By: Molly Strawn-Carreno, Director of Brand Growth, aytm & Co-Director of Membership, Insights Association West Chapter
Every AI capability in market research arrives with a trust condition attached. This guide takes the six set out in the Insights Association's Code of Standards and Ethics and works them section by section in conversation with Howard Fienberg, Senior VP, Advocacy at Insights Association, and grounded in aytm's own practice.
Speed without trust is just noise
Every new AI capability in market research answers the same promise: faster. Faster fielding, faster analysis, faster decisions. But faster than what, the consent flow that tells participants what they're part of, the review that catches a biased instrument before launch, or the conversation with a client about what AI actually changed in their data?
Speed only earns its keep when trust keeps pace, and the IA's September 2025 Code update is the industry's response. It folds AI provisions into the framework that has governed responsible research for decades, on the theory that new technology raises old questions in louder voices, and in conversation with Howard Fienberg, the Code's practical edge sharpens.
"Ethical practice isn't a constraint on AI-driven research. It's the infrastructure that makes AI-driven research worth trusting." — Howard Fienberg
Transparency is a three-way obligation
When AI shows up in a study, three audiences need transparency, and the Code addresses each one directly:
- Code §2.4 requires researchers to notify participants when AI-based avatars or chatbots could be perceived as human.
- Code §8.3 requires transparency to clients about data origins and any AI use that may represent a risk to research quality or accuracy.
- And Code §4.2 requires AI-generated data to be distinguished from human-derived data, with the tool's purpose, technique, model type, accuracy, and source disclosed.
"Surprises in research results are fine—that's what research is for. Surprises about each other's practices are where trust breaks down." — Howard Fienberg
Disclosure must be built into the workflow rather than bolted on after the fact. On aytm, that takes shape across the platform. For example, Conversation AI runs AI-moderated qualitative interviews at quantitative scale with §2.4 disclosure in the participant flow- every Skipper output traces what was AI-generated versus human-derived, and research data never enters public AI training sets. The reasoning: informed participants produce better data, clients defending findings need to trace their origins, and the credibility of research depends on knowing where the data came from.
Human judgment is what makes AI useful
The tempting narrative says, “let the machine handle it.” Let AI draft the survey, run the analysis, generate the report, and while every step is technically possible, every step taken without human oversight is a step away from defensible research. AI is an amplifier, and what it amplifies depends entirely on what it's pointed at. If it’s aimed at expertise, it amplifies expertise; if it’s aimed at whatever lives in the training data, it amplifies that.
Code §4.3 captures the principle directly: No AI system used in research should operate exclusively without human judgment embedded in its lifecycle. Speed doesn't override rigor, rigor is what speed is for in the first place.
At aytm, we’ve noticed that as research tools go self-serve, the people running studies aren't always trained researchers, and the human-in-the-loop principle only works when the human can actually exercise meaningful judgment.
"As AI absorbs more of the mechanical work, the decisions that require expertise, context, and judgment become the entire job." — Howard Fienberg
This shapes how Skipper, aytm's AI suite, is built. Skipper Draft turns a research brief into a methodologically sound survey for human review, while Skipper Smart Review evaluates surveys before launch and flags logic errors, excessive length, and unclear phrasing. Every output is transparent, editable, and under the researcher's authority so that the human in the loop is genuine, not nominal cover.
Data provenance is now an existential question
For most of research history, provenance was simple: questions, answers, analysis, with a clear chain of custody from one to the next. Generative AI breaks that simplicity by opening at least three new forks:
1. Data that looks like survey responses, but came from a model
2. Samples augmented with synthetic respondents
3. Predictions about how a segment would have answered questions it never saw.
Code §4.2 requires every AI-generated data point to be identified, and Code §4.1 governs personal data in AI training sets while protecting against PII reverse-engineering through inference.
"The burden sits with the researcher, not as a one-time assessment, but as an ongoing responsibility. Participant trust depends on it." — Howard Fienberg
aytm's position on this is structural: Real data is the best AI asset. PaidViewpoint, Traffic Sentinel, and Data Centrifuge form an end-to-end provenance chain, including a dedicated LLM detection layer that flags AI-generated text in the data itself.
The policy-practice gap is a ticking clock
The most common ethics failure in AI-driven research is quiet. The gap between what documentation says and what teams do.
Howard has seen this pattern before. Across nineteen years at Insights Association, he's spent time explaining that organizations can't copy and paste a privacy policy from somewhere else, no matter how good the source document looks, because a policy without internal alignment generates more risk than no policy at all. AI plays the same dynamic out faster, since the tools change faster.
"The disconnect between what you're putting out into the world and what you're actually doing creates real exposure—regardless of intent." — Howard Fienberg
The fix isn't more paperwork, it's alignment. Privacy policies that address AI processing, consent flows that disclose AI involvement, methodology reports that identify AI-assisted steps. aytm operates this way by default: SOC 2 Type II audited annually, GDPR and CCPA compliant, with 100% link encryption against an industry average of 75.4%.
Why bias compounds
Most organizations treat AI bias as a setup problem, something to evaluate before deployment and check off the list, but the reality is messier. Bias enters across the lifecycle, with each layer compounding the one before: Training data underrepresents populations, AI-drafted instruments embed cultural assumptions, analysis surfaces patterns from historical biases, and reporting tools reinforce assumptions where they should be challenging them.
Code §4.3 requires regular evaluation of AI models for intentional and unintentional biases, including bias from demographic or cultural variables. aytm runs quality evaluation continuously across the data lifecycle: Data Centrifuge during and after fielding, DraftScore predicting respondent fatigue before launch. Thus, evaluation behaves less like a milestone gate and more like a system that runs.
Speed makes respect harder, not easier
AI compresses timelines, which is the headline benefit, but every hour saved on the production side becomes pressure on the one element of research that can't be automated: the participant's experience.
When studies are faster and cheaper to run, the temptation is to collect more—more questions, more demographic detail, more data points per respondent. Each addition is small, but the cumulative burden is not, and it shows up downstream as fatigue, abandonment, and quality removal.
"Surveys that are too long, studies that collect insanely detailed demographic information that isn't relevant to the results […] add time and burden on the respondents and create greater risk." —Howard Fienberg
That burden is the upstream cause of low-quality data downstream, which means ethics and data quality are the same conversation. At aytm, we build infrastructure to head it off before it happens. DraftScore predicts respondent burden before launch, Traffic Sentinel filters traffic at the source, PaidViewpoint provides verified respondents, and Data Centrifuge cleans during and after fielding.
|
aytm 2025 Benchmark
|
Result
|
|
Quality removal rate
|
0.4%
|
|
Abandon rate
|
3.0%
|
|
Median study length
|
5.3 min
|
|
Actual incidence rate
|
80.0%
|
That's infrastructure delivering on Code §1.1's mandate to balance subjects, integrity, and objectives.
Ethics as infrastructure
Every AI capability has a trust condition. Speed requires transparency, scale requires human judgment, generation requires traceability, adoption requires policy alignment, automation requires ongoing bias evaluation, and efficiency requires genuine respect for participants.
"These aren't constraints on AI adoption. They're the conditions under which AI adoption produces research worth acting on." — Howard Fienberg
The September 2025 Code update validates this institutionally. The profession is setting its own standards through self-regulation, and the way to protect that autonomy is to use the Code proactively, early and often, as a competitive feature rather than compliance overhead.
At aytm, this is operational. Skipper exists because we made the same bet the Code formalizes: AI handles the computational work, researchers control the decisions that require expertise. The organizations that demonstrate ethical commitment through their actual practices earn the trust that makes everything else possible, and in the AI era, trust is the only kind of speed that compounds.
If you’re interested in hearing our full conversation with Howard Fienberg, head over to the Curiosity Current podcast, and have a listen here.