AI in Hiring. AI in Content. Oh My...

AI is being used across nearly every stage of the hiring process, and it’s showing real benefits.

In applicant review, AI can review more candidates in less time than humans can, giving companies the ability to consider more people for an opportunity.

During interviews, AI can give candidates more flexibility to interview at hours that work for them instead of needing to take time off or accommodate a recruiter’s limited openings. Ribbon reported that nearly 1 in 4 of its AI interviews happen between 10p.m. and 2a.m. their local time.

More opportunity, more accommodation.

However, the lack of transparency around how AI is evaluating candidates throughout the process still creates skepticism.

During applicant review, how and why does AI reject candidates? How does it determine whether someone is qualified or not?

During an interview, what is the AI actually looking for? How is it determining fit outside of good or bad answers? Does it understand nuance and social cues the way a human would? If not, does that make the evaluation fair?

That skepticism is showing up in candidate behavior. Greenhouse reports close to 40% of candidates have walked way from hiring processes that involve AI.

Regardless of the answers, it’s important to remember that humans are still behind these systems. Companies decide what they want the AI to evaluate, what they believe makes someone qualified, how different criteria should be weighed, and how much influence the AI’s output has on the final decision.

While AI may be doing the evaluating, humans are still deciding what the evaluation should value.

Suddenly, snitching on potential AI usage in content is a feature.

LinkedIn, Substack, and Claude have introduced new features and system-level tools that expose or identify potentially AI-assisted published content.

LinkedIn has introduced a ‘Seems like AI Slop’ feature that allows users to report individual posts they believe are low quality, AI-generated content. The feature is supposed to help LinkedIn understand what users are identifying as AI slop and use that feedback to improve what gets distributed in people’s feeds.

The issue here is that just because a post seems low quality or AI-generated, doesn’t mean it is. Some people may prefer this, but overall, I think it can be abused and can lean toward being unfair. Especially to people simply trying to build an audience and find their footing online.

Substack has introduced a new ‘Scan for AI text’ feature that uses an AI detection tool called Pangram to estimate how much of a piece of content may have been written by AI. Those results are then shown to the writer’s readers.

Substack does allow writers to disable the feature. However, it still kind of snitches on writers who disable it because when readers click ‘Scan for AI text’, they’re told that AI detection has been disabled for that publication. Now you’ve potentially created skepticism from readers simply because a writer chose not to participate.

The issue with Pangram is that it’s not 100% accurate. Pangram reports high accuracy rates for its own detector, but those numbers come from Pangram’s own research. Independent research on AI detection has found limitations on these tools, especially as models, writing styles, and editing techniques change.

Personally, I turn this feature off. Regardless of whether it’s writing, coding, or research, I use AI. I’ve broken down how I use AI in this publication before. I write, then give it to an LLM to help make it clearer and more concise. I’m all about AI assistance, not AI doing all the work for me.

Then there’s Claude’s new watermarking feature, where hidden, machine-readable marks are embedded into text generated by Claude models. The watermark is designed to survive things like copying and pasting the text somewhere else. The watermark isn’t something that can be found by highlighting a paragraph or inspecting the page, it’s built into the generated text itself.

Anthropic implemented this in response to the EU’s new AI transparency requirements, but decided to roll this out globally rather than only in Europe.

The purpose isn’t to declare that anyone who uses Claude to produce content is a fraud. People use Claude in all kinds of ways, including to edit or improve content they originally created. The watermark is meant to provide a way to identify that Claude was involved.

What makes me curious is where this goes in the future once other companies can use that information. A watermark could eventually become another indicator used by publishers, platforms, schools, employers, or other organizations to identify undisclosed AI use or to even investigate potential fraud. This could potentially change how comfortable people are using AI to help create content in the first place.