Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Writing Your Job Descriptions Now
Transcript
- Lucas: So there's this staffing company called EmployStream that processes about 3,000 job descriptions a week for warehouse and light industrial roles. And as of about four months ago, they stopped having humans write most of them. Luna: Wait — they just let AI take over the copy for actual job postings that real people see? Lucas: Yeah. They trained a model on their own archive — roughly 50,000 job descriptions that had historically performed well in terms of application rates. Now a recruiter inputs three things: job title, location, hourly wage range. The AI generates a full description in under 30 seconds. Luna: Fifty thousand is a pretty solid training set. But I'm wondering what 'performed well' actually means here. Did they optimize for applications filled out or for hires made? Lucas: That's the interesting part. They optimized for click-through rate — basically, the percentage of people who see the posting and actually start an application. Not for retention, not for quality of hire. Just raw conversion. Luna: Which means the AI is learning to write the equivalent of clickbait headlines for jobs. Lucas: Exactly. So you end up with language that's overly positive — 'fast-paced, fun environment' for a job that's mostly standing on a concrete floor for nine hours. The AI doesn't know the actual working conditions. It just knows that certain phrases correlate with people clicking 'apply'. Luna: Does EmployStream disclose that the description was ai generated? I'd want to know if I'm applying based on language written by a model that's never set foot in a warehouse. Lucas: They don't label it. And that's becoming a legal question. The EEOC put out guidance in 2023 saying that employers are still liable for discriminatory outcomes even if an AI tool wrote the job ad. So if the model learns to favor certain gender-coded language — like 'aggressive' or 'dominant' — that could create a disparate impact claim. Luna: Right, the 'ninja' and 'rockstar' problem. Those terms had documented bias issues long before generative AI showed up. Lucas: And the AI can amplify them. One experiment I saw — a researcher fed a model a set of neutral job descriptions and asked it to 'improve' them. The AI introduced words like 'competitive', 'driven', 'assertive' — all coded as masculine in prior linguistic research. The researcher ran the same prompt three times and got slightly different biased phrasing each time. Luna: So the consistency problem cuts both ways. It's biased, but not consistently biased, which makes it harder to audit. Lucas: Precisely. And that's where a lot of HR teams are stuck right now. They want the efficiency gain — because writing a job description from scratch can take forty-five minutes to an hour. But they're not sure how to review AI output for bias without essentially rewriting the whole thing, which kills the time savings. Luna: Have any companies actually been sued yet over ai written job postings? Lucas: Not specifically. But there's a class action pending against a large retailer that used an AI screening tool — not the job description part, the resume filter. The plaintiff's argument is that the tool systematically excluded older workers. If that case sets a precedent, it's a straight line to job descriptions that use ageist language. Luna: Quick thing — and I realize this is a bit meta — but episodes like this one, where we really dig into a specific use case, that's exactly the kind of thing that makes this show possible. We don't run ads, and a small group of listeners helps fund it through buy me a coffee dot com slash fexingo. It's not a big ask — just mentioning that that's what keeps these deep dives ad-free and independent. Lucas: Yeah, and we really appreciate those who do. It lets us spend time on stories like this one instead of chasing sponsors. So thank you. Luna: Anyway — back to the job descriptions. You mentioned that some companies are trying to build their own guardrails. What does that look like in practice? Lucas: A few approaches. One is to fine-tune the model on descriptions that have been vetted by legal and HR for inclusivity. So instead of training on all historical postings, you curate a set of 'exemplary' ones. Another is to add a second AI that reviews the first AI's output for problematic language. Luna: An AI to check the AI. Classic automation stack. Lucas: Right. And some vendors are doing that. Textio, for example — they've been in the augmented writing space for years, pre-generative AI. They now offer a layer that flags biased or unclear language in real time. But the recruiter still has to decide whether to accept the AI's suggestion or override it. Luna: Which puts the human back in the loop, which is probably where they should be. But then you lose some of the speed advantage. Lucas: Yeah. So the trade-off is real. And it might be worth it — because job descriptions are effectively the front door of your company. If the door is written by a model that doesn't know what the actual job is like, you attract the wrong people or you mislead the right ones. That leads to higher turnover and lower morale. Luna: Is there any data on whether ai written descriptions actually perform better than human-written ones in terms of hiring outcomes? Lucas: Early results are mixed. One study from a recruiting analytics firm found that ai generated descriptions had a 12 percent higher click-through rate but a 7 percent lower completion rate — meaning more people started applying but fewer actually finished the application. That suggests the AI might be overselling the role and then disappointing candidates once they see the details. Luna: So you get more leads but worse conversion. That might not actually be better for the recruiter. Lucas: Exactly. And if you're paying per applicant or per start, you might be spending more money for lower-quality candidates. Some companies are now A/B testing human vs. AI descriptions for the same role in different markets. I've seen preliminary results where the human-written descriptions produce higher retention at 90 days. Luna: Which is the metric that actually matters. Not clicks, not applications — retention. Lucas: Right. And the AI can't optimize for that unless you train it on retention data. But retention data is much harder to collect because it involves linking the job description to employee outcomes months later. Most companies don't have that data pipeline in place. Luna: So we're back to the classic AI problem: garbage in, garbage out, but with a glossy language model on top. Lucas: Yeah. And I don't want to sound overly critical — the technology is genuinely useful for saving time on boilerplate. If you're posting 3,000 warehouse job descriptions a week, like EmployStream, the efficiency gain is real. The concern is that companies adopt it without auditing what the AI is actually producing. Luna: What about smaller companies? Are they using these tools too? Lucas: Increasingly. The big HR platforms — Workday, SAP SuccessFactors — are adding generative AI features directly into their job description modules. So a small business that uses Workday might soon have ai generated descriptions as a default setting. They won't have the resources to run bias audits like a large enterprise might. Luna: That's a bit concerning. The compliance burden shifts onto the vendor, but the liability stays with the employer. Lucas: Exactly. And the vendors are adding disclaimers — 'this content is ai generated and should be reviewed before use' — but in practice, how many busy HR managers are going to carefully review every line of a description for 50 roles? Luna: Not many. Especially if they're already short-staffed. So the AI becomes the de facto writer, and the human becomes a rubber stamp. Lucas: That's the risk. And one more thing — job descriptions are also used internally for promotions and transfers. If the AI writes those too, you could have a situation where internal candidates are steered away from opportunities because the language doesn't match their resume keywords. Luna: So it's not just an external hiring issue. It affects current employees' career paths. That's a much bigger surface area for potential bias. Lucas: Right. And I think that's the part that's not getting enough attention. Everyone's focused on the flashy part — AI writing job ads — but the downstream effects on internal mobility and pay equity are where the real impact could be. Luna: So what would a responsible adoption of this technology look like? Lucas: I'd say three things. One: train the model on data that includes outcomes — not just click rates but retention and performance. Two: have a human review process that's actually feasible, maybe a checklist of flagged terms that the reviewer can approve or reject quickly. Three: audit the output regularly for demographic bias, and be transparent with candidates that the description was ai assisted. Luna: That last one feels like the simplest and the least likely to happen. Companies don't want to admit they're using AI because it might make them seem impersonal. Lucas: Yeah. But ironically, being upfront about it could build trust. If a candidate knows the description was drafted by AI but reviewed by a human, they might actually feel more confident that the company is being thoughtful about its process. Luna: That's a good point. Transparency could be a differentiator in a tight labor market. Lucas: Exactly. And some companies are starting to do it. Buffer, the social media management platform, has a public AI policy that covers job descriptions. They state clearly when AI is used and how the output is reviewed. It's early, but it's a model others could follow. Luna: So we're at this inflection point where the technology is good enough to be useful, but not trustworthy enough to be left alone. And the next few months of court cases and EEOC guidance will probably shape how it evolves. Lucas: Yeah. I think by the end of this year, we'll see either a major lawsuit or formal regulatory guidance that changes how companies approach ai written job descriptions. And then the conversation shifts from 'should we use it' to 'how do we use it compliantly.' Luna: And hopefully with an eye toward fairness, not just efficiency. Lucas: Right. Because a job description is often the first impression a candidate has of a company. If that impression is generated by an algorithm that doesn't understand the work, both the company and the candidate lose. Luna: It's a reminder that the tools we build reflect the priorities we set. If we prioritize speed over accuracy, we get fast but flawed output. If we prioritize fairness, we have to build that into the training data from the start. Lucas: Well said. And on that note, I think we'll keep an eye on this space. There's a lot more to come.