Lauren McCluskey: good vs good enough

A self-funded study on AI moderation, the 30% that cheaper panels miss, and why the researcher has to be the guardian of meaning.

Lauren McCluskey is living and breathing good work, to the point that when she found herself with a lull in workflow last summer, she self-funded her own research project to understand whether AI moderation is suited to sensitive topics like women’s menopause journeys. She set out to study recruitment differences in AI-moderated work, but the human-versus-AI question kept surfacing through the data, giving new evidence to old intuitions.

Lauren has worked with AI-assisted analysis since the hallucination-heavy early days. Her study is informed by genuine curiosity and a non-adversarial attitude towards the technology, something she uses daily to elevate her work and deliverables, and which is particularly valuable for independents and smaller teams who can now reach a level and speed previously out of reach.

Is good work, good enough?

When I asked about her views on good work, and whether AI moderation feels like it, she invited me to think about what she identified as the latest buzzword: “good enough work.”

Clients have a standard of what is good enough, and that bar is very different for every client. What’s the client’s definition of good enough? Is good work different from good enough? What is the gap between those two?Lauren McCluskey

There’s a general consensus that if we can get “70% there,” it’s generally good enough for clients, especially if that lower bar is achieved through cost efficiency, at an eighth of the cost of traditional methods. Her study lets her examine good versus good enough on two levels: recruitment, which she designed the study to test, and moderation, which the data surfaced anyway.

Lauren isn’t anti-AI, but she has a clear theory for its application. For tactical concept work, panel AI moderation is definitely good enough. For emotional and strategically risky work, it is not. Her study straddles both: methodologically simple concept work, on an emotionally loaded topic that required high human involvement.

Research on research: same result, different depth

Lauren’s project does “research on research,” comparing recruitment styles rather than human-versus-AI moderation directly. She compared AI-moderated conversations with 101 participants from an automated recruitment platform against 28 “pure qual” participants sourced via a high-touch human recruiter, across three distinct menopausal life stages. The study used a sequential monadic design to present three concepts: territories with distinct imagery and taglines.

One of the headlines: panel and traditional qual netted out in largely the same place in terms of which concept performed best. But what sat underneath showed very different depth and quality of response. Cheaper panel recruitment falls squarely in the good-enough category. It got 70% of the way there, but lacked the 30% extra depth that more expensive pure-qual recruitment offered.

Anyone worth their weight, who’s been around, knows our research is only as good as the people we invite into it. There is a meaningful difference between panel sourced within an application and a higher-touch human recruiter.Lauren McCluskey

The panel provided the “what,” the winning concept, but the human-recruited group provided the much richer feedback needed to understand the nuance behind the preference. Those participants offered what Lauren calls “emotional breadcrumbs”: visceral, personal details that the AI was often unable to follow or probe further.

The moderator as guardian

The project confirmed some assumptions: AI moderation is closer to quant than to pure qual, and qual moderation is much more than asking the next question on the guide.

A good moderator stays alive to what’s happening in the room and makes small judgement calls in the moment. They notice when someone is contradicting themselves, when an emotional clue is worth following, or when the plan needs to be interrupted because something more interesting has appeared. It’s also about knowing when not to push too quickly. Sometimes the work is to sit with ambiguity long enough for people to reveal what they really mean.Lauren McCluskey

She’s also clear that the part of the lifecycle everyone overlooks is analysis.

People talk a lot about AI moderation, but not enough about what happens afterwards, in the analysis. You can unleash this technology on people, but then our role has to shift. Polished outputs can hide a lot of mess underneath. In that sense, researchers become a kind of guardian. We are there to protect the meaning, the context, and the integrity of what people actually said.Lauren McCluskey

If moderation is surrendered, some meaning is never generated. No amount of analysis can recover it.Lauren McCluskey