The judgment pipeline

Executive summary

The AI's draft comes back in ninety seconds: open-ends coded, themes ranked, three implications drawn. The senior researcher reviewing it needs about two minutes to find the flaw, an implication resting on an assumption the data can't carry. She catches it because she spent two decades doing this exact work by hand, and the error pattern is as familiar to her as a mispriced item on a receipt. She fixes it, ships the deck, and the business decides well. Nothing about the moment looks like a crisis. But one thing in the room is invisible: the junior analyst who would have spent this quarter coding those open-ends, building the same reflex through the same tedium, was never hired. The task that would have trained her just ran in ninety seconds. A working senior researcher at PepsiCo named the moment precisely this quarter: AI is a powerful accelerant in a domain where she can easily spot where it went wrong, and she can spot it only because of expertise the next generation may never get the chance to build.¹¹

Q1 2026 asked what makes a research claim real. Practitioners kept returning to the harder question underneath it across our podcast's thirteen Q2 2026 conversations. From a Stanford lab, a supplements maker's insights function, a careers community, a research-platform founder's desk, and PepsiCo's flavor-foresight team, we heard the same durable human contribution named: judgment. Choosing the question worth asking, pressure-testing the output, tying the finding to the business decision it exists to move.^12,8,10,7,11 Even the quarter's policy voice arrives at the same place from the regulatory side: the industry's own 2025 ethics code now holds that no AI system used in research should operate without human judgment embedded in its lifecycle.⁵

The quarter's distinct discovery is what that consensus quietly assumes. Judgment is a manufactured asset with a supply chain: entry-level grunt work, apprenticeship beside someone senior, field contact with real consumers, and live ground-truth feedback that tells you when you were wrong. Q2's guests, mostly without prompting, described AI adoption dismantling that supply chain at every stage, in the same motion that it multiplies demand for the finished product: eliminating the entry rungs, automating the teaching tasks, and substituting simulation for field contact.^6,9,3,2 The industry is consuming judgment faster than it's producing it. Everyone agrees a human must stay in the loop. Almost no one can say who is training that human. That's the pipeline problem behind the AI problem.

The quarter's worry isn't whether a human stays in the loop but the prior question that consensus keeps forcing.

Who is training the judgment the loop depends on?

The answer Q2 returned, in thirteen independent voices and one converging shape: five gaps in the judgment supply chain, each requiring a different repair.

The five gaps:

Join us as we parse through these five gaps, explore counter-currents, define an emergent lexicon for practitioners, and chart a trajectory for the industry. Let's begin.

Section 01 · Grunt work vs. groundwork

The tasks AI automated were also the profession's only training program.

The beginner tasks were the apprenticeship

Every profession has a layer of work that its seniors are glad to be done with. Market research's version was specific: the phone room, the open-end coding spreadsheet, the junior project director role, the first-draft questionnaire. In the past eighteen months, AI and automation have absorbed most of that layer, and the organizational logic is impeccable. The work was slow, expensive, and repetitive, and the machines do it faster. The quarter's most alarmed voice, a marketing professor who founded a national student research competition, argues that the logic misses what that layer actually was. Market research knowledge has always been transferred apprenticeship-style, experienced researchers training younger ones on the job, and the beginner tasks were the medium of that transfer. Automate them away and the pipeline is severed: in ten years, when today's senior researchers retire, there will be no one trained to take over their companies.⁶

A bystander problem, and the rungs already gone

She gives the industry's inaction a name borrowed from social psychology: a bystander problem. Just as a crowd assumes someone else will help the person who has fallen, the industry assumes someone else is fixing the collapsing talent pipeline, so nobody actively does, even as layoffs and freelancing dismantle the structures that trained newcomers.⁶ The entry rungs that once let people fall into research careers, phone-room survey work, phone-room manager, junior project director, have been automated or eliminated, and a path that once required no bachelor's degree now demands an expensive master's. That's a quiet cost transfer with an ethical edge: students go into debt to subsidize the training employers used to provide. And the price of entry filters the pipeline down to those who can afford advanced degrees, draining the diversity of thought the field's own consulting clients learned the hard way to value.⁶ Even willing students hit a wall; landing a market research internship is now harder than breaking into digital, events, or influencer marketing. And she offers a precedent for where the road leads, one she attributes to the architecture profession: as she tells it, architecture cut junior hiring twenty years ago and now faces a shortage of licensed architects as its senior cohort retires. The causal story is hers to defend, but the shape of the warning is clear enough.⁶

The erosion has a second front, less discussed because it looks like progress. Accessible self-service research platforms now let people in digital, management, or any adjacent function run research themselves, which is good for research volume and quietly corrosive to the traditional entry points that big-agency work once provided.⁶ The signals that used to route talent into the field are degrading in parallel: with AI making it easy to cheat to a 4.0, she argues, GPA is no longer a reliable indicator of competence, so academia and industry need other ways for students to prove they can do the work. Her own answer is a third-party proving ground, a national competition in which students do real research for a major brand, evidence of competence a transcript can't provide.⁶ The pipeline problem, in other words, isn't only that the rungs are gone. The ladder's signage is failing too.

The manual struggle was the learning

The learning-science argument underneath is the quarter's most quietly radical claim. We still teach children arithmetic by hand despite calculators, because manual reasoning builds the problem-solving muscle the tool can't replicate. Most students, she argues, need the manual struggle, the stumbling and getting back up, to build the synapses that underpin real understanding; AI lets them skip exactly the step that produces the learning.⁶ The PepsiCo senior director supplies the corroborating evidence from the working world, and it cuts because it's a confession as much as an observation. She can let AI accelerate her work and skim its output for errors precisely because of the expertise she built doing the work manually. She has no confident answer for how younger researchers will develop that discernment without the same apprenticeship.¹¹ The Stanford professor gives the same idea its most durable formulation: the human contribution that survives AI is personal, apprenticed knowledge, a connoisseur's judgment that can't be transferred by curriculum.¹² Apprenticed is the operative word. It names a production process, and the process is being shut down.

The repairs are apprenticeship-shaped, and so is the dissent

The proposed repairs are all, revealingly, apprenticeship-shaped. The co-founder of a research learning community argues that learning the craft should look more like the trades, hands-on and in the field beside experienced practitioners, and less like a standardized academic program that's outdated by the time the curriculum ships.⁹ The marketing professor's fix is deliberately small and scalable: every freelancer and small firm offers a low-paid internship of five to ten hours a week for a semester, with the crucial condition that someone senior actually takes the intern under their wing. An intern left to figure it out alone has a part-time job, not an apprenticeship.⁶ The supplements-maker insights leader locates the repair inside everyday management: middle and upper managers must deliberately put junior talent on the spot, the follow-up question, the example, the board-presentation moment that AI can't answer for them. The critical-thinking muscle keeps getting worked even while AI runs backstage.⁸ A brand and insights leader who has run research across QSR and specialty grocery adds the generational framing: foundational research knowledge is being lost as new cohorts enter the field, and senior professionals carry a responsibility to mentor before that institutional memory disappears.²

The through-line has an internal dissent worth keeping in view, because it changes what the repair is for. The learning-community founder concedes the training vacuum, and then some; researchers today, he says, aren't really being trained to do anything in particular, and that vagueness is itself part of the profession's problem.⁹ But he rejects the collapse scenario outright. In his account, AI does to research what the tractor did to farming: it collapses the cost of production so far that the volume of research rises by a couple of orders of magnitude. And, like the cheese factory that created technician and QA roles, the AI research economy creates new human jobs: technicians who configure and update the tools, auditors who validate massive flows of AI-conducted interviews.⁹ If he's right, the industry's problem isn't a shrinking workforce; it's a workforce that must be trained for different rungs. But notice what his scenario still requires: someone competent enough to audit the flood. The abundance case and the collapse case disagree about demand and agree completely about the bottleneck. Both end at a training problem nobody currently owns.

Section 02 · Stop asking vs. ask better

Nobody in the record thinks consumers lie; the fight is over which tools get to work the contradiction.

The case every camp can claim

A meal-kit brand positions its product as a way to cook like a chef, and the positioning quietly repels the very buyers it was written for. The research that fixed it found the actual driver was nothing aspirational at all: dinner in twenty minutes or less. The intimidating promise lost to the modest one, and the research-derived language outperformed the aspirational language decisively.² That small case study is the quarter's cleanest specimen of its liveliest methodological dispute, because every camp in the dispute can claim it. Consumers said they wanted one thing and responded to another, and the disagreement is over what a researcher should do about that.

The camp that would retire the question

The first camp says the instrument is the problem. The CEO of a live-market experimentation platform puts the correlation between how people say they'll behave and how they actually behave at 20 to 30%, a figure this paper treats as her claim rather than established fact (see Notes and methodology), and draws the blunt conclusion: the industry should stop asking people what they would pay and whether they would buy.³ Her reasoning is more interesting than the slogan. Language itself is an unreliable vehicle, the same word meaning different things to different people, and respondents aren't lying; they genuinely don't know what they'll do at the moment of action. The replacement she proposes is the buying driver test: run live ads against a real segment and measure which value proposition actually causes a click and a purchase. Her evidence for why this can't be simulated by asking is a live experiment for a food-delivery brand in which plain home delivery from a big-box store significantly outperformed 30% cash back, a result she argues no survey and no marketer's intuition would have predicted.³

The camp that works the contradiction

The second camp says the contradiction is the finding. The supplements-maker insights leader has built a working method on exactly the data the first camp would discard. In her telling, agile research shifted the value of quant from any single data point to the pattern across many, and an off-pattern point is the tip of an iceberg. When consumers say they want cleaner ingredients or sustainability and then choose price at the shelf, they aren't lying; they're in conflict, because no one has offered them a product that lets them satisfy both values at once. And a discrepancy that recurs across demographics, seasons, retailers, and years, consistency in the inconsistency, is a cue to investigate rather than a data-quality error to discard.⁸ The strategic payoff is the strongest version of the argument: the first brand that resolves the tension wins the category's preference, and over time the resolved tension becomes the new point of parity. Executives who wave the gap away as consumers lying decline exactly the investment a competitor will eventually make.⁸ Where the experimentalist reads the gap as noise to be routed around, she reads it as a map of unmet demand.

Her method for working that map is deliberately artificial. Unconscious frictions surface when you place consumers inside engineered environments and constraints, conjoint and max-diff exercises, shelf tests, time pressure, and watch how consistent or inconsistent their behavior turns out to be, because respondents don't know they answered one thing and behaved another way.⁸ It's a middle path worth naming: it accepts the experimentalist's premise that stated preference can't be taken at face value, and it keeps the survey apparatus the experimentalist would retire, redeployed as an instrument for producing contradictions on purpose. The specialty-grocery leader adds the organizational stakes for whichever camp prevails: in many organizations, a confident opinion still routinely defeats well-researched data, a persistent dysfunction insights teams counter only by building fact-based credibility with decision-makers.² The say-do debate is, among other things, a contest over which kind of evidence a confident opinion finds hardest to override.

Watching with the sound off, and the accusation all three retire

A third position bypasses the argument by bypassing language altogether. The consumer insights VP at a major household-products company argues that the most powerful innovation opportunities come from unarticulated moments no consumer can voice, needs that live at the level of joy, surprise, and delight rather than the level of stated preference. Her discipline is to watch consumer behavior with the sound turned off, observing what people do and experience rather than what they say, and asking whether the task was as rewarding as it could be rather than as fast as it could be. Her proof case is deliberately unglamorous: brightly colored, scented trash bags, a genuine hit that no amount of asking would ever have designed, because no consumer has ever requested a cherry-blossom-scented bin liner.⁴

What unites the three camps matters as much as what divides them, because it retires a decades-old accusation. Nobody in this quarter's record believes consumers lie. The experimentalist describes vitamin buyers who tell themselves they're motivated by longevity while their behavior reveals a wish to look attractive now, self-deception rather than dishonesty, and the show's host sharpened the point: the gap is invisible from the inside, and respondents genuinely believe their own stated intentions.³ The PepsiCo senior director completes the picture with the self-described clean eater whose pantry is full of junk food and whose soda is an emotional comfort tied to home; surfacing exactly that illogical-but-meaningful truth, she argues, is the job of insights.¹¹ The say-do gap, on this quarter's evidence, has stopped being an embarrassment to be minimized and become raw material to be worked. The schism, and it's a genuine one, is over which tools get to work it.

Section 03 · Correlated vs. grounded

Validation against a corrupted benchmark validates the corruption.

The benchmark critique moves upstream

Q1 closed with the synthetic-respondent market separating into legitimacy tiers based on what the models were built from. Q2 moved the debate upstream, to the benchmark itself. The experimentation-platform CEO opens the new line of critique: the industry's favorite proof of synthetic validity, the 80 to 95% correlation with surveys that she describes as the standard vendor claim, proves only that both may be wrong in the same way. Synthetic audiences trained on survey data inherit the survey's foundational flaw: they model self-report rather than behavior.³ She then adds the quarter's most consequential supporting claim, and it demands careful handling: a paper she attributes to MIT estimating that 40 to 50% of survey respondents are now bots generated by large language models. If that figure holds, synthetics may correlate with surveys partly because the same models are generating both sides of the comparison. This paper treats the citation strictly as the guest's account, flagged for verification before any downstream use (see Notes and methodology), but the structure of the argument doesn't depend on the exact number. Validation against a corrupted benchmark validates the corruption.³

She grades her own side's homework with the same pen. Synthetics trained on first-party behavioral data are more grounded than survey-trained ones, but inherently backward-looking, able to answer questions about what's already in market and silent on new spaces, new partnerships, new launches.³ And the honest case for synthetics, in her telling, was never cost. It's real-time availability, an answer at the moment of a live argument about launch direction, which is a qualitatively different thing from a cheaper version of a slow process.³ The stakes are scaling with the automation around them: rolling out agentic marketing workflows without a validated, real-time representation of the customer means the same machinery that enables personalization can ship the wrong message to the wrong segment at enormous scale before anyone notices.³

The skeptic with the most at stake in being wrong

The skeptic's corner is occupied by the practitioner with the most at stake in being wrong. The PepsiCo senior director works in flavor foresight, the forward edge of innovation, and her objection to synthetic data is temporal: it's a model of past behavior, and relying on it means perpetually replicating the past. It can't project the drastic, unforeseen change that economic shocks, politics, or a pandemic produce; COVID is her canonical example of consumer behavior no model could have predicted. For forward-looking work she watches restaurants, chefs, and away-from-home channels instead, tracking how signals trickle down to CPG.¹¹ Her vendor heuristic follows: the signal that an AI tool adds value is that it accelerates the analysis of real data. The warning sign is personas and synthetic respondents replacing real people, where it's unclear whether anything is saved or whether the answers can be trusted at all.¹¹

Maybe there was never a gold standard

Then the Stanford professor arrives and quietly undermines the ground everyone else is standing on. Research, he argues, is in a reproducibility crisis across many domains; running the same human study twice doesn't return the same answer, so human research shouldn't be presumed the gold standard against which AI is evaluated.¹² That's the strongest pro-simulation argument this podcast has recorded, and it's strong precisely because it isn't boosterism. His case for simulations is that research is genuinely hard, and that a stack of structural problems, online fraud, unreachable samples, questions too sensitive to get permission to ask, might yield to simulation. AI's real promise, in his telling, is doing research better, not merely faster and cheaper.¹² He's equally clear about scope: simulations do best on simpler questions the internet already knows a lot about, and, like an airplane model, added detail makes a model harder to keep useful, not easier. And for commercial purposes he's willing to say what academics rarely say aloud: prediction without explanation can be enough. The computational linguists who just tried every possibility beat the theorists who insisted on first explaining language, and that pattern is now playing out in marketing, medicine, and economics.¹² He resists the caricature his own argument invites, though: prediction and understanding aren't antithetical, and enough accumulated prediction can, in some cases, climb back into the theoretical space of explaining why.¹²

Grounding as a maintenance activity

Read as a triangle rather than a spectrum, the three positions stop canceling out and start specifying a discipline. The professor is bullish because the benchmark is weaker than the industry admits. The senior director is skeptical because the future is wilder than the training data. The experimentation CEO is conditional, and her condition is the synthesis the quarter converges toward: someone on the team must own keeping synthetics correlated, regularly running live market experiments to compare against. Correlations decay as the market evolves; they need to be refreshed rather than assumed.³ Grounding, on this view, isn't a property a synthetic model has; it's a maintenance activity a team performs. Her companion test for trustworthy AI is the ability to declare ignorance: a system that confidently says it doesn't know, and directs you to run a live experiment, is worth more than one that generates a plausible fabrication.³ Her third requirement is provenance: every AI-generated insight ships with the evidence underneath it, the winning experiment variants, the relevant ads, the verbatim customer quotes, so users can trust the answer without having to trust the black box.³

The marketing professor adds the stakes: the whole point of research is to help people make sound decisions before they invest heavily. A flawed launch driven by an AI hallucination is exactly the outcome verification with real people exists to prevent.⁶ And the supplements-maker insights leader points at where the grounding will come from commercially: as public quantitative data becomes ubiquitous and undifferentiated through AI, the brands that distinguish themselves will be the ones investing in deep, proprietary qualitative data to feed their models. The usual flow reverses: qual feeds the quant, the LLMs, and the agents.⁸ The quarter's answer to the correlation wars is neither embrace nor abstinence. It's a loop: simulate, then verify against live humans, then re-simulate, with the loop staffed by exactly the judgment through-line I worries no one is training.

Section 04 · Writing questions vs. choosing questions

The craft's center of gravity is relocating from executing research to curating it.

Question-writing loses its professional standing

The bluntest sentence in the quarter's record comes from the head of learning and research at an insights career community, and it lands on a professional identity many researchers still print on business cards: the skill of writing questions is quickly becoming outdated, because today's large language models already draft them well. A researcher whose self-image is expert question writer, she argues, is no longer really an insights professional.¹⁰ What replaces it is business fluency: understanding how finance, marketing, and operations work and fit together. The work AI still can't do is tying an insight back to a business outcome, seeing the nuance and the politics in an organization, and translating what the data says into what a company must actually do. Her definition of the craft's durable core is a single distinction: the data is what it says; the insight is why it matters to the business outcome.¹⁰

The question itself is the durable contribution

The Stanford professor gives the same relocation its philosophical form. The most durable human contribution as AI advances, he argues, is the question itself, deciding what's important to try to know about. That's a connoisseur's judgment, personal and apprenticed rather than objective, and far harder for AI to replace than the sampling, interviewing, and statistics that will be automated away.¹² The show's host compressed it into the quarter's best coinage: the researcher as curator of questioning, valued for selecting and shaping which questions are worth asking almost more than for answering them.¹² The supplements-maker insights leader supplies the operational version, and it brackets the AI workflow neatly: problem framing on the inputs, asking AI the right questions, delegating the right tasks, designing the right agents; and judgment on the outputs, pressure-testing whether the AI's reasoning is logically sound. Garbage in, garbage out survives the technology transition intact.⁸ Her caution about the output side deserves its own sentence, because it names the failure mode that will actually catch teams: hallucination is the easy case, while the subtle failure is an answer resting on an untested assumption, or a concept that means something different to the AI than to your organization, smuggling in baggage a human must spot.⁸

The wider convergence, and the whiplash in education

The convergence on this through-line is the widest of the quarter, eight episodes, and the minor voices fill in the career ladder. The research-platform founder draws the line at delegation of the question itself: you can't hand AI your business question any more than you can hand it the decision, and asking AI to design the whole project is lazy in a way a real researcher must stay in the loop to catch.⁷ The marketing professor watches the same inflection hit students, who arrive having Googled answers their whole academic lives and must learn, often for the first time, to define the right question when the answer can't be looked up.⁶ The specialty-grocery brand leader translates fluency into a career instruction: marketers and insights people who don't understand P&L fundamentals are limited partners to the operators and franchisees they serve, and learning the financial reality of partners should be a deliberate investment.² The insights manager at an industrial manufacturer closes the loop reflexively: researchers should apply their core professional skill, asking the right questions, to their own technology purchases, not just to client work.¹

The education system feeding the profession is caught between the poles, and the careers-community head documents the whiplash from both directions. Working researchers' sharpest anxiety right now is not knowing which skills to invest in, whether to learn specific AI tools, storytelling, or business fundamentals, because job descriptions, interview conversations, and the job's actual demands keep diverging.¹⁰ Universities get the same static: industry advisory boards tell programs to stop focusing on Excel and PowerPoint, and students return from entry-level roles reporting that Excel and PowerPoint are the entire day.¹⁰ The supplements-maker insights leader's prescription reaches furthest upstream: with the world's knowledge at every student's fingertips, education's great remaining job is critical thinking itself, philosophy, logic, and the Socratic method, training people to judge information rather than merely retrieve it.⁸

The argument inside the consensus

But the consensus contains a genuine internal argument, and it's worth rendering at full strength because it decides how much of the craft survives. The learning-community founder insists that study design remains a defensible specialty: even with AI tools, product managers and designers can't do research as well as researchers can, because designing the thought process for getting someone to open up in the right way is a real skill most non-researchers will never develop.⁹ His famous formulation of the stakes appears in this quarter's lexicon: bad research doesn't stink. Unlike bad design, which announces itself, bad research fails invisibly, which is exactly why the craft layer can't be waved away.⁹ His sorting of the profession is generational advice in disguise: researchers who define themselves as people-talkers will experience AI as an existential threat, while those who see themselves as study designers and interpreters will experience it as a power multiplier.⁹ That's a direct answer to the careers-community head, from inside the same consensus. She says the writing is commoditized; he says the designing isn't, and the difference matters enormously to what universities should teach and what juniors should practice. The quarter doesn't resolve it. What both agree on is the direction of travel: away from execution as identity, toward judgment as identity, which is to say, straight into the pipeline problem of through-line I.

Section 05 · Installed vs. absorbed

The tools work; what fails is the organization's ability to metabolize them.

The process was never rewritten

The quarter's tooling conversations share a starting premise that would have sounded strange two years ago: capability is no longer the constraint. The tools work. What fails is absorption, and the head of learning and research at the insights career community has the cleanest diagnosis of why. Adoption breaks down because the process was never rewritten when the new tool arrived; organizations assume the tool plugs into the existing workflow, but its inputs, outputs, and timing are all different, and the process has to change to match.¹⁰ Tools get tested in a vacuum rather than against the rest of the stack, so a tool that shines alone creates friction the moment it has to mesh with everything else. And too many pilots begin with a directive she summarizes as go find something with AI, which isn't a problem statement at all; it's fear of missing out wearing a strategy costume.¹⁰

The evaluation checklist matures

Evaluation criteria are maturing in response, and the quarter assembled a practical checklist across four episodes. The careers-community head puts interoperability and exit cost at the top: what systems the tool must plug into, whether it has APIs, and whether you can export your data and leave, a long-term question most evaluations ignore until it's expensive.¹⁰ The research-platform founder argues the first day should deliver a quick win through a small pilot, because switching a stack is slow and the tool has to prove value on something immediate; his adjacent test is that a tool requiring a month of training to onboard probably isn't really a tool.⁷ The industrial manufacturer's insights manager starts earlier still, defining the specific problem and the must-have, nice-to-have, and avoid lists before any vendor call, and she contributes the quarter's most useful market taxonomy: the AI research-tech ecosystem has split into research-native builders, longtime practitioners working with technologists, and IT-world builders retrofitting tools from outside the discipline, and an evaluator's job is to tell them apart with hard methodological questions.¹ She also names the discipline almost nobody practices: recognizing when a previously valuable tool has been superseded and retiring it while it still works.¹

Why absorption fails even with good tools

Two structural notes explain why absorption fails so often even with good tools and good intent. The first is a bandwidth illusion: brands, suppliers, and research-tech each assume the others have more time than they actually do, when brand-side researchers spend their days wrangling stakeholders rather than combing data and suppliers are orchestrating internal teams of their own, so nobody's process rewrite gets resourced because nobody believes the others need one.¹⁰ The second is who's doing the selling: buyer frustration runs hottest where research-tech is staffed by technology people who can't speak basic research terminology or place their tool in a researcher's actual day, the IT-world half of the vendor taxonomy above.¹⁰ The platform founder's design principle answers both: apply AI to accelerate the specific steps where researchers already spend the most time, his own highest-value example being the auto-coding of open-ends that once meant days of manual spreadsheet work, rather than forcing agents in everywhere just to be able to claim you have them.⁷

The payoff when absorption succeeds is concrete, with the usual caveat that one team's experience is an anecdote rather than a benchmark. Her two-person team, after adopting an LLM analysis tool she credits with cutting open-end analysis time by more than half, grew from roughly one survey a month to three to ten, capacity multiplied rather than headcount replaced.¹ The claims are hers, the direction is consistent with the abundance case from through-line I, and the mechanism is exactly what the careers-community head prescribes: a tool matched to a named problem, absorbed into a rewritten process.

Fear or earned mistrust

The through-line's genuine disagreement is about the humans who resist, and it matters because the two diagnoses prescribe opposite treatments. The Stanford professor reads resistance as fear: the people who'll use AI in research organizations are already worried about their jobs, and a worried person is a poor experimenter, so leaders must introduce the technology in a way that makes people comfortable enough to experiment, precisely because nobody yet knows how best to use these tools.¹² His prescription is manufactured psychological safety, plus his own operating posture of leaning in, running pilots, and evaluating tools on whether they hold up on their face and stay consistent with decades of collected data.¹² The careers-community head rejects the premise. Researchers' reluctance, in her account, isn't primarily fear for their jobs at all; it's well-earned mistrust, built over years of being oversold and over-promised tools that never delivered the time savings on the label.¹⁰ If he's right, the remedy is safety: model experimentation, celebrate failed pilots, lower the personal stakes. If she's right, the remedy is credibility: vendors and leaders who make smaller promises and keep them, because the audience being asked to trust is a profession of professional skeptics who have been burned before. A leader who treats mistrust as fear will offer reassurance where the team is owed evidence. The quarter leaves both diagnoses standing, and the honest reading is that every insights organization contains both populations and would do well to know which is which.

Counter-currents

The abundance dividend

Against the quarter's scarcity mood runs a sustained argument that AI's real effect on research is expansionary. The learning-community founder points out that most executives can't answer fundamental, knowable questions about their own businesses, why they win or lose customers, why churn moved, purely because research has been too expensive to run, and that when AI collapses the cost, a vast body of currently unaffordable but valuable research becomes economical. His forecast is volume up a couple of orders of magnitude, with new technician and QA roles created along the way.⁹ The industrial manufacturer's two-person team living at three to ten surveys a month is the small-scale existence proof.¹ And the Stanford professor adds the qualitative version of the dividend: the first thing people imagine about AI is faster, cheaper, and job-threatening, but the more important promise is doing research better, because research is genuinely hard.¹² The abundance case doesn't refute the pipeline worry; it sharpens it, because abundance without trained judgment is exactly the flood of unaudited output the pessimists fear.

There is no gold standard

The Stanford professor's reproducibility argument deserves to stand alone, because it undercuts the validation vocabulary the entire quarter relies on. If running the same human study twice doesn't return the same answer, then validate synthetics against human data is a circular instruction, and the industry's whole correlation discourse is measuring agreement with a moving target.¹² What's notable in the record is the silence around it. No other guest engages the argument directly; the practitioners who invoke human data as ground truth simply don't address whether the ground is stable. That silence is itself a finding. The industry has a live, unanswered challenge to its deepest quality assumption sitting in plain view.

Surveys are future-proof

The quarter's loudest instrument-level claims, the experimentation CEO's call to stop asking and the industrial-manufacturer insights manager's judgment that survey research is a dying breed as the primary path to insights,^3,1 drew a calm structural rebuttal from the research-platform founder. A survey, he argues, is fundamentally a tool for hearing back from your customers at scale, and businesses will always want to hear from their clients directly rather than ask an LLM whether their client is happy. The format will keep evolving, as it has for 150 years, but the underlying purpose persists through every format change.⁷ Even the dying-breed voice concedes the core: quantitative research will never become irrelevant, because some decisions require statistically defensible confidence numbers that qualitative work alone can't replace.¹

AI isn't special

The quarter's policy voice, the senior advocate at the industry's trade association, pushes against the transformation rhetoric from an unexpected direction: regulatory pragmatism. Policymakers can't even agree on how to define AI; he cites a California attempt to regulate automated decision-making that was originally drafted so broadly it would have regulated spreadsheets out of existence. The better path, for regulators and researchers alike, is applying long-standing principles, transparency and consent above all, to a new tool, and tasking existing authorities with the parts that match their expertise rather than inventing an AI-specialist regulator.⁵ His advice to researchers mirrors his advice to legislators: treat AI as a new tool with new capabilities that still answers to first principles, and adjust from your strengths and weaknesses rather than reorganizing everything around the tool.⁵ In a quarter preoccupied with what AI changes, his is the standing reminder of how much it doesn't.

Emergent lexicon

Bad research doesn't stink

The observation that quality failures in research are invisible in a way other professional failures aren't. Bad design announces itself on contact; bad research looks exactly like good research until the launch underperforms, so quality problems go undetected precisely where the stakes are highest.

The bystander problem

Industry-wide diffusion of responsibility for the collapsing talent pipeline: everyone assumes someone else is fixing it, so no one does, even as layoffs and freelancing dismantle the structures that trained newcomers. The term imports the social-psychology framing deliberately; the remedy it implies is visible, named commitment.

The screenome

The five-second-resolution record of an individual's screen life, from the Human Screenome Project's method of capturing a device screenshot every five seconds for up to a year. Its Q2 currency is what it revealed: people spend roughly seven to ten seconds on any one thing before switching, stitching radically different content into narratives unique to them, which means aggregate media metrics like time spent on a platform have, in the researcher's phrase, melted away.

Consistency in the inconsistency

A say-do discrepancy that recurs across demographics, behavioral groups, seasons, retailers, or years. The recurrence is the tell: a one-off contradiction is noise, but a patterned one marks an unresolved consumer conflict, and the first brand to resolve it wins the category.

Research with the sound turned off

Observing what consumers do and experience rather than what they say, on the premise that breakthrough innovation lives in unarticulated moments of potential satisfaction that no consumer can voice. The operative question shifts from was this task as fast as it could be to was it as rewarding as it could be.

Decision engines

What social platforms have become: the place where learning, comparing, validating, and buying now happen, absorbing the jobs once done by search engines, product reviews, and store associates. The term reframes social from an attention channel to the full decision journey.

Exit cost

The neglected tool-evaluation criterion: whether you can export your data and leave, or whether everything gets stuck on the system. Paired with interoperability, it converts tool selection from a feature comparison into a long-term architecture decision.

Research-native vs. IT-world builders

The two camps of AI research-tech vendors: longtime market research practitioners building with technologists, and technology professionals retrofitting non-research tools for the insights market. The distinction gives evaluators their first sorting question, and hard methodological follow-ups do the rest.

Curator of questioning

Host coinage for the durable researcher role: the professional valued for selecting and shaping which questions are worth asking, almost more than for answering them. It compresses the quarter's widest convergence into a job description.

AEO, AI engine optimization

Earned visibility inside AI assistants' answers: structuring content so that when someone asks an AI which platform or product is best, the system can pull and rank you. A successor discipline to SEO, raised by the show's host and corroborated by the guest's forecast that AI chatbots will monetize brand placement the way social platforms did.

Scorecard — last quarter's predictions

Q1 closed with five named through-lines and three formal Year-ahead predictions. One quarter is too short to settle year-horizon calls, but Q2's thirteen conversations are the first evidence against which they can be checked; the read below is directional, not conclusive. Two of the predictions are graded inside the through-lines they belong to; the third is graded on its own line.

Held, with the axis shifted: Plausible vs. checkable. The legitimacy critique of synthetic respondents sharpened exactly as Q1's trajectory suggested it would: survey-trained synthetics inherit self-report's flaws, and the vendor correlation claims got named and challenged directly.^3,11 But Q1's predicted market disambiguation of the synthetic data label didn't visibly arrive. Instead the argument moved upstream, from output legitimacy to benchmark legitimacy, when the quarter's academic voice challenged the assumption that human research is ground truth at all.¹² The through-line held; the fault line moved.

Held, professionalizing: Statistic vs. story. Q2 added mechanics to Q1's persuasion thesis. Narrow a million data points to the five that make the decision clear, with everything else in the appendix; pre-seed controversial recommendations in elevator moments before the formal readout; define what evidence type the audience will find compelling before gathering it; and study professional influencers, since the stakeholders researchers persuade respond to the same engagement drivers as social audiences.^8,3,4 What was a stance in Q1 is becoming a toolkit.

Stalled as stated, transformed in substance: A number vs. a replicable number. Explicit replicability-guardianship language nearly vanished from the Q2 record, and honesty requires saying so. What replaced it is the deeper version of the same worry: not whether anyone will guard the numbers but who will be able to. The guardianship question became the judgment-pipeline question, who trains the researcher capable of catching the error,^11,6 and the question-ownership question, who still knows which number matters.¹⁰ The prediction's letter stalled; its spirit migrated into this edition's thesis.

Quiet quarter: Reluctant vs. selective. Faint echoes only. The household-products insights VP debunked the hard natural-versus-traditional consumer bifurcation in favor of job-to-be-done navigation, applied cohort-versus-life-stage discipline to generational claims, and reported Gen Z redefining what clean enough means, with nearly half of Gen Z saying they look forward to cleaning against 18% of older generations, figures from the company's own research.⁴ Real signal, but segmentation reform wasn't a Q2 through-line, and this scorecard says so rather than forcing it.

Accelerated most: Study completed vs. decision influenced. Q2 institutionalized what Q1 observed. Insights as accountable co-creator from the brief onward, with shared ownership of results on high-risk projects; insight defined outright as why it matters to the business outcome; internal cross-functional partners named as the insights team's real customers; and the discipline turned reflexively inward, applied to researchers' own tool purchases.^8,10,2,1 The decision-linkage test Q1 flagged as an emerging standard now reads like the quarter's default operating assumption.

Not yet evidenced: The AI sampling threat will produce a visible quality incident at the industry level. Nothing in the thirteen Q2 conversations surfaces a high-profile research failure publicly attributed to AI-assisted survey fraud; the prediction stands open on its original horizon.

Accelerating

• The qualitative renaissance. Deep qual as the differentiating fuel for AI models, at AI-lowered friction. The supplements-maker insights leader forecasts a three-to-five-year window in which ubiquitous public quant data stops differentiating anyone and proprietary qualitative depth becomes the scarce input, with AI collapsing qual's historic cost and cycle time; the industrial manufacturer's team and the research-platform founder are already operating the low-friction version.^8,1,7
• Live behavioral validation as a standing function. The re-validation loop, someone on the team responsible for keeping models honest against live market behavior because correlations decay, is moving from project to permanent job description.³
• Business fluency as the researcher's core credential. Understanding finance, operations, and P&L reality is displacing methodological breadth as the career-defining skill, from the careers-community head's ten-year advice to the specialty-grocery leader's franchisee-trust playbook.^10,2
• AI-disclosure law as operating reality. The trade-association advocate counts California, New York, and Maine bot-disclosure statutes already on the books, twenty states with comprehensive privacy laws in effect and likely more by year's end, and provenance-disclosure requirements now written into the industry's own ethics code; all counts are his and should be re-checked against current law at publication.⁵ Compliance is no longer a future scenario for AI-moderated research.
• Talent-pipeline alarm going public. The bystander problem now has a proposed mechanism attached, a visible public pledge of internship hours through an industry outlet, moving the issue from conference-hallway lament toward named commitment.⁶

Subsiding

• Question-writing as professional identity. LLMs draft well enough that the expert question writer, as a self-image, is losing its claim to professional status. The durable identity is moving up a level, to question selection and business translation.¹⁰
• The survey as default first instrument. Contested, and this paper's Counter-currents carry the rebuttal, but the directional signal is real: the experimentation camp routes around asking entirely, and even a working insights manager calls survey research a dying breed as the primary path to insight.^3,1
• Tool-stack maximalism. The five-or-six-tool research stack is giving way to consolidation logic and exit-cost scrutiny; the evaluation questions are now about interoperability and the cost of leaving, and the ownership question of retiring superseded tools is on the table.^7,10,1
• The 50-page deck as delivery format. The influencer-craft argument and the five-data-points discipline converge on the same verdict: the stakeholders being persuaded respond to engagement, and the appendix is where completeness now lives.^4,8
• Polish as a trust signal. In content and, by extension, in research delivery: overproduced material increasingly reads as corporate filtering, while imperfect, conversational material reads as credible. The digital-transformation strategist's account of creator content outperforming brand production is the consumer-side version of a shift insights teams are starting to feel in their own readouts.¹³

The year ahead

Within four quarters, at least one major insights employer or supplier launches a formal, publicly announced apprenticeship or rotation program explicitly framed as AI-era judgment training. Signal: the repairs proposed across the quarter are uniformly apprenticeship-shaped, from trades-style learning to scaled micro-internships to managers deliberately putting juniors on the spot, and the demand side of the abundance scenario needs the same trained auditors.^6,9,8

Correlation claims lose currency: at least one synthetic-respondent vendor pivots its marketing from percentage correlation with surveys to live or behavioral re-validation as the headline feature. Signal: the benchmark critique now has two independent prongs, self-report circularity and bot contamination of the training benchmark, and the practitioner heuristic already treats acceleration of real data as the value signal.^3,11

A state AI-notification law produces the first enforcement action or public compliance scramble involving AI-moderated research interviews. Signal: bot-disclosure statutes are on the books in California, New York, and Maine by the advocate's count, and California's broadly drafted automated-decision rules already sweep in incentivized research subjects in ways most researchers haven't noticed.⁵

An industry body operationalizes the visible-commitment mechanism: a public internship and training pledge with named firms, within twelve months. Signal: the mechanism is specified, publicly pledged hours through an industry outlet with participation as a point of pride, and the competitor-collaboration norm it requires already exists in the industry's data-quality work.^6,7

A major brand tracker adds a continuous conversational-qual layer, marketed explicitly as feeding the brand's AI models. Signal: the qualitative-renaissance forecast names the mechanism, qual as proprietary model fuel; the call to replace two-to-three-year foundational studies with continuous iterative ones names the cadence; and conversational AI inside surveys is already shipping.^8,1,7

What to do with this

Two routes out of the quarter, calibrated to where the reader sits.

Stay in the conversation. Subscribe to the Curiosity Current quarterly white paper (one issue per quarter, delivered to inbox) and explore the underlying corpus: every claim in this paper traces back to a topic page and a named contributor. The platform is built to be read down, not just across.

Where we sit. aytm's published positions on the questions this quarter raised are gathered at aytm.com. Three threads in particular: experienced human judgment as a scarce, appreciating asset as AI scales; synthetic data earning trust through standing validation against real human response, not a one-time correlation; and the say-do gap read as aspiration interrupted by context, with empathy as the instrument that captures it. Real data is the best AI asset, aytm's synthetic-data whitepaper, is the next artifact in this conversation.

References and contributors

1. [1] Sarah Haftings, Insights Manager, Shurtape Technologies. Tool vetting, research-native vs. IT-world builders, AI capacity gains, qual-forward outlook (April 7, 2026). Episode
2. [2] Gloria Reardon, Brand Marketing / Insights Leader, The Fresh Market. Menu adoption curve, meal-kit positioning case, franchisee trust, P&L fluency, mentorship (April 14, 2026). Episode
3. [3] Kate O'Keeffe, CEO & Co-Founder, Heatseeker. Live market experiments, say-do correlation critique, synthetic re-validation loop (April 21, 2026). Episode
4. [4] Oksana Sobol, Vice President of Consumer Insights, The Clorox Company. Enduring trends, sound-turned-off observation, generational discipline, influencer craft (April 28, 2026). Episode
5. [5] Howard Fienberg, Senior Vice President of Advocacy, The Insights Association. AI disclosure law, privacy patchwork, 2025 ethics code AI provisions, regulatory pragmatism (May 5, 2026). Episode
6. [6] Brooke Reavey, Professor of Marketing; Founder & President, Marketing Research Competition, Dominican University. Bystander problem, entry-rung collapse, manual-learning case, internship repairs (May 12, 2026). Episode
7. [7] Marin Mrša, Founder & CEO, Peekator. Survey future-proofing, stack consolidation, quick-win pilots, AI co-creation limits (May 19, 2026). Episode
8. [8] Deborah Mendez, Consumer Insights Leader, Pharmavite. Consistency in the inconsistency, insights as co-creator, problem framing and judgment, qualitative renaissance (May 26, 2026). Episode
9. [9] Alec Levin, Co-founder & CEO, Learners. Abundance case, apprenticeship-style learning, study design as defensible craft, bad research doesn't stink (June 2, 2026). Episode
10. [10] Zontziry Johnson, Head of Learning and Research, Insights Career Network. Question-writing commoditization, business fluency, absorption failures, exit cost, earned mistrust (June 9, 2026). Episode
11. [11] Kerry Ellen Schwartz, Senior Director of Global Foods, PepsiCo. Expert error-catching, synthetic-data skepticism, consumer contradiction, AEO corroboration (June 16, 2026). Episode
12. [12] Byron Reeves, Chaired Professor (Communication, Symbolic Systems, Education), Stanford University. Reproducibility critique, screenome, prediction vs. explanation, the question itself (June 23, 2026). Episode
13. [13] Vladimer Botsvadze, Digital Transformation & Marketing Thought Leader. Decision engines, trust over polish, social as decision journey (June 30, 2026). Episode

Notes and methodology

Indexed record as instrument. This white paper treats the indexed record of thirteen Q2 2026 episodes of the Curiosity Current podcast (the corpus, 335 indexed claims) as the instrument of analysis. The unit of analysis is the phenomenon, the recurring pattern, emergent tension, or convergent observation, not the individual guest. Practitioners appear as data points; their arguments are examined for convergence, divergence, and explanatory weight against the quarter's through-lines.

Attribution format. In-body attribution uses role and organization where the source's institutional vantage point carries analytical weight. Full personal names appear only in References and contributors. Footnote numerals correspond to the episode's position in the quarter's broadcast order and are reused on subsequent citations of the same source. Inline topic references link to the topic page's markdown view, the canonical citation surface for the underlying claims; lexicon terms link to their definitions in this document. Host observations are used only as color or coinage, per the convention established in prior editions.

Where-we-stand framing. Each through-line closes with a brief synthesis of aytm's published position on the question the through-line raises. These syntheses draw from positions already in this corpus (primarily aytm CEO Lev Mazin's prior-quarter claims) or from aytm-published material at aytm.com. The body of each through-line remains editorially neutral. The Where-we-stand and Where-we-sit copy is drawn from Lev's published positions and was editorially signed off before publication.

Verification flags. Several specific figures in the corpus are flagged for independent verification before use as cited evidence:

All flagged claims are attributed clearly to their sources in the body text. None should be treated as independently verified fact.

Scope of the thesis. Two episodes sit beside the judgment-pipeline thesis rather than inside it, and this paper hasn't forced them. The digital-transformation conversation (June 30) is a brand-marketing and social-media episode with little research-craft content; it contributes to the Emergent lexicon and one Subsiding note.¹³ The policy conversation (May 5) contributes its human-in-the-loop code provision to the thesis and anchors a Counter-current and the regulatory trajectory notes, but its core lives in the legal landscape rather than the craft debate.⁵

Cross-references. Convergences and tensions noted in this paper are drawn from the Curiosity Current indexed episode record. Q2 2026 introduces thirteen contributors. References to Q1 2026 through-lines are drawn from that edition's indexed episodes and published white paper.

Quarterly scope. All references to episode counts, contributor counts, and claim counts in this paper are scoped to Q2 2026 and don't represent living totals from the full Curiosity Current indexed record.

Episode links. Full episodes are available on the Curiosity Current podcast feed on Apple Podcasts, Spotify, and YouTube.

Curiosity Current is produced by aytm. This white paper was compiled by the Curiosity Current editorial team from the quarter's indexed episodes. Published Q3 2026.

Produced by aytm

Curiosity Current is made by aytm, the consumer-insights company. About aytm