Are we being too uncritical about AI research in employer branding?

I attended EB Uncut this week, the brilliant Unconference organised by EB Space (i.e. largely by the awesome Claire de Souza).

There was a great range of topics to discuss, and some remarkable diversions from script, leading us all to consider our own mortality. But, you know, in a good way.

Naturally, AI was a sub topic. In the in-house branding and agency roles that everyone in the room was from, the possible short-cuts and benefits that AI affords us cannot be ignored.

AI double standard?

On the one hand: AI was rejected for producing creative expression of the employer brand. It flattens, it sameifies. It produces less, not more, distinctive work.

And that means you fail the first test of effective employer branding. Not whether people believe you, not whether it educates or changes minds. But simply whether it stands out.

In a world of infinite choice, would I choose to engage with that? The answer, if it is AI generated is clear: No, why would I?

On the other hand: there was a real confidence in using AI for research. The confidence to me felt surprising. Especially when the acknowledged effect of that was to remove the research from the role of the brander or the agency / consultant that supports them. And therefore to take away their depth of understanding.

Of course, I will feel that diminution of the research role harder. It’s equally clear that if we are dealing with a large volume of, basically, language, it’s natural to have confidence in a large language model to deal with it.

Are we too uncritical of AI Research?

What I didn’t detect was the same critique of AI in the research use case.

If the creative outputs are same-y and indistinct, why would one not assume that the factual insights won’t come out in just the same way? Because that’s what I see.

Obviously, I use AI as part of my research processes. I’d be a fool not to.

Immediately on the launch of ChatGPT, I saw the potential. Initially for doing qualitative research at real scale. Once web access became more practical and reliable, AI could go to many more sources of information, far more quickly, than I could hope to as a human.

AI Limitations

I also noticed the AI limitation early. I had written a detailed report where AI helped me find the source material, but the reading, interpretation and judgement were mine. It was good work. Then the report needed a summary of summaries. I asked AI to do that bit.

When I read it, I thought it had done a decent job. But that was because I already knew the source material. I could fill in the blanks. To a client reading it cold, it read well and said very little. That is one of the dangers.

AI can produce something that feels like a finding because it is fluent. But fluency is not evidence.

Another time, I had a report with some striking ideas. There was one thought in particular that felt interesting and potentially valuable. But I could see the question coming: “What triggered you to think that?” And I realised I could not properly unengineer the route back to the evidence. That is also a problem.

Not because the idea was necessarily wrong. It may have been right. But if you cannot show how you got there, you weaken your own argument. Your conclusions come across as guesswork.

Why AI doesn’t always work for you

Of course, much of the problems I’ve described can be addressed with better prompting. But there still remains the chance that AI will do one of five things.

The first two we’re familiar with:

· Fabricated facts: specific details (numbers, dates, names, percentages) that sound precise but were never retrieved from anywhere.

· Fake citations: a source that looks real: author, journal, year, even a DOI. But the paper doesn't exist.

But we also need to be aware of these other three:

· Confident nonsense: Making up a whole answer, not just one fact. AI creates all the parts of an answer at the same time, so they fit together neatly. The neat fit makes the answer look true.

· Knowledge boundary errors: Answering questions about a reality it cannot possibly know, and so pivots from the specific ask to a generic industry default, or even something entirely irrelevant.

· Maths and logic failures: The facts may be correct, but the reasoning that joins them does not hold: an absolute conclusion from partial evidence, or arithmetic that, well, doesn’t add up.

The key idea here is that AI often generates plausible text, not verified truth. The risk is we mistake fluency as evidence of accuracy.

Fluent = samey

Employer brand research is not just about finding the average employee experience. It is about finding the texture. The tensions. The contradictions. The throwaway comment that explains everything. The minority view that reveals the culture more clearly than the majority answer. The phrase someone almost uses before correcting themselves.

That is where the useful fodder lives. AI is not naturally drawn to that. It is drawn to the centre. That means you get the average answer. The tidy answer. The answer that sounds like the right answer.

The box, not the whiskers. The plant, not the flowers.

And that creates a strange risk. The very thing we reject in AI creative -the sameness, the smoothing, the loss of distinctiveness - can creep into the research as well.

Only this time, it is harder to spot. With creative work, blandness is visible. With research, blandness can look like confidence.

That is where I think the caution needs to be. Not “don’t use AI”. That would be counter-productive. Use it. But do not outsource judgement to it.

AI-Human Hybrid

What I encourage everyone to think about is a hybrid approach.

AI can organise transcripts or other sources, it can suggest themes, challenge assumptions and find gaps. But the judgement ought to be yours and the final interpretation absolutely should be.

AI can accelerate your findings, but it should never be the final authority. A summary tells you what appears to be there. An insight tells you what matters. They are not the same thing.

How to safeguard yourself

Raw evidence needs treating carefully. If it’s core to your case, it needs to be read by a human too, carefully, to avoid any misinterpretation.

Quotes can be difficult. AI can quote what people meant to say, or the gist of a whole interview – but it may not be a verbatim. Check with care.

Data too is often misinterpreted or mangled. I’ve had wonderful successes using AI to support with proper data analysis of huge data sets. And I’ve equally had it give me a response rate that is plausible, rather than accurate. You need QA as well.

Be careful about what’s a summary and what’s an insight. AI outputs should be treated as potential avenues, not your final destination.

Conclusion

The danger is that if we put too much trust in AI as a research tool, it will make our understanding of people feel more certain than it really is. And that will diminish our employer brands.

Summary: A balanced argument for using AI in employer brand research cautiously: AI can speed up analysis, but it can also flatten nuance, overstate confidence, invent evidence and turn distinctive human insight into generic summary. This connects with live employer brand debate about how AI reads and represents employer content.

Questions this article answers:

1. Should employer brand teams use AI for research?

2. What are the risks of AI research in employer branding?

3. Why can AI-generated insights feel convincing but be weak?

4. How should agencies use AI without losing human judgement?

5. What safeguards make AI-supported research more reliable?