At the end of 2023 I spent a week at AWS re:Invent, and AI was everywhere: every keynote, every booth, every hallway conversation. Responsible AI sat right at the front of it — an industry announcing guardrails almost as fast as it announced models. Fifty thousand people talking about the biggest technology shift in a generation.

I flew home thinking about a suffix.

Spanish can take almost any noun and make it small enough to love — pequeño becomes pequeñito, and the word stops describing size and starts describing affection. The piece I wrote on that flight argued that language is the most personal technology humans have, and that any model claiming to understand people has to capture what a suffix can carry.

The worry was concrete. Models inherit the biases of whatever they’re fed, and the era’s cautionary tale was a hiring tool trained on a decade of mostly male résumés that learned to prefer them. My conclusion looks almost quaint now: keep strong researchers working in tandem with the tools, because the human element isn’t optional.

Language is the most personal technology humans have.

So how did that piece age? Better than I’d like, honestly.

The crude failures got engineered away. Today’s models rarely trip over the obvious stereotypes; there are eval suites and red teams where there used to be shrugs. That part of 2023 already feels like another era.

But bias didn’t leave. It got polite.

The sharp edges moved into softer places: fluency that’s strongest in English and thins out from there, and nuance that survives the dictionary but not the workflow. Ask a model what pequeñito means today and it will give you a small essay on diminutives and affection; the vocabulary problem is genuinely solved. But run the same word through real work — translating a few hundred open-ends, rolling a focus group up into themes — and the warmth is the first thing to go. The model knows what the suffix means. It doesn’t carry it.

The model knows what the suffix means. It doesn’t carry it.

Most of all, bias moved into the averages. Trained on everything, a model returns the center of everything — and growth rarely comes from the center of a distribution. It comes from the edges: multicultural consumers, younger cohorts, shoppers whose category language hasn’t made it into anyone’s training data. Those are the people an averaged answer flattens first, and they’re exactly the people we’re hired to hear. A flattened answer doesn’t announce itself; it reads smooth, presents well, and misses quietly.

All of that quiet drift happened during the loudest stretch this industry has ever produced: new models every quarter, agents and reasoning, prices collapsing while capabilities climbed, demos of everything.

Insights spent those years working out in public what the technology was actually for. Two ideas got a full trial: that a language model with a demographic prompt was close enough to a person, and that the whole thing could be waited out. Both returned the same finding from opposite directions — the technology is real, and so is the need for real people underneath it. Most of the field ended up somewhere in the middle; the interesting differences are in how each shop got there.

The pressure reached our clients too — budget owners asking why research still takes weeks, procurement forwarding synthetic-sample pitches, every vendor arriving with an AI story. Since 2023, every client call I’ve been on has ended the same way: so… what’s your AI story?

Underneath the noise, what those two years actually changed is simple: answers got cheap. Anyone can generate a plausible read on any question in thirty seconds. And when answers get cheap, judgment gets expensive — knowing which answer to trust, which question was worth asking, which smooth paragraph is confidently wrong. That’s why everyone has a deck now and fewer have receipts. Ours turned out to be the work itself: six hundred–plus studies a year, 28% of the Fortune 500 in the last two years.

And when answers get cheap, judgment gets expensive.

Our answer to those loud years started somewhere unfashionable: with rules, published before we shipped a single AI tool. A short list of commitments, posted on crresearch.com, that every tool, vendor, and use case gets checked against — a test suite that is never allowed to fail.

The commitments:

• AI as an enhancement — not a replacement
• Transparency in AI applications
• Bias mitigation & data integrity
• Your data never trains AI models
• SOC 2 audited — annually

Each line means something specific in practice. “Enhancement, not replacement” means a senior researcher’s name is on every study, and that person, not a model, decides what the data means. “Transparency” means clients know when and how AI touched the work, every time. “Bias mitigation and data integrity” is the Chapter 1 worry turned into review steps instead of good intentions. And “your data never trains AI models” means exactly that — no exceptions, with an annual SOC 2 audit sitting behind the claim.

Half of those commitments are refusals, and that’s deliberate: in a market selling acceleration, what a partner declines to do says the most about its judgment. We’ve operated this way since 1959 — and when we launched the KidzEyes® panel in 2000, it was built COPPA-compliant back when researching children online at all sounded radical. The instrument was new; the standards weren’t. The balance is the product.

Principles are cheap until they ship. Here’s where AI currently earns its place at C+R — every tool built so a human stays decisive at the moment that matters:

The toolkit

What they have in common is that each one exists to catch the small thing. SmartProbe follows up in real time because the real answer usually lives one question past the checkbox. Accelerated A+R turns themes around in hours, but an analyst verifies them before they reach a client, because unverified speed is just being wrong faster.

And where AI doesn’t earn its place, it doesn’t go. SnapIQ delivers custom insights in three to four business days with no AI badge required, because sometimes the job is just speed and rigor. Deep custom qual and quant still carries the weight for the questions that deserve it: segmentation, brand tracking, in-home usage tests, ethnographies, online communities and panels.

As for the people from Chapter 2 — the ones models average away — we’ve spent decades building dedicated practices around them: multicultural, youth and family, LGBTQ+, shopper, B2B, generational — practices whose history runs through names like CultureBeat®, alive and well today — because hearing those consumers accurately was never something you could bolt on at the end. When the fieldwork needs to happen in Spanish, or Korean, or across a dozen markets at once, the answer is in-language researchers, not a translation layer.

The method fits the problem — an even mix of qual and quant across the year, senior-level attention on every single study.

Unverified speed is just being wrong faster.

Toolkit aside, the conclusion I’d write today is nearly the one I wrote in 2023, just with more evidence behind it. Language is still personal. Bias still hides where the data thins. And two loud years of tooling never changed the part that matters: someone still has to decide what the data means, and that someone should be a person.

What I’d add is about sequence. Everyone ends up with AI tools eventually; what separates research partners is what they built first. We published our guiding principles before we piloted our first tool. We piloted SmartProbe before we launched a client-facing suite. We grounded SmartPersonas in real segmentation work before we scaled AI-moderated qual and accelerated reporting. Careful is a feature, not a speed limit.

Careful is a feature, not a speed limit.

That sequence is worth interrogating anywhere — including here. What are the rules, and when were they written? What happens to client data, specifically? Which decisions does a human make, and which have quietly been handed to a model? Any partner with good answers won’t mind the questions.

Two years from now I expect half this toolkit to look quaint. I don’t expect the checklist to.

Curious what careful, AI-forward research could look like for your brand?

AI in Market Research, AI Smart Solutions, AI SmartPersonas