A generative AI tool can draft a termination letter that cites a labour code section that does not exist, summarise a candidate’s background with an award they never received, or answer an employee’s leave query with a policy the company never wrote. The output reads clean, sounds authoritative, and is wrong.
This failure mode has a name: hallucination, the tendency of large language models to produce fluent, confident text that is factually false or entirely invented.
Indian HR has adopted these tools faster than almost anyone. India leads the world in workplace generative AI use, with 92% of Indian employees using AI at work several times a week against a global average of 72%, per Boston Consulting Group’s AI at Work 2025 report. That speed cuts both ways. When a recruiter, an HR business partner, or an employee-facing chatbot leans on a tool that fabricates, the error does not stay contained in a draft. It reaches a candidate, a payslip, a disciplinary file, or a regulator.
Why Language Models Invent Facts
Hallucination is not a bug that a patch will fix. It sits in how the technology works. A large language model predicts the most statistically probable next word given everything before it, and probability is not truth. The model has no internal fact-checker and no concept of “I don’t know.” When it lacks a real answer, it generates a plausible one with the same confidence it uses for a correct one.
Researchers have shown the problem cannot be fully engineered away. In 2024, computer scientists Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli formally proved that hallucination is inherent to large language models and cannot be completely eliminated. A separate paper reached the same conclusion through a different mathematical route. The realistic goal is mitigation and traceability, not a hallucination-free tool, and any vendor promising the latter is selling something the math does not support.
The rates are not trivial. Even legal AI tools built specifically for accuracy using retrieval-augmented generation still hallucinate between 17% and 34% of the time on benchmark queries, according to Stanford’s Human-Centered AI Institute. General-purpose tools of the kind most HR teams actually use offer weaker guarantees, and the errors cluster in exactly the rare, high-stakes cases where human judgment matters most.
The exposure is heaviest precisely where adoption is fastest. Indian workers use these tools more than any other national workforce BCG surveyed, which means a Wipro recruiter, an HDFC Bank HR partner, or an Infosys shared-services team is more likely than a global peer to be running a first draft through a model that can invent a fact without signalling that it has. High usage without matching verification habits turns a technical quirk into an operational one.
Where Hallucinations Bite In The HR Function
The damage is worse where an HR decision touches a person’s livelihood, a legal obligation, or a permanent record. AI in Indian HR has moved from pilots to embedded infrastructure, with the Capterra India 2025 HR Software Trends Survey reporting that 72% of Indian organisations have integrated AI features into their HR software. The wider the surface, the more places a fabricated fact can enter unnoticed.
Hiring and Screening
Résumé-parsing and candidate-summarisation tools condense long documents into short verdicts, and condensation is where invention creeps in. A model can attribute a skill, a certification, or a past employer that the résumé never mentioned, or misread an ambiguous date into a fabricated tenure. A recruiter working a high-volume TCS or Infosys fresher pipeline who trusts the summary over the source document can reject a qualified candidate or advance an unqualified one on invented evidence.
The candidate-facing side carries its own risk. Recruitment chatbots that answer questions about roles, compensation bands, or notice periods can state figures no one approved. When the chatbot speaks in the company’s name, the company owns what it says, a principle Indian HR teams should weigh before pointing one at applicants. A structured, source-grounded approach to AI in recruitment keeps the tool assisting the decision rather than making it.
Compliance and Legal Drafting
Legal drafting is the sharpest edge because fabricated law looks exactly like real law. A model asked to cite provisions of the Industrial Relations Code, 2020, or obligations under the POSH Act, 2013, can produce section numbers, clause language, and case references that are entirely invented.
The most cited cautionary tale is corporate: in 2025, Deloitte Australia refunded part of an AU$440,000 government contract with the Department of Employment and Workplace Relations after a report it delivered contained fabricated academic references and a made-up quote attributed to a judge. The same failure applied to an Indian termination letter, a POSH inquiry finding, or a compliance filing produces legal exposure, not just embarrassment.
Employee-Facing Answers
HR chatbots handling leave balances, benefits eligibility, or grievance procedures are trusted by employees precisely because they carry the employer’s authority. A hallucinated answer about maternity entitlement under the Maternity Benefit Act, 1961, or about EPF withdrawal rules, misinforms someone making a real decision about their money or their family. The employee acts on it, and the correction, if it comes, comes late.
The Indian Regulatory Frame
India now has a governance framework that speaks directly to this risk, and it puts the obligation on the deployer. On 5 November 2025, the Ministry of Electronics and Information Technology unveiled the India AI Governance Guidelines under the IndiaAI Mission, built on seven principles including “People First” and accountability. Human oversight and final human control over AI systems sit at the centre of the framework, which means an HR team cannot treat a tool’s output as a decision-maker in its own right.
Data protection law compounds the duty. The Digital Personal Data Protection Act, 2023 governs how candidate and employee personal data is processed, and an AI tool that fabricates or misattributes information about a data principal creates accuracy and rights problems on top of the operational error. As such, certain obligations sit together for Indian HR:
| Obligation | Source | What It Means For HR |
| Human oversight and accountability | India AI Governance Guidelines, 2025 (MeitY) | A person, not the tool, owns every AI-influenced HR decision |
| Accuracy and lawful processing of personal data | DPDP Act, 2023 | Fabricated candidate or employee data is a compliance failure, not just a mistake |
| Fairness and non-discrimination | India AI Governance Guidelines, 2025 (MeitY) | Hallucinated or biased outputs that affect hiring must be checked and correctable |
Employee trust tracks these gaps closely. Confidence in workplace AI remains uneven in India even amid record adoption. As such, understanding why Indian employees still don’t fully trust AI is part of deploying it responsibly, not just quickly.
Building Guardrails That Hold
No control removes hallucination, so the working model is containment: assume the tool will sometimes fabricate, and design the workflow so a fabrication cannot reach a person or a record unchecked. Several practices do that work without slowing HR to a halt.
Keep A Human On Every Consequential Output
A person verifies before anything AI-drafted becomes a decision, a message, or a filing, and the more the output affects someone’s livelihood or legal standing, the harder that check. Résumé summaries get read against the résumé. Drafted letters get read against the actual law. This is the practice the MeitY guidelines encode, and it is the single most reliable defence.
Ground Tools In Real Sources And Demand Citations
A tool that answers from a controlled document set, and shows which document each answer came from, is far easier to catch out than one generating from open-ended memory. Where an HR chatbot cannot show its source, its answer gets treated as a draft, not a fact. The black box problem and the case for explainable AI in Indian HR is the same argument from a different angle: an output you cannot trace is an output you cannot trust.
Audit Before Trusting, Then Keep Auditing
Bias and accuracy in an HR tool get tested before it touches real candidates and re-tested on a schedule, because model behaviour drifts. A documented process for running an AI bias audit for Indian HR doubles as a hallucination check when it includes accuracy sampling against known-correct answers.
Write The Policy Down
A tool used without written rules is a tool used inconsistently, and inconsistency is where errors slip through. A clear internal policy on which tasks may use AI, which may not, and who signs off turns individual caution into an organisational standard. An AI policy is now non-negotiable for the modern workplace, and hallucination risk is a core reason.
In The End…
One audit this quarter is enough to establish a baseline. The tool that drafts letters, screens candidates, or answers employees on law and money is the one to test, using ten queries where the correct answer is already known, a few of them deliberate edge cases.
The rate at which the tool invents a fact, a figure, or a citation is the number that matters, and it maps directly to which outputs currently need a human check before they leave the building.
The rule follows from that number. It names the HR tasks where an AI draft may go out only after a person verifies it against the source, and it names the person who signs. India adopted these tools first and fastest.
The teams that keep their advantage will be the ones that pair that speed with the discipline to check what the machine hands them, because a fabricated section of the POSH Act reads exactly like a real one until someone looks it up.
FAQs
What is AI hallucination in HR?
AI hallucination is when a large language model produces fluent, confident text that is factually false or entirely invented. In HR, this can mean a generative AI tool citing a labour code section that does not exist, attributing a credential a candidate never earned, or answering an employee query with a policy the company never wrote.
Can AI hallucination be eliminated?
No. In 2024, computer scientists Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli formally proved that hallucination is inherent to large language models and cannot be completely eliminated. The realistic goal is mitigation and traceability, not a hallucination-free tool.
How often do AI tools hallucinate?
Even legal AI tools built specifically for accuracy using retrieval-augmented generation still hallucinate between 17% and 34% of the time on benchmark queries, according to Stanford’s Human-Centered AI Institute. General-purpose tools that most HR teams use offer weaker guarantees.
Where does AI hallucination pose the biggest risk in HR?
The risk is heaviest where an HR decision touches a person’s livelihood, a legal obligation, or a permanent record: hiring and candidate screening, compliance and legal drafting, and employee-facing chatbots answering questions about leave, benefits, or grievance procedures.
What do Indian regulations require for AI use in HR?
The India AI Governance Guidelines, released by MeitY on 5 November 2025, place human oversight and final human control at the centre of AI use, meaning a person must own every AI-influenced HR decision. The Digital Personal Data Protection Act, 2023 adds an obligation for accuracy and lawful processing of candidate and employee personal data.
