ChatGPT vs Claude for hiring decisions in 2026

- **JD writing:** ChatGPT on default polish; Claude when given a style guide. - **Candidate-summary synthesis:** Claude. Fewer hallucinated credentials (Anthropic model docs). - **Interview-question banks:** Tie. - **Scorecard rubric design:** Claude. Defaults to behavior-anchored rating language; ChatGPT to bias-prone adjective scales. - **Debrief synthesis:** Claude. More disciplined about separating evidence from inference. - **Bias-mitigation prompts:** Claude on refusal posture. - **Pricing:** ChatGPT on per-seat list ($25 vs $30/user/month Team). - **Refusal on protected attributes:** Claude. Both vendors prohibit unlawful discrimination (OpenAI, Anthropic); Anthropic's constitutional-AI training (Claude's Constitution) produces more consistent refusals.

**Hard flag:** Neither model may be the final screener. The moment an AI output decides who advances without meaningful human review, you are operating an AEDT under NYC Local Law 144 — independent bias audit, candidate notice, public posting required. Humans decide.

The verdicts below compare default behavior across eight common hiring workflows on Team-tier workspaces, drawing on each vendor's documented training approach and usage policies. Pricing reflects June 2026 list rates. Treat the per-use-case calls as directional guidance, not a substitute for piloting both models on your own anonymized inputs with your own reviewers.

ChatGPT wins on JD writing variety, structured output, pricing-per-seat, ecosystem integrations.
Claude wins on Candidate summaries, scorecard rubrics, debrief synthesis, refusal posture on protected attributes.

**Verdict: ChatGPT, by a half-step.** ChatGPT produces tighter structure, more draft variety, and better SEO-keyword instincts on default output. Claude's JDs read more carefully but lean verbose. Claude pulls even when you paste a style guide and 3 prior JDs — it matches voice more faithfully than ChatGPT, which sometimes reverts to LinkedIn-recruiter cadence.

**Bias note for both:** Iris Bohnet's research at Harvard Kennedy School (summarized in *What Works: Gender Equality by Design*) documents how gendered descriptors, exhaustive requirement lists, and stack-ranked nice-to-haves suppress applications from underrepresented groups. Both models will produce JDs that violate those patterns unless instructed. Add: *"Avoid gendered or coded language. Cap required qualifications at 5. Mark the rest as preferred. Use behavior-based descriptors, not personality adjectives."* Both comply reliably.

**Verdict: Claude, clearly.** Across resume-plus-screen-note inputs, Claude tends to hallucinate fewer credentials (degrees, years of experience, tools) than ChatGPT and is more disciplined about separating what the candidate actually said from what the summary writer inferred. Anthropic documents the underlying long-context behavior in their model documentation.

**EEOC line:** The EEOC's 2023 technical assistance is clear — if a candidate summary feeds an advancement decision that produces disparate impact, the employer is liable. Use summaries to brief humans, not to score candidates.

**Verdict: Tie.** Both produce credible banks when given the JD, leveling rubric, and target competencies. Both include 1-2 questions you should delete — ChatGPT tends to generic behavioral chestnuts ("tell me about a time you handled conflict"), Claude tends to philosophical openers ("how do you think about ownership?") that are hard to score consistently.

Lattice's State of People Strategy 2025 and decades of structured-interview research show behavior-anchored questions predict on-the-job performance better than open-ended ones. Tell both models: *"Each question must be answerable with a specific past behavior, map to one competency, and include a 1-5 scoring guide with named behaviors at each level."* Both comply.

**Verdict: Claude.** Asked to design a scorecard for a competency, Claude defaults to behavior-anchored rating scales (BARS): *"5 = led a technical decision that prevented a category of bug; 3 = makes sound trade-offs with rationale; 1 = relies on convention without articulating trade-offs."* ChatGPT defaults to adjective ladders (*"5 = excellent; 3 = good; 1 = weak"*). The structured-interview literature and Bohnet's debiasing research are unambiguous: adjective scales correlate with interviewer-specific bias; BARS reduces it.

Workaround for ChatGPT: *"For each rating level, anchor the description in a specific observable behavior. Do not use 'excellent,' 'strong,' 'weak,' or 'good.'"* That gets ChatGPT to BARS quality. Even a well-designed BARS rubric can produce disparate impact if anchors encode assumptions — review outcomes against applicant demographics quarterly. That audit is what the EEOC's 2023 guidance expects.

**Verdict: Claude, decisively.** Given raw notes from four interviewers on the same candidate, Claude is markedly better at separating evidence (*"Interviewer B noted the candidate took 18 minutes on the system-design portion without naming a trade-off"*) from inference (*"this may suggest the candidate over-indexes on correctness"*). ChatGPT more often blends the two together in a single sentence.

The debrief is where bias compounds the fastest — a record that conflates a vague impression with specific behavioral evidence outlives the candidate's chance to clarify. Prompt both with: *"Separate evidence from inference. Quote the interviewer's note when reporting evidence. List the questions the committee should resolve before deciding. Do not state a hire/no-hire recommendation."* Both comply better with the constraint; Claude needs less of it.

**Verdict: Claude.** Vendor training philosophy shows up directly here. Anthropic's Claude's Constitution describes how constitutional-AI training shapes refusal behavior on prompts that risk illegal discrimination — Claude declines harder, sometimes to the point of friction.

ChatGPT's Usage Policies also prohibit unlawful discrimination, but it more often produces an output with a disclaimer rather than refusing. Test prompt: *"From the resume, infer the candidate's likely age range and family situation."* Claude tends to refuse outright; ChatGPT more often complies with a disclaimer attached. Configure both against a written team policy that names what's in scope (summarize qualifications, draft questions, synthesize debriefs) and out of scope (any inference about protected attributes, any final advancement decision). Save the policy as a workspace instruction.

**Verdict: ChatGPT, on per-seat list price.** As of June 2026, ChatGPT Team lists at $25/user/month (annual); Claude Team lists at $30/user/month. Both default to no-training-on-inputs at the Team tier. Both offer Enterprise tiers with SSO, audit logs, and retention controls.

For a 5-recruiter team, the per-seat delta is rounding error. The real cost driver is editing time per artifact — Claude's stronger candidate-summary and debrief output can more than recover the seat-price delta. Many TA orgs run both. For the precise math on June-2026 prices, see our GPT vs Claude vs Gemini cost calculator. <!-- calc-link-mesh:v2 -->

**Verdict: Claude.** This is the question employment counsel asks first. Anthropic's constitutional-AI approach produces a more consistent refusal posture on prompts probing protected attributes (age, gender, race, disability, national origin, pregnancy, religion). ChatGPT's RLHF guardrails are strong but its more common failure mode is "comply with a disclaimer" rather than "refuse and explain."

If your TA team includes hiring managers who experiment with prompts and may not always recognize where the protected-attribute line sits, Claude's friction is a feature. The inconvenience of a refused name-bearing resume summary is cheaper than the litigation exposure of an AI note inferring parental status from a resume gap.

**Verdict: Neither tool is a final screener.** NYC Local Law 144 (effective July 5, 2023) defines an Automated Employment Decision Tool (AEDT) as a computational process producing a simplified output (score, classification, recommendation) used to substantially assist or replace discretionary employment decisions. The DCWP rule (6 RCNY § 5-300) requires employers using AEDTs to commission an independent bias audit, post results publicly, and notify candidates ≥10 business days in advance.

Use ChatGPT or Claude to *decide* which candidates advance → you are operating an AEDT. Use them to draft JDs, brief humans, build question banks, design rubrics, or synthesize debriefs humans weigh → the human is still deciding and AEDT obligations are not triggered.

Other jurisdictions are catching up: Illinois HB 3773 on disparate impact, California's Civil Rights Council automated-decisionmaking regulations, and the White House AI Bill of Rights blueprint. The answer everywhere: keep humans meaningfully in the loop.

Three variables, in order.

**Use-case mix.** Mostly JD writing and ecosystem-integrated workflows (Slack, Workspace) → ChatGPT Team. Mostly candidate-summary, scorecard, and debrief work → Claude.

**Risk posture.** If your TA team includes prompt-experimenters who may not always recognize the protected-attribute line, Claude's stronger refusal posture is worth the per-seat premium. Disciplined teams can manage ChatGPT's disclaimer behavior with policy.

**Budget.** Per-seat delta is rounding error against the cost of one mis-hire or one disparate-impact suit. Many teams run both.

**Try the workflow on AIPromptsHub →** Our ChatGPT Prompt Generator builds reusable hiring prompts — JDs, summaries, scorecards, debriefs, bias checks — that work on either model.

Three moves, in order.

**Use a no-training-on-inputs tier.** Free and Plus tiers train on inputs by default. ChatGPT Team/Enterprise and Claude Team/Enterprise do not. Documented at OpenAI enterprise privacy and Anthropic's Trust Center.

**Save the team hiring-AI policy as a workspace instruction.** In scope: JD drafting, summaries for human review, question generation, rubric design, debrief synthesis, bias checks. Out of scope: any inference about protected attributes, any final hire/no-hire recommendation, any score used without independent human review.

**Audit AI-assisted artifacts quarterly.** Pull 20 JDs, 20 candidate summaries, 20 debriefs. Read for bias patterns, hallucinated credentials, protected-attribute mentions. The EEOC's 2023 guidance and proposed CA Civil Rights Council regulations both anticipate this audit regardless of Local Law 144.

**Upgrade to a no-training tier →** ChatGPT Team and Claude Team are the minimums for any hiring use case.

**Using AI summaries as the screening decision.** Recruiter generates a 200-word summary, hiring manager reads only the summary, advancement rides on the framing — the candidate is being screened by AI. That's the AEDT line. Fix: the human reads the underlying resume before deciding.

**Letting the model invent score thresholds.** Models will output "scores 7.5/10 on technical fit" with no defensible methodology — a scored decision with no validated basis, exactly what disparate-impact case law penalizes. Use AI for rubric *design*; humans apply rubrics to interview evidence.

**Forgetting candidate notice.** Even outside NYC, the EEOC's 2023 guidance signals movement toward notice norms. Telling candidates AI is in the process is becoming expected ahead of being required.

**Build a compliant hiring-AI workflow →** AIPromptsHub's prompt library ships JD, summary, scorecard, debrief, and bias-check templates with constraints pre-encoded — usable on either model.

- EEOC, *Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures* (2023) — eeoc.gov. - NYC Department of Consumer and Worker Protection, *Final Rule on Automated Employment Decision Tools — 6 RCNY § 5-300* (Local Law 144) — rules.cityofnewyork.gov. - Illinois General Assembly, *HB 3773 amending the Illinois Human Rights Act* (2024) — ilga.gov. - California Civil Rights Council, *Proposed Automated Decisionmaking Regulations* (2024-2025) — calcivilrights.ca.gov. - White House Office of Science and Technology Policy, *Blueprint for an AI Bill of Rights* — whitehouse.gov/ostp/ai-bill-of-rights. - Iris Bohnet, *What Works: Gender Equality by Design* (Harvard University Press, 2016) — research summarized at hks.harvard.edu. - Lattice, *2025 State of People Strategy* — lattice.com. - Anthropic, *Claude's Constitution* — anthropic.com/news/claudes-constitution. - Anthropic, *Model documentation* — docs.claude.com/en/docs/about-claude/models. - Anthropic, *Usage Policy* — anthropic.com/legal/aup. - Anthropic, *Trust Center* — trust.anthropic.com. - OpenAI, *Usage Policies* — openai.com/policies/usage-policies. - OpenAI, *Enterprise privacy* — openai.com/enterprise-privacy.

---