India’s hiring market in 2026 is a paradox. Application volumes are running roughly 40% above pre-pandemic levels, yet 54% of recruiters tell LinkedIn that fewer than half of what reaches them is relevant to the role on offer. Some of the noise comes from candidates spraying AI-written CVs; a quieter share comes from the job descriptions those candidates are responding to.
Across India Inc, generative AI now sits at both ends of the hiring funnel. The JD was produced by a model, the resume was shaped by another, and the match score comes out of a third. For a document that doubles as a legal commitment, a brand artefact and the first evidence in any future bias audit, the trade-off has gone largely unexamined.
The failure is not that AI writes bad prose. It’s structural. A probability engine, trained on a decade of existing job ads, flattens roles toward the median of its training data, imports the bias of that corpus, and produces text that is grammatically correct but factually underspecified. The resulting failures are already visible in Indian recruitment data when job descriptions and AI are combined with consideration.
The Adoption Curve That Made This Inevitable
AI in Indian hiring is no longer in pilot mode. The tools now sit across the full recruitment process, with 63% of recruiters already using them and roughly 80% planning to increase that use through 2026, according to LinkedIn research. JD drafting is one of the first places that use has landed, because it’s a bounded, repeatable writing task managers would rather not do themselves.
The practical workflow is now familiar. A hiring manager types a short brief into ChatGPT, Gemini or an in-ATS assistant. Within seconds, a JD comes back with sections for responsibilities, requirements, qualifications and perks. It reads well, covers the obvious ground, and almost never gets challenged by the recruiter who posts it. The speed is real. The quality control is largely absent.
The Model Writes What the Model Has Read
The sharpest critique of this workflow comes from a 2025 study by Sugat Chaturvedi of Ahmedabad University and Rochana Chaturvedi of the University of Illinois Chicago. The researchers ran more than 40 million simulations across several open-source LLMs using 332,044 real job postings from India’s National Career Services portal. Their headline finding: most models “reproduce stereotypical gender associations and systematically recommend equally qualified women for lower-wage roles.”
The bias traces back to the language of the JDs themselves. Male references, technology terms and money-related language in a description correlated with male callbacks; female references correlated with female callbacks. The models didn’t invent the pattern. They absorbed it from the Indian job-ad corpus and amplified it.
That corpus is the same training ground now generating most new Indian JDs through consumer AI tools. The bias doesn’t require a specialised hiring model to reach a candidate’s screen. It enters at the drafting stage, through text a human reads and typically doesn’t rewrite.
Where the Gender Signal Enters
The specific linguistic patterns driving the bias are documented and consistent across studies. The problem is that generic LLMs reproduce those clusters by default, which means they now sit inside most JDs drafted this quarter without any human having chosen them.
- Masculine-coded clusters: “aggressive,” “driven,” “dominant,” “confident,” “competitive,” “ambitious.”
- Feminine-coded clusters: “collaborative,” “supportive,” “interpersonal,” “nurturing,” “committed.”
- Requirement inflation: “ninja,” “rockstar,” “10x” framings that correlate with lower female application rates.
- Experience thresholds pulled from the median of past ads rather than from what the role actually needs.
The commercial consequence is specific. Women make up roughly 34% of India’s IT-BPM workforce, per Nasscom, with representation thinning sharply above middle management. A JD that leaks even mild male-coded language reduces the pool further. The ratio shifts slowly at the leadership level, as our analysis of women in Indian management documents.
Precision Without Accuracy with Job Description and AI
The second problem is quieter but shows up in the signal-to-noise crisis the LinkedIn data describes. AI-drafted JDs are grammatically crisp and structurally complete. They’re also, often, not accurate to the role as it will actually be done. The hiring manager’s brief gets re-expanded by the model into every adjacent responsibility the training data associates with that title.
A single prompt for “Senior Backend Engineer” returns a JD that stacks responsibilities drawn from a hundred past ads, not from the specific team doing the hire. Candidates read the inflated requirement list and either over-apply (if they can plausibly tick any box) or self-select out (if they can’t tick most). Either way, the pool that lands in the ATS doesn’t match the role.
| Common AI JD Distortion | What It Looks Like | Downstream Effect |
| Skill inflation | Lists every tool in the ecosystem rather than the three the role uses | Qualified candidates skip; unqualified ones apply |
| Experience padding | “5+ years” when 2 would do, because the median past ad said 5 | Shrinks the pool, raises CTC expectations |
| Hollow responsibilities | “Drive strategic initiatives” with no objective | Signals a role no one has designed |
| Generic perks block | Boilerplate culture statements | Weakens employer brand, repels senior candidates |
| Missing team context | No reporting line, team size, or stack | Removes the signal strong candidates screen for |
The fix isn’t to abandon AI. It’s to treat its output as a first draft that fails without a specific-knowledge pass. Guidance on what that pass looks like is covered in our walkthrough on writing an effective job description.
The DPDP Gap: Automation Doesn’t Dilute Employer Liability
The third problem is legal. India’s Digital Personal Data Protection Act, 2023, now in force after its rules were notified in 2025, treats AI processing of personal data as “processing” within the scope of the Act. It doesn’t, however, contain an equivalent of the EU GDPR’s Article 22, which gives individuals the right to meaningful information about solely automated decisions.
That gap matters for hiring. An Indian candidate rejected on the basis of a JD-generated screen cannot currently demand an explanation of how the model weighted their application. What they can do, through other routes, is challenge the employer. Indian courts and labour authorities have signalled that automation doesn’t dilute employer responsibility, and that principles of natural justice still apply to algorithm-driven workplace decisions.
Several existing statutes already reach into AI-generated JDs without an AI-specific law:
- The POSH Act, 2013, on language and treatment linked to gender.
- The Equal Remuneration Act, 1976, on wage discrimination that an AI-drafted JD could inadvertently signal.
- The Rights of Persons with Disabilities Act, 2016, on exclusionary language around physical requirements.
Penalties under the DPDP Act run up to ₹250 crore per contravention, based on the nature and severity of the breach. The exposure sits with the employer whose logo is on the JD, not the vendor whose model drafted it. A clear workplace stance on these tools, treated as policy rather than toolchain preference, is the baseline most Indian companies still lack. See our take on why an AI policy at work is now non-negotiable.
What’s Actually Broken Underneath
Each of the problems above traces to the same mismatch. A JD is a precision document. It’s a public, often legally binding statement of what work a person will do, what capability they need, and what the organisation will give them in return. LLMs are precision-adjacent probability engines. They generate the most statistically plausible next token given everything they’ve read. Those are different jobs.
The output of the model, left unchecked, trends toward the median of its training corpus. For Indian hiring, that median is a decade of job ads written in a labour market with documented gender imbalance, title inflation and generic prose. When the generated JD is posted without rework, the hiring process inherits all of it, plus a new layer of plausible-sounding specificity that was never ground-truthed against the actual role.
In the End…
AI-drafted JDs aren’t going away, and the honest editorial position is that most of them shouldn’t. The volume of hiring India Inc runs makes some degree of automation necessary. What has to change is what sits between the model’s output and the post button.
A workable discipline looks like this:
- Treat every AI-generated JD as a draft, never a final. Build a 10-minute human rewrite into the posting workflow, owned by the hiring manager and reviewed by HR.
- Run the draft through a gender-decoder tool before posting. Flag male-coded clusters and inflated experience thresholds.
- Document who signed off on the final version. Under the DPDP Act, employer liability lands on a named person, not a tool.
- Strip skill inflation. Keep only the capabilities the role will actually test for at the interview stage.
- Add the specifics the model can’t know: reporting line, team size, current stack, stage of the problem the role is solving.
- Keep a dated archive of every posted version. If a diversity audit asks later how a role was described, there will be a record.
The problem isn’t that recruiters are using AI. It’s that AI is being asked to do a job it isn’t built for, in a market where the downstream costs show up in diversity numbers, hiring quality and legal exposure. The teams that get the balance right in 2026 will be the ones that stop treating the model’s draft as a finished product.
FAQs
Does AI actually introduce gender bias into job descriptions?
Yes. A 2025 study by researchers at Ahmedabad University and the University of Illinois Chicago ran more than 40 million simulations across open-source LLMs using 332,044 real postings from India’s National Career Services portal. The models reproduced stereotypical gender associations and systematically recommended equally qualified women for lower-wage roles. The bias traces to the Indian job-ad corpus the models were trained on.
Is using AI to draft job descriptions legal in India under the DPDP Act?
Yes, but with caveats. The DPDP Act, 2023 treats AI processing of personal data as processing within scope. Unlike the EU’s GDPR Article 22, it doesn’t give candidates a right to meaningful information about solely automated decisions. However, the POSH Act, the Equal Remuneration Act, and the Rights of Persons with Disabilities Act all still apply to the language in a JD, regardless of who or what wrote it.
Who is liable if an AI-generated JD leads to a discrimination claim?
The employer, not the AI vendor. Indian courts and labour authorities have signalled that automation doesn’t dilute employer responsibility, and principles of natural justice still apply to algorithm-driven workplace decisions. Liability lands on a named person inside the organisation who signed off on the posted version.
What’s the financial exposure under the DPDP Act?
Penalties run up to ₹250 crore per contravention, based on the nature and severity of the breach. That exposure sits with the employer whose logo is on the job post.
Which words and phrases should I strip from an AI-drafted JD?
Masculine-coded clusters like “aggressive,” “driven,” “dominant,” “competitive,” and “ambitious” correlate with lower female application rates. So do framings like “ninja,” “rockstar,” and “10x.” Watch for inflated experience thresholds the model pulls from the median of past ads rather than from what the role actually needs, and generic skill lists that stack every adjacent tool instead of the three the job uses.
What’s the minimum workflow discipline for AI-generated JDs?
Treat every AI draft as a first pass, not a final. Build a 10-minute human rewrite into the posting workflow owned by the hiring manager. Run the draft through a gender-decoder tool. Add specifics the model can’t know, including reporting line, team size, current stack, and the stage of the problem the role solves. Document who signed off, and keep a dated archive of every posted version for future audit.
