Science News & Breakthroughs May 30, 2026

How ChatGPT and Claude Can Help HR Managers: Real Examples from the Workplace

Language models can speed up HR drafting, but screening and ranking candidates carries bias and legal risk. A researcher-style look at pilots, validation and accountability.

SC
ScienceTrace Science Desk
•
 6 min read
•
 1,185 words

Human-resources work is text-heavy: job descriptions, policy documents, candidate files, training plans, announcements. That makes it an obvious place to try language models, and a useful case study for anyone interested in how to evaluate such tools responsibly. HR is also a domain where errors affect people's livelihoods, so the questions researchers ask of any automated system, such as how it was validated, where it is biased and who is accountable, apply with particular force. This article separates the low-risk drafting uses from the high-risk screening uses, and describes how an organisation can pilot either one in a way that produces evidence rather than anecdotes.

Two kinds of task, two kinds of risk

It helps to sort HR tasks by what happens when the tool is wrong.

  • Drafting tasks (a job description, an announcement about a holiday schedule, a first outline of an onboarding plan) produce a document that a human reads before it is used. Errors are visible and correctable, and the cost of a mistake is usually small.
  • Evaluative tasks (ranking resumes, summarising candidates, flagging "outdated" policy sections) produce judgments that may be accepted without being rechecked. Errors are invisible by default and may fall unevenly across groups of people.

Most of the practical benefit is in the first category, and most of the legal and ethical exposure is in the second.

Drafting: where a model reliably helps

Asking for a job description for a data analyst with experience in Python, SQL and a business-intelligence tool yields a serviceable first draft in seconds. The value is in having a structure to edit rather than a blank page. The review step matters more than the generation step. Check that required qualifications are really required: models tend to inflate requirements by reproducing typical postings, and inflated requirements can discourage qualified applicants. Check that the language is not gendered or exclusionary, and that salary and benefits statements match what the organisation actually offers, because a model cannot know them.

The same applies to interview question banks. A request for behavioural questions for a customer-service role returns common questions such as "Tell me about a time you handled an angry customer." These are generic. Their quality as a selection instrument depends on structure: the research tradition in personnel selection has long found that structured interviews, with the same questions and a scoring rubric for every candidate, predict job performance better than unstructured ones (Schmidt and Hunter, 1998). A model can draft the questions; the organisation must still define the criteria and the rubric.

Document review: useful, but verify

Summarising a long employee handbook, or comparing it with a new regulation, is a reading task. A model can produce an index of sections and flag passages that look inconsistent. But a statement such as "this section is outdated" is a claim that needs a legal or policy reviewer to confirm, and the model has no knowledge of your jurisdiction's current employment law unless you supply the text. A sound procedure is to require the tool to quote the passage it is concerned about and name the document it compares against, then have a person confirm each flag. Count how many flags were correct; that ratio is the number you need to decide whether the process saves time overall.

Resume screening: the high-risk case

Uploading 200 resumes and asking for a ranking sounds efficient, but it is where the evidence is most cautionary. Automated screening systems learn or apply patterns from past hiring that can encode historical bias. A widely reported example is Amazon's experimental recruiting tool, which was reported by Reuters in 2018 to have been abandoned after it was found to penalise resumes associated with women. Reviews of algorithmic hiring tools have found that vendors' claims about bias mitigation are often difficult to verify from public information (Raghavan and colleagues, 2020).

General-purpose language models add their own problems. Output can vary between runs on the same input, so two identical resumes may be ranked differently. The criteria are implicit in the prompt and in the model's tendencies, so they are hard to audit. And a ranked list creates anchoring: reviewers who see a model's top ten tend to stop looking.

A pilot design that produces evidence

If an organisation wants to test assisted screening, treat it as a small study:

  1. Write the job-related criteria first, before the tool is involved.
  2. Take a past batch of applications with known outcomes, remove names and other identifying fields, and have human reviewers score them against the criteria.
  3. Run the tool on the same set, several times, and measure agreement with the human scores and agreement between runs.
  4. Compare outcomes across relevant groups where this can be done lawfully and ethically, and look for systematic differences.
  5. Decide in advance what level of agreement and consistency is acceptable, and what happens when the tool is wrong.

Many organisations will find that the tool is best used to extract structured facts (years of experience in a named skill, with the supporting sentence) rather than to rank people.

Privacy, law and accountability

Resumes contain personal data. Uploading them to a third-party service raises data-protection questions about consent, purpose, retention and cross-border transfer; check your jurisdiction's rules and the provider's terms. Regulation is also moving. The European Union's AI Act (Regulation 2024/1689) places AI systems used for recruitment and for decisions affecting employment among its high-risk categories, and New York City's Local Law 144 requires bias audits of automated employment decision tools used there. Under the GDPR, Article 22 restricts decisions based solely on automated processing that significantly affect individuals. Whatever the local rules, the principle is straightforward: a person must be able to explain and own every decision about hiring, promotion or discipline.

Limitations and open questions

  • Public, peer-reviewed measurements of how general-purpose language models behave when ranking real applications are limited, and results depend on prompt wording and model version, both of which change.
  • It is unclear how much time is saved once verification effort is counted. Savings in generating text may be offset by the time needed to check it.
  • Effects on candidates, such as applicants using models to write resumes that are then screened by models, are not yet well understood.

What practitioners and researchers should watch

  • Run-to-run consistency: the same input, ten runs, how much does the ranking move?
  • Sensitivity to irrelevant details such as names, formatting or university prestige.
  • Whether reviewers change their minds when they disagree with the tool, or merely defer to it.
  • Whether the documented workflow lets a rejected candidate or an employee obtain a meaningful explanation.

References and further reading

  • Schmidt, F. L. and Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin, 124(2).
  • Raghavan, M., Barocas, S., Kleinberg, J. and Levy, K. (2020). Mitigating bias in algorithmic hiring: evaluating claims and practices. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAT*).
  • Dastin, J. (2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters, 10 October 2018.
  • Regulation (EU) 2024/1689 (Artificial Intelligence Act).
  • Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22.
  • New York City Local Law 144 of 2021 on automated employment decision tools.
#ChatGPT #Claude #HR #Artificial Intelligence #Workplace #Human Resources
All Scientific Breakthroughs