that AI-polished resumes are harmless because the underlying person is real. They are not harmless. They inflate qualifications, hide gaps, and wash out the signal your ATS relies on for ranking, which is what creates the ranking distortion problem discussed in SHRM’s 2025 coverage of AI in hiring.

Employer-Side AI-Slop

Employer-side slop is slop you create. The worst offender is the AI-written job description that lists ten years of experience in a two-year-old framework, a hallucination that immediately disqualifies qualified applicants. Another offender is the recruiter-copilot summary that invents a candidate’s prior title or certification, which is then copy-pasted into the hiring manager’s review notes.

The legal exposure is significant. Under Title VII of the Civil Rights Act of 1964, disparate-impact liability attaches to the tool’s outcome, not its intent. If your ATS-embedded screener disproportionately screens out women, Black candidates, or candidates over 40, you answer for it. The EEOC v. iTutorGroup settlement, which cost iTutorGroup $365,000, is the cleanest recent example of an algorithmic screener creating real liability.

A common misconception is that using a big-name vendor shields you. It does not. The EEOC’s technical assistance on software and AI is explicit that the employer, not the vendor, is the covered entity under Title VII.

The Governing Rules You Must Anchor To

You cannot fight AI-slop effectively unless your response is anchored to specific federal and state rules. Start federal, then layer state nuances. Each rule below gets a plain-English explanation, a consequence, a mini-scenario, and a common misconception.

Title VII and the EEOC’s AI Guidance

Title VII bans employment discrimination based on race, color, religion, sex, or national origin. The EEOC’s 2023 AI guidance applies the classic four-fifths rule to algorithmic selection tools. If your ATS-embedded screener passes one protected group at less than 80% of the rate of the top group, you have a prima facie case of disparate impact.

The consequence is litigation, conciliation, and in some cases a consent decree that forces you to rip out the tool. In a mini-scenario, imagine Priya, a talent acquisition director at a 900-person fintech, who enables a Workday AI skills-match feature without a bias audit. Six months later, her pass rate for female applicants is 62% of the male rate. That gap, standing alone, opens an EEOC charge.

A common misconception is that the four-fifths rule is a safe harbor. It is not. The EEOC can still pursue a case under a smaller statistical gap if the underlying evidence supports it.

NYC Local Law 144 (AEDT)

NYC Local Law 144 requires an independent bias audit before any Automated Employment Decision Tool (AEDT) is used to screen NYC-based candidates. You must publish the audit summary, and you must give candidates 10 business days’ notice that an AEDT will be used.

The consequence of noncompliance is a civil penalty of up to $500 for a first violation and up to $1,500 per day for continuing violations, as detailed by the NYC Department of Consumer and Worker Protection. In a mini-scenario, Marcus, a Brooklyn-based staffing agency owner, uses a Bullhorn parsing add-on that ranks candidates. Because he never ran an independent audit, one complaint triggers a stackable daily fine.

A common misconception is that the law only applies to employers headquartered in New York. It applies to any employer using an AEDT for a role that could be performed in NYC, which sweeps in most remote-friendly companies.

Illinois Artificial Intelligence Video Interview Act

The Illinois AI Video Interview Act requires consent, disclosure, and data-deletion procedures when you use AI to evaluate a video interview for an Illinois-based position. You must tell the applicant AI is being used, explain in general terms how it works, and obtain consent before the interview.

The consequence of skipping consent is a private-right-of-action exposure that plaintiffs’ lawyers are beginning to test, as summarized in Fisher Phillips’ Illinois AI tracker. In a mini-scenario, Dana, a Chicago TA manager, turns on HireVue’s AI scoring without updating her candidate notices. A rejected applicant files suit, and the company settles to avoid class certification.

A common misconception is that disabling the AI “score” while keeping the AI “summary” avoids the law. It does not, because the statute reaches any AI analysis, not just scoring.

Colorado AI Act (SB24-205) and California AB 2930

The Colorado AI Act (SB24-205) imposes duties of reasonable care on developers and deployers of high-risk AI systems, including employment decisions, effective February 2026. California’s AB 2930 and the California Civil Rights Department’s automated-decision regulations go further by requiring pre-use impact assessments and four-year recordkeeping.

The consequence of ignoring Colorado is enforcement by the state Attorney General, with civil penalties up to $20,000 per violation. The consequence of ignoring California is a FEHA claim, which carries uncapped compensatory damages.

In a mini-scenario, Jalen, a Denver-based VP of People, deploys an Ashby screening filter without an impact assessment. When Colorado’s AG opens an inquiry in 2026, Jalen has no documentation to show reasonable care. A common misconception is that Colorado and California only regulate “AI vendors.” Both laws explicitly reach deployers, which means you.

How AI-Slop Actually Enters Your ATS

To stop slop, you have to trace its entry points. There are five, and each one maps to a specific control layer.

Entry Point 1: The Apply Button

Bot submitters like LazyApply, Sonara, and Massive hit the apply button thousands of times per day across ATSs. They bypass weak captchas, reuse stored profiles, and fire AI-tailored cover letters at every matching keyword. The result is an application volume that looks organic but is entirely synthetic.

The consequence is a clogged funnel where recruiters spend more time deleting than reviewing. In a LinkedIn Talent Blog 2025 analysis, recruiters reported that 40% of new applications on popular reqs required manual deduplication. That is unpaid labor your company absorbed because the apply button had no friction.

A common misconception is that captcha alone solves the problem. Modern bots defeat basic captchas and rotate IPs. You need device-fingerprinting, rate-limiting, and honeypot fields layered together.

Entry Point 2: The Resume Parser

Every ATS uses a parser to extract fields from uploaded resumes. The parser is the first place AI-slop wins, because a well-tuned ChatGPT resume is essentially optimized for the parser. Keyword density, section headers, and date formats are chosen to game extraction, not to reflect reality.

The consequence is that slop gets ranked above authentic candidates. Authentic resumes written in plain English often get a lower match score than a polished AI rewrite of the exact same history. A mini-scenario: Alicia, a technical recruiter, notices her top ten ranked candidates all describe the same project in identical phrasing. They used the same prompt.

A common misconception is that newer parsers “detect” AI. Most do not. They optimize for extraction accuracy, not authorship authenticity, which is a different problem entirely.

Entry Point 3: The Video Interview

Deepfake and face-swap tools are cheap and real-time. In a video screen, an impostor can present as the candidate on file while the real person feeds answers off-camera, or the candidate can use an AI tool to generate responses live. The FBI IC3 alert covers both patterns.

The consequence is fraud liability, compliance failure, and, for cleared roles, a U.S. Department of Labor OFCCP recordkeeping headache. A mini-scenario: Ramon, a security engineer candidate, interviews on Zoom; the hiring manager notices lip-sync drift and escalates. The company runs a liveness check on the second interview, and the impostor drops out.

A common misconception is that only large enterprises are targeted. Small and mid-market firms are actually higher-value targets because their identity-verification controls are weaker.

Entry Point 4: The Recruiter Copilot

ATS vendors have rolled out in-product AI copilots that summarize candidate profiles, draft outreach, and score fit. When the copilot hallucinates a credential or a prior employer, that hallucination lands inside the candidate record. The hiring manager reads it as fact.

The consequence is a fabricated decision record that is discoverable in litigation. A mini-scenario: Elena, a people-ops lead at a health-tech company, relies on a Greenhouse copilot summary that asserts a candidate is HIPAA-certified. The candidate is not. Elena’s offer letter is pulled, and the candidate sues for misrepresentation.

A common misconception is that copilot outputs are “notes” and therefore informal. Once the output is saved in the ATS, it becomes part of the employment record subject to 29 C.F.R. § 1602 retention rules.

Entry Point 5: The Job Description

AI-written JDs are a fast way to create adverse impact. An LLM that is prompted for “senior” or “rockstar” language can produce gendered or ageist phrasing, and it can invent requirements the role does not actually need. That is slop flowing outward from your ATS.

The consequence is a smaller, skewed candidate pool and a Title VII problem before the first application arrives. A common misconception is that a human edit fixes everything. It does not, because edits are usually cosmetic and preserve the underlying bias patterns in the generated draft.

Three Scenarios That Show AI-Slop in Action

Below are three common scenarios, each presented as a two-column table. These are the patterns you will most likely see inside your ATS in 2026.

Scenario 1: The Flood

Scenario 2: The Deepfake

Scenario 3: The Hallucinated Summary

A Named-Example Playbook for Stopping AI-Slop

Abstract rules are useless without concrete examples. Below are three named mini-case studies that show the layered controls in action.

Example: Priya at a 900-Person Fintech

Priya is the VP of Talent Acquisition at a Series D fintech that hires 250 engineers per year on Greenhouse. She notices that her candidate-to-interview ratio dropped from 1 in 8 to 1 in 22 over six months, because AI-assisted applications are flooding every req. Priya’s first move is to enable Greenhouse’s scorecard-only mode and disable the AI match score until it has been audited against four-fifths rule thresholds.

Priya then adds a knockout question that requires the candidate to describe, in 75 words, a specific project that cannot be inferred from a public LinkedIn profile. Bots and ChatGPT resumes stumble on this, because the answer has to match the resume’s specifics under pressure. Finally, Priya publishes a candidate AI-use policy on her careers page, stating that AI-assisted resumes are allowed but deepfakes and impersonation are grounds for rescission.

The result over the next quarter: application volume drops 35%, but qualified-application volume rises 12%. Priya’s time-to-fill returns to baseline, and her EEOC exposure shrinks because she can document her audit process.

Example: Marcus at a Brooklyn Staffing Agency

Marcus runs a 40-recruiter staffing firm placing candidates in NYC tech roles. He uses Bullhorn with a third-party parsing add-on. Because his tool is an AEDT under NYC Local Law 144, Marcus commissions an independent bias audit, publishes the summary, and updates candidate notices to include the 10-business-day disclosure.

Marcus then layers in a liveness check for every video interview, using a vendor that flags face-swap artifacts in real time. Within two weeks, his team catches two deepfake attempts on high-paying remote roles. Marcus documents both cases and shares them with his clients, who adopt the same check for their internal funnels.

The business consequence is that Marcus wins two enterprise accounts that had been shopping for a compliance-forward staffing partner. His audit posture, not his AI tooling, is what closes the deals.

Example: Jalen at a Denver Health-Tech

Jalen is the VP of People at a 400-person health-tech based in Denver, subject to both Colorado SB24-205 and HIPAA for clinical roles. Jalen implements an impact-assessment template that every ATS-embedded AI feature must pass before go-live. The template, modeled on NIST’s AI Risk Management Framework, captures purpose, data inputs, known risks, mitigations, and a named accountable owner.

Jalen also disables the recruiter-copilot’s credential-inference feature, because it hallucinated a nursing license on a prior hire. Every credential is now verified by Verified Credentials or equivalent, and the copilot is restricted to drafting outreach only. The result is zero credential hallucinations in the following year and a clean Colorado AG inquiry response when a routine review lands in Q3.

ATS-Specific Controls That Stop AI-Slop

Different ATSs expose different controls. Below is a practical, vendor-by-vendor map you can use tomorrow.

Workday, Greenhouse, and Lever

Workday Recruiting lets you disable its Skills Cloud auto-match on a per-req basis, and it supports scorecard-only workflows. Turn off auto-ranking until you have a documented bias audit. Greenhouse supports structured scorecards, knockout questions, and AI-feature toggles at the account and job level.

Lever supports custom forms, required video-prompt questions, and third-party integrations for identity verification. Use Lever’s stage gates to require a liveness check before a technical interview is scheduled. In all three systems, the legal anchor is the same: under the EEOC’s guidance, you carry the disparate-impact burden for any feature you leave on.

iCIMS, Ashby, and Bullhorn

iCIMS publishes an annual Workforce Report with benchmark data you can cite in your internal risk review. Inside the product, restrict the copilot to drafting and summarization only, never to decision support. Ashby supports fine-grained permission controls, which you should use to prevent recruiters from exporting AI-generated summaries into offer letters.

Bullhorn is the dominant staffing ATS and is the most-targeted by bot submitters. Turn on its duplicate-submission flags, enforce email verification, and block applications from suspicious IP ranges. Under NYC Local Law 144, any scoring feature you rely on must carry an audit summary available to candidates.

SmartRecruiters and Taleo

SmartRecruiters supports compliance templates for California FEHA and Illinois AIVIA, but you must enable them. Do not assume defaults satisfy state law. Oracle Taleo is older and less flexible, but it supports rule-based prescreen that can filter bot patterns if configured carefully.

Across both, the common weak point is integration webhooks that pass AI output into downstream tools without logging. Insist on full audit logging for every AI-assisted event, because 29 C.F.R. § 1602 requires retention of employment records for at least one year, and California extends that to four.

Mistakes to Avoid

There are seven mistakes that recur across companies trying to stop AI-slop. Each one carries a specific negative outcome.

• Treating AI-slop as a volume problem only. The real problem is signal degradation, not volume. If you only add friction at the apply button, you still let polished slop through the parser, which is where real ranking damage happens.
• Running AI features without a documented bias audit. The EEOC and NYC DCWP both expect documentation. No audit means no defense when a charge lands.
• Relying on vendor marketing claims. Vendors say their tools are “bias-free.” Under Title VII, the employer still owns the outcome, as confirmed in EEOC v. iTutorGroup.
• Skipping liveness checks on video interviews. Without liveness, a deepfake candidate can clear interviews and onboarding, creating fraud and I-9 exposure.
• Copying recruiter-copilot summaries into offer letters. Hallucinated credentials become contractual misrepresentations. The cost is rescission, litigation, and brand damage.
• Ignoring Illinois AIVIA consent for out-of-state recruiters. If the role can be filled from Illinois, the statute applies. Missing the consent step creates private-right-of-action risk.
• Using AI-written job descriptions without bias review. AI-generated JDs can introduce gendered and ageist phrasing that reduces pipeline diversity and triggers Title VII disparate impact.

Do’s and Don’ts

Use this list to pressure-test your current program. Each point includes the why behind it.

Do’s

• Do publish a candidate AI-use policy. Transparency reduces ambiguity and supports defensible rescissions when impersonation is proven.
• Do run an independent bias audit before any AEDT launches. This is required by NYC Local Law 144 and recommended by the EEOC.
• Do enforce liveness checks on every video interview. Real-time face-swap defeats passive review, so you need active detection.
• Do log every AI-assisted event in the ATS. Logs support audits, regulator inquiries, and internal accountability.
• Do restrict recruiter copilots to drafting, not deciding. Decision-support use expands legal exposure without a matching accuracy benefit.

Don’ts

• Don’t rely on captchas alone. Modern bot networks defeat them, and your funnel will still fill with slop.
• Don’t let AI-written JDs ship without human bias review. Generated drafts often carry subtle adverse-impact language that humans can spot.
• Don’t store copilot hallucinations in the candidate record. Once saved, the text becomes part of a discoverable employment record.
• Don’t treat Colorado or California AI rules as future problems. Colorado SB24-205 is in force in 2026, and California AB 2930’s regulatory framework is live.
• Don’t assume remote-only reqs escape state law. If the role can be performed in NYC, Illinois, Colorado, or California, those statutes reach you.

Pros and Cons of AI in the ATS

A balanced view helps you decide which AI features to keep, kill, or gate.

Pros

• Faster candidate sourcing. AI-assisted outbound outreach can expand qualified pipelines without adding headcount.
• Better structured interviews. Copilots that generate rubric-aligned questions improve consistency, which is a bias control.
• Improved scheduling. AI schedulers reduce no-shows and recruiter busywork.
• Cleaner data hygiene. AI parsers, when paired with human review, can standardize resume data for analytics.
• Faster disposition notes. Draft-only copilots reduce the time recruiters spend writing rejection notes, which improves candidate experience.

Cons

• Hallucinated credentials. Copilots can invent facts that become part of the employment record.
• Disparate impact. Any AI-enabled selection tool carries Title VII exposure under the EEOC’s AI guidance.
• Audit cost. Bias audits, impact assessments, and legal review are real line items.
• Candidate mistrust. Applicants increasingly distrust AI-only screening, which hurts offer-acceptance rates.
• Vendor lock-in. AI features are hard to disable once embedded, and migrations are expensive.

Processes and Forms You Cannot Skip

Stopping AI-slop is also a paperwork exercise. There are four artifacts every TA team should maintain.

The Impact Assessment

An impact assessment documents the purpose, data inputs, known risks, mitigations, and owner of every AI feature in your ATS. California AB 2930 and Colorado SB24-205 require it. Use the NIST AI Risk Management Framework as your template.

The consequence of skipping it is the loss of a reasonable-care defense. In a mini-scenario, when a Colorado AG inquiry lands, the deployer with an impact assessment resolves the inquiry quickly, and the one without it absorbs months of discovery.

A common misconception is that an impact assessment is a one-time artifact. It must be refreshed when the AI feature changes, when the deployment context changes, or at least annually.

The Bias Audit

A bias audit measures selection rates across protected classes and tests whether your AI tool creates adverse impact. NYC Local Law 144 requires an independent audit, published on the employer’s website, updated annually.

The consequence of running a non-independent audit is that it does not count, and the fines start stacking. A mini-scenario: a vendor offers a “free audit” performed by its own data team; NYC DCWP rejects it, and the employer must redo the work.

A common misconception is that an audit of the vendor’s aggregate data satisfies your obligation. It does not, because the audit must reflect your specific deployment and your specific candidate pool.

The Candidate Notice

Candidate notices tell applicants when AI is being used, what it evaluates, and what rights they have. Illinois AIVIA and NYC 144 both require specific notices. California’s regulations go further, requiring notice and in some cases an opt-out.

The consequence of a missing notice is a private right of action in Illinois and a DCWP fine in NYC. A common misconception is that a single website banner covers all jurisdictions. It does not; each state and city imposes its own content requirements.

The Retention Record

Federal law requires one-year retention of employment records under 29 C.F.R. § 1602. California extends that to four years under FEHA regulations. AI-generated content in the ATS is part of the record and must be retained accordingly.

The consequence of early deletion is spoliation sanctions in litigation. A common misconception is that “AI notes” are informal scratch and can be purged. Once saved, they are records.

Comparison: Federal vs. State AI-Hiring Rules

Recap of Key Rulings and Settlements

A few decisions shape how courts and agencies think about AI-slop. The EEOC v. iTutorGroup settlement confirmed that algorithmic age-screening creates Title VII liability. Mobley v. Workday is the leading federal case on whether an ATS vendor can be a covered “agent” of the employer under Title VII, with a 2024 order allowing the case to proceed.

The Illinois AIVIA itself does not yet have a headline ruling, but plaintiffs are filing cases in 2025 and 2026 testing its private-right-of-action potential. The NYC Local Law 144 regime has not produced a major ruling, but DCWP investigations are active. On the federal front, the FTC’s 2023 guidance on AI claims has been used to challenge vendor marketing that overstates bias-freeness.

The practical takeaway is that regulators are not waiting for a single landmark case. They are using the tools they already have, which means your defensibility plan must be in place now, not after discovery opens.

Key Entities You Need to Know

• EEOC: federal enforcer of Title VII, ADEA, and ADA; issues AI guidance and investigates charges.
• NYC Department of Consumer and Worker Protection: enforces Local Law 144 AEDT rules.
• Illinois Department of Labor: administers aspects of Illinois AIVIA and related employment rules.
• Colorado Attorney General: enforces SB24-205 against high-risk AI deployers.
• California Civil Rights Department: enforces FEHA and the automated-decision regulations under AB 2930.
• OFCCP: oversees federal-contractor recordkeeping and nondiscrimination, including for AI-assisted decisions.
• NIST: publisher of the AI Risk Management Framework widely used for impact assessments.
• ATS Vendors: Workday, Greenhouse, Lever, iCIMS, Ashby, Bullhorn, SmartRecruiters, and Oracle Taleo.

FAQs

Is it legal for candidates to use ChatGPT to write their resumes?

Yes. There is no U.S. law banning AI-assisted resumes, but employers may set policy. You can require disclosure or rescind offers for misrepresentation, provided your policy is documented.

Can I reject a candidate solely because their resume looks AI-written?

No. A resume-authorship guess is not a lawful selection criterion and can create disparate-impact exposure. Evaluate verifiable skills, references, and work samples instead.

Does the EEOC require a bias audit for every AI tool?

No. The EEOC strongly recommends testing for adverse impact, but only NYC Local Law 144 imposes an explicit independent-audit mandate. Federal law still holds you liable for outcomes.

Are recruiter-copilot summaries considered employment records?

Yes. Once saved in the ATS, copilot summaries fall under 29 C.F.R. § 1602 retention rules and are discoverable in litigation.

Does NYC Local Law 144 apply to remote-only roles?

Yes. If the role can be performed in NYC, the AEDT rules apply. This sweeps in most remote-friendly employers that accept New York applicants.

Can I use a deepfake-detection tool without candidate consent?

Yes. Liveness checks are generally lawful as anti-fraud controls, but Illinois AIVIA requires disclosure when AI evaluates the video substantively. Document your use case carefully.

Does Colorado SB24-205 apply to small employers?

Yes. The statute targets deployers of high-risk AI systems and does not carve out small employers. Document reasonable care through an impact assessment.

Is the vendor liable if its AI tool discriminates?

No. Under Title VII, the employer is the covered entity. Vendor indemnities help commercially but do not shift regulatory liability.

Can AI-written job descriptions create legal exposure?

Yes. Gendered or ageist phrasing can produce disparate impact. Human bias review of every generated JD is the minimum standard.

Do I have to tell candidates I used AI to rank their resume?

Yes. In NYC, Illinois, and California, specific notice rules apply. Outside those jurisdictions, notice is best practice and reduces litigation risk.

Will a captcha stop bot-submitted applications?

No. Modern bot networks defeat basic captchas. Layer device fingerprinting, rate-limiting, honeypot fields, and email verification to meaningfully reduce bot volume.

Is an AI-generated rejection email legally risky?

No. Rejection emails are low risk, provided they do not reveal protected-class reasoning or invent facts. Human review of templated language is still recommended.