The synthetic-expert disclosure gap: AI respondent panels are entering buy-side workflows before the taxonomy exists

A research memo lands on a portfolio manager's desk. It cites five expert perspectives on data-center power constraints. Four are transcripts of recorded calls with named, vetted experts. The fifth is a model-generated response from a synthetic persona trained on prior interviews with utility executives. The memo does not distinguish between them. The LLM that summarized the memo cannot distinguish between them either. Neither can the compliance officer signing off on the fund's use of primary research.

This is the gap we want to name. The industry has spent five years building provenance metadata for content assets, source URLs, capture timestamps, transcription confidence, C2PA credentials for images and audio. It has not built provenance metadata for respondents. As synthetic-expert products move from B2B market research into buy-side primary research pipelines, the field that distinguishes a human expert on a recorded call from a synthetic persona generating a plausible response is missing, and the vendors that ship it first will define the taxonomy the rest of the industry inherits.

The product category is now real, not speculative

A year ago, synthetic respondents were mostly an academic curiosity and a market-research experiment. That is no longer where the products live.

NewtonX's launch of B2B Synthetic Personas is the clearest signal of the crossover. NewtonX runs a verified-expert network. Its core promise, the one it has spent a decade building, is that the respondent is real, vetted, and identity-confirmed. Sitting a synthetic-persona product on top of that panel is a deliberate move: it says the same buyer who wants a verified human expert also wants, in some workflows, a model-generated response calibrated against that verified-expert history. The two products live under the same roof.

Prolific's synthetic participants is the parallel move on the academic and consumer-research side. Prolific's positioning is that synthetic samples are useful for pilot instruments, hypothesis generation, and low-stakes iteration, with real human samples reserved for confirmatory work. Qualtrics, Yabble, Evidenza, and Rally sit in adjacent territory, each pitching some version of a synthetic audience to enterprise research teams. Evidenza and Rally have raised capital specifically on the AI-generated buyer-panel thesis.

What matters for the expert-network industry is not whether these products work. It is that they exist, they are sold to the same procurement layer that buys expert calls, and they produce artifacts (transcripts, quotes, structured responses) that look, on the page, indistinguishable from artifacts produced by human expert calls. A quote from a synthetic persona and a quote from a vetted expert both render as a paragraph of text with a name attached. The medium does not carry the metadata.

The agent runtime is where the disclosure gap becomes acute

The buy-side reads content differently than it did in 2019. A modern research workflow at a large hedge fund or a top-decile mutual fund composes outputs from a mixed input stack: internal notes, earnings call transcripts, SEC filings, sell-side research, expert-network call transcripts, industry primers, and increasingly, LLM-generated intermediate artifacts. Agentic research platforms including Hebbia, Rogo, Bridgetown, and AlphaSense Generative Search are the tissue that connects these inputs and produces synthesized outputs.

Inside that runtime, every input is a chunk of text with some metadata. The metadata typically covers source (which document, which URL), citation (page number, timestamp), and increasingly confidence (extraction quality, retrieval score). It does not cover respondent type. The chunk that contains a paragraph from a human expert call reads, to the model, exactly like the chunk that contains a paragraph from a synthetic-persona response. Both are text. Both are attributed to a named or role-labeled respondent. Both are treated with equal weight in the composition step.

This matters because the buy-side workflow has a compliance layer above it. When an analyst uses primary research in an investment memo, the compliance question is about what the respondent knew, whether they were subject to trading restrictions, whether they held MNPI, whether the network complied with its own vetting and screening protocols. That entire framework assumes the respondent is a person. It has no answer to the question: what does compliance look like when the respondent does not exist?

The gap is not hypothetical. Once a synthetic-persona response enters the context window of an agentic research platform and gets composed into a summarized output, downstream the analyst reading the summary has no way to tell which sentences originated from a human expert and which originated from a model. The provenance chain breaks at the composition step, and no one has proposed a standard for keeping it intact.

Why existing provenance standards do not cover this

INFLXD has covered content credentials before, particularly in the context of media asset provenance. The C2PA 2.0 specification is the most developed standard in production today. It attaches cryptographically signed manifests to media assets that record capture device, edit history, and originating tool. It is the right foundation for a broad class of provenance questions.

It does not solve this one. C2PA describes the asset. It does not describe the respondent inside the asset. A C2PA-credentialed audio recording tells you the file was captured by a specific device and has not been edited since capture. It does not tell you whether the voice on the recording belongs to a human respondent or a synthetic voice generated by a model. And in the text case, which is where most of buy-side research actually lives, C2PA is not the operative layer at all. A transcript that quotes a synthetic persona is not an asset with a cryptographic manifest; it is a paragraph in a Word document or a chunk in a vector database.

The missing field is not asset provenance. It is respondent-type disclosure: a structured, machine-readable flag on every quote, transcript segment, or attributed response that answers whether the respondent is a verified human, an unverified human, a synthetic persona, or a hybrid (a human respondent whose answers have been augmented or paraphrased by a model). That flag needs to travel with the content through the composition step, into the LLM context window, and into the downstream memo. It does not exist today.

The three respondent-type categories that will need labels

The taxonomy will settle. We think it settles around three primary categories, with a fourth edge case.

Verified human. A named respondent with confirmed identity, current or recent employment history, and a recorded exchange (call, structured interview, survey) whose transcript can be traced to a specific real person. This is the current expert-network product. It is what compliance teams currently understand.

Unverified or panel human. A real human respondent whose identity is confirmed to the extent required by the panel vendor's screening process, but who is not named or not identity-verified to the standard of an expert network. This is where much of survey research and consumer research lives. It has always been distinguishable from expert-network calls on price and process, but the outputs (a quote, a rating, a percentage) can look similar in a memo.

Synthetic persona. A model-generated response, produced by an LLM trained or prompted to simulate a specific respondent profile (a CIO at a mid-market manufacturing firm, a purchasing manager at a Tier 1 auto supplier, a nurse practitioner in a rural hospital system). No human answered the specific question. The response is a plausible reconstruction based on prior human data and prompt engineering.

The fourth edge case: hybrid. A human respondent whose answers have been paraphrased, summarized, or reformulated by a model. The substance came from a human. The wording did not. This category will grow, and it is the hardest to label cleanly because the human-to-model contribution ratio varies from call to call and from sentence to sentence.

Each of these needs a machine-readable flag that travels with the content. The flag needs to survive transcription, chunking, retrieval, and composition. That is a nontrivial engineering problem, and it is where the first-mover advantage sits.

The regulatory picture is narrow but not absent

Regulation is not the leading edge here, but it is not silent either.

The FTC's August 2024 rule on fake reviews and testimonials bans, among other things, the sale or purchase of reviews or testimonials created by AI when they are presented as coming from real people. The rule is aimed at consumer-facing commerce, not at B2B research, but the underlying principle (an AI-generated testimonial passed off as a human one is a deceptive practice) generalizes. A research vendor that sells synthetic-persona output without labeling it as synthetic is closer to the FTC's frame than the industry has fully absorbed.

The SEC's marketing rule for investment advisers covers testimonials and endorsements in adviser communications and would apply to any adviser using synthetic-respondent content in marketing material. It does not, on its own, reach into the internal research process. But the adviser rule sets a floor for how synthetic respondents can appear in adviser-published content, and that floor will pull the internal process along with it.

ESOMAR's 2024 guidance on synthetic sample in market research is the most developed industry-standard document in the space today. It sets expectations for how synthetic sample should be disclosed, validated, and used alongside human sample in market-research work. Its scope is market research, not buy-side primary research, and it has not crossed into the compliance frameworks that hedge funds, PE firms, and corporate strategy teams use for expert-network work. That crossover is the gap we expect to be filled next.

Our read is that regulation will lag vendor practice by 18 to 36 months. The vendors that adopt a clear disclosure schema early, publish it, and get customers to standardize on it will define the taxonomy that regulators eventually codify. This is the pattern from every prior provenance debate: privacy notices, cookie disclosure, algorithmic transparency, model cards. Practice sets the template; regulation ratifies it.

What a respondent-type disclosure field looks like in practice

If we sketch what a working disclosure schema would carry, it is not complicated. The field needs to travel at the segment level (per quote, per transcript chunk, per attributed response), not at the document level. Document-level flags are too coarse: a memo that mixes human and synthetic respondents needs per-quote granularity, or the flag is useless downstream.

At minimum, the field carries: respondent-type (verified human, unverified human, synthetic persona, hybrid), respondent-identity-hash (a stable identifier that can be used for deduplication and audit without exposing PII), source-medium (recorded call, structured interview, survey response, model-generated), and generation-method (for synthetic and hybrid cases: the model family, the prompt approach, the training or grounding dataset in general terms).

The hard part is not the schema. The hard part is getting the field to survive the composition step. When an agentic research platform pulls chunks from a vector database and composes them into a summarized output, the current standard is to preserve source and citation. Preserving respondent-type requires the platform to (a) accept the field as a first-class metadata attribute on every chunk, (b) propagate it into the composed output as inline annotation or structured footnote, and (c) surface it in the UI in a way the reader actually notices.

None of that is happening today at scale. It is the engineering work the first-moving vendor will do.

Who ships this first, and why the incentive is uneven

The incentives to ship respondent-type disclosure are not evenly distributed. Vendors that sell verified-human expert calls have a straightforward incentive: labeling makes their product visibly distinct from a synthetic alternative, and it reinforces the premium they charge. An expert network that ships a clean respondent-type flag is essentially advertising its own value proposition inside every downstream memo.

Vendors that sell synthetic-persona products have a more complicated incentive. Clear labeling constrains the use cases where their product can substitute for a human respondent. It also protects them from the FTC-style deceptive-practice risk and from customer blowback when a synthetic response gets treated as a human one and produces a bad decision. The mature synthetic-persona vendors we expect to see thrive are the ones that lean into labeling early, because their long-run product is not synthetic-passing-as-human; it is synthetic-known-as-synthetic, priced accordingly, used for the workflows where it is actually appropriate.

Agentic research platforms have the most complex position. Their product is composition. A respondent-type field is one more piece of metadata they need to preserve and surface, and every field they preserve is engineering work. But the customer they sell to (compliance-sensitive buy-side research teams) is exactly the customer that will demand this field once a single high-profile incident occurs. The platform that ships it before that incident becomes the standard-bearer. The platforms that wait become the ones catching up.

Our base case is that a verified-expert-network vendor ships the first credible schema, that one or two agentic research platforms adopt it within six months to protect their compliance story, and that the synthetic-persona vendors follow because the alternative is regulatory attention they do not want. That is how the taxonomy gets written: not by a standards body, but by the first three or four vendors whose disclosure format becomes the de facto interoperability layer.

What we would watch next

The signals that would confirm or refute this thesis are specific. A verified-expert-network vendor publishing a per-segment respondent-type schema and getting one large buy-side customer to require it in procurement is the strongest possible confirmation. An agentic research platform adding a respondent-type filter or annotation in its UI is a second-order confirmation. An enforcement action by the FTC against a synthetic-audience vendor for undisclosed use in a B2B context would accelerate the timeline sharply. ESOMAR extending its 2024 guidance explicitly to primary research used in investment decisions would signal that the market-research and buy-side worlds are starting to converge on a shared frame.

The counter-signals: continued vendor silence, no schema publication, no procurement-clause changes, and continued composition of mixed human and synthetic inputs into buy-side memos without any per-segment labeling. If that pattern holds through 2026, the first incident becomes a matter of when, not whether, and the disclosure regime that emerges will be reactive and less well-designed than the one the industry could build now.