AI BS Detector: Real AI Research Capability vs. Thin Wrappers (July 2026)

Jul 7, 2026 by Ethan Pidgeon

I'll be frank: the words 'AI-powered insights' mean almost nothing right now without a follow-up question. AI washing market research vendors has become a full-time job for serious buyers, because the gap between a system that actually reasons over sourced evidence and one that generates fluent-sounding text from thin context is invisible in a standard demo. The good news is it's not invisible if you run the right four tests yourself.

TLDR:

• AI washing ranges from outright fabrication to thin API wrappers rebranded as proprietary AI; tiers 1 through 3 cover most of what gets sold as AI research.
• Specialized RAG tools hallucinate between roughly 17 and 33 percent of the time, per a Stanford study, and a fabricated stat in a board deck becomes consensus by the next meeting.
• Ask any vendor to click a specific claim through to the source page with a retrieval date. If it requires a follow-up email, you have your answer.
• Run four self-directed tests in one afternoon: consistency check, known-answer test, no-good-answer test, and source pull. Log results across vendors in a shared sheet.
• Merciv attaches a three-tier confidence score (High, Directional, or Exploratory), a named source, and a retrieval date to every output, with uploaded documents kept outside model training.

What AI Washing Actually Is

AI washing is overstating or misrepresenting AI capabilities in product marketing to win customers or investment. The term borrows from greenwashing, and the mechanic is the same: the claim outruns the product.

The behavior sits on a range. At one end, outright fabrication: models that do not exist, or accuracy figures with no published methodology. In the middle, relabeling: if-then rules and scripted workflows repackaged as machine learning, dashboards rebranded as agents. At the softer end, thin wrapping: a third-party model called through an API, dressed in a proprietary interface, and sold as house-built AI, per the CFA Institute's 2025 report on AI washing.

For an insights buyer, the working definition matters because every category below assumes it. A vendor claim is not evidence. Retrievable architecture, published limitations, and reproducible outputs are. Assessing the best AI tools for market research starts with exactly these criteria.

Why Market Research Tools Are a Prime Target

Research software sells on trust, not benchmarks. Unlike a database or payments system, the output is a narrative, and fluent narrative is hard to falsify in a demo. That gap is where AI washing thrives.

Three conditions stack in the category's favor:

• Buyers rarely have a data engineer in the room during evaluation, so architecture claims (retrieval, graph reasoning, agentic orchestration) go unchecked.
• The deliverable is prose. A confident summary reads the same whether it came from a retrieval system with citations or a general model guessing from thin context.
• Vendor incentives reward the biggest verbs available: synthesizes, predicts, decides. No public accuracy leaderboard exists for consumer insights tools, and no industry body requires vendors to publish against a shared eval set.

Self-description is close to unreliable by default.

The Scale: From Real AI to Rebadged Algorithms

Treating "AI" as a single label collapses a category that behaves nothing alike in practice. The useful move is to place any vendor on a five-tier ladder, from the crudest automation to systems that reason over structured evidence.

Tiers 1 through 3 cover most of what gets sold as an AI research product. Genuine synthesis lives at tier 5, and the real test is whether a vendor can show you the retrieval path for a specific answer, a pattern covered in depth in why internal RAG for consumer insights fails.

Red Flags That Signal AI Washing

Most AI washing shows itself in the demo if you know where to look. The signals below do not require a data engineer to spot, and any one is enough to warrant a harder second meeting.

• The rep cannot click a specific claim in an output and land on the source page with a retrieval date attached. "We cite sources" collapses when tested at the sentence level.
• Accuracy is quoted as a headline number with no evaluation framework, no test set, and no published limitations. Ask what the model gets wrong; a real answer exists.
• The same prompt, run twice, returns materially different findings. Run-to-run drift means the system is guessing on retrieval, not reasoning over a fixed evidence set.
• "Hallucination-free" is used as a marketing line. No production LLM earns that phrase.
• Descriptions of the AI stay abstract. Ask which retrieval paths run, in what order, and where code executes versus where the model generates.
• Licensed data is claimed without contracts named or refresh cadence disclosed. If the vendor cannot describe which feeds refresh weekly and which sit static, the corpus is likely the open web.

Walk in with this list. The vendors who pass will thank you for asking.

Questions to Ask Any AI Research Vendor Before You Sign

The strongest evaluation move is a fixed question set, delivered in the same order to every vendor, with answers written down. Capability shows through in what the rep can produce on the spot.

• Trace a finding to source. Pick any sentence in a sample output. Ask the rep to click through to the exact page, with retrieval date. If it takes a follow-up email, that is the answer.
• Run it on our data. Curated demos prove nothing. Ask for a pilot on your own decks, POS extract, or review pull, with a question set you write. AI market research should be faster and defensible, not merely fluent.
• Show me a low-confidence answer. A credible system separates high, directional, and exploratory findings, and will say when evidence is thin. Fluent certainty on a shallow corpus is the failure mode.
• What happens to our uploaded documents? Ask for the training-data policy in writing: whether prompts, files, and outputs are excluded from training, and whether that exclusion extends to third-party model providers.
• Has any AI claim been independently validated? A published eval, a customer-run benchmark, a third-party audit. If the only proof is the sales deck, the claim is marketing.

Send the same list to three vendors. The gaps between answers will do most of the work for you, especially if you are already weighing alternatives to traditional consumer research beyond the quarterly report.

The Hallucination Problem Vendors Won't Lead With

Hallucination in a research context is not one failure. It shows up in three shapes, and vendors lead with the least damaging one.

• Factual hallucination: the model invents a statistic, market size, or quote.
• Citation hallucination: the source itself does not exist. A plausible title, a plausible year, no underlying document.
• Misgrounding: the citation is real, but does not support the claim attached to it. The hardest to catch, and the most common in enterprise settings.

A Stanford study in the Journal of Empirical Legal Studies found specialized RAG tools hallucinated between 17 and 33 percent of the time, which is one key reason to evaluate Merciv vs. ChatGPT for consumer research on citation reliability.

A fabricated stat that lands in a board deck acquires institutional weight. Six weeks later, it is quoted back to you as consensus, which is why board-ready consumer insights without black-box AI demand full source traceability.

Source traceability is infrastructure, not a bonus feature. If a finding cannot be clicked through to a specific page, on a specific date, with the surrounding paragraph visible, "AI-powered" says nothing about whether the number is real.

The Regulatory Net Is Tightening

Enforcement has caught up to the marketing. In 2025, the FTC brought at least a dozen AI washing cases under Operation AI Comply, targeting companies selling machine learning capabilities backed only by manual work.

On the securities side, at least 46 AI-related class actions had been filed in the US since 2020 as of early 2025, with more filed since, the majority alleging AI washing. Then-SEC Chair Gary Gensler (who left office in January 2025) formally coined the term in March 2024, and cases have continued across administrations.

For buyers, this reframes the exposure. A vendor whose AI claims cannot withstand a regulator's question will not withstand your board's, either.

How to Run Your Own Capability Evaluation

Vendor demos are curated. A trial run on your own terms is not. Four tests, run in one afternoon, will tell you what the sales deck cannot.

• Consistency check. Run the same prompt five times in fresh sessions. Materially different findings means the retrieval is guessing, not reasoning over a fixed evidence set, a pattern that also inflates the cost of an in-house insights copilot.
• Known-answer test. Ask a question you already have the answer to from a recent report. Compare the tool's output against the underlying document line by line.
• No-good-answer test. Ask something the corpus cannot support, like a niche competitor's private financials. A credible tool admits the evidence is not there. A washed one invents.
• Source pull. Open any cited claim and read the surrounding paragraph. If the passage does not support the claim as written, the tool is misgrounding.

Run the same four on every vendor. Log results in a shared sheet.

How Merciv Tackles the AI Washing Problem

The tests in this piece are the same ones we built Merciv to pass. Every output carries a three-tier confidence score (High, Directional, or Exploratory), a named source, a retrieval date, and a click-through to the underlying page. Uploaded documents and licensed syndicated research stay inside a walled-garden tenant and never enter model training, a standard the best consumer insights platforms for enterprise teams should all meet.

We do not claim zero hallucinations. We claim every finding carries a citation and a confidence score, so any number in the deck can be pressure-tested by your CMO before the meeting starts. The audit trail is the deliverable.

Final Thoughts on Cutting Through AI Washing in Market Research

The question is not whether a tool calls itself AI. The question is whether your team can click through a finding, read the source, check the date, and defend the number in a board meeting. That audit trail is what separates a research tool from a confident-sounding guess. Merciv's enterprise offering is built around that standard. If your team is running this evaluation now, the four tests above apply to any live pilot as well. Run them on your own data at merciv.com/enterprise to see how the answers hold up against a real corpus rather than a prepared demo.

FAQ

How do I test whether an AI market research vendor's claims are real before signing a contract?

Run four tests in one afternoon on your own data, not the vendor's curated demo: send the same prompt five times in fresh sessions and check for materially different answers; ask a question you already know the answer to and compare outputs line by line; ask something the corpus cannot support and see whether the tool admits it; then open any cited claim and read the surrounding paragraph to verify the passage actually supports what's written above it. Send the same test set to every vendor in your shortlist and log results in a shared sheet. The gaps between answers will tell you more than any sales call.

What's the fastest way to spot AI washing in a vendor demo for market research tools?

Ask the rep to click a specific sentence in a sample output and land on the exact source page with a retrieval date — if it requires a follow-up email, that's your answer. Two other tells: run the same prompt twice and check whether findings shift materially (run-to-run drift means the system is guessing, not reasoning over a fixed evidence set), and ask what the model gets wrong — a real answer exists for every credible system, and a vendor who can't name a limitation is describing marketing, not a product.

What are the three types of AI hallucination that matter most when assessing AI claims in market research tools?

Factual hallucination (an invented statistic or quote), citation hallucination (a plausible-sounding source that does not exist), and misgrounding (a real citation that does not actually support the claim attached to it) are the three failure modes — and misgrounding is the hardest to catch because the source exists and the output reads as credible. A Stanford study published in the Journal of Empirical Legal Studies found specialized RAG tools hallucinated between 17 and 33 percent of the time. A fabricated figure that lands in a board deck acquires institutional weight fast; six weeks later it gets quoted back as consensus, which is why clicking through to the underlying paragraph is the only reliable check.

ChatGPT vs Merciv for assessing AI claims in consumer research: where does each one ceiling?

ChatGPT is genuinely useful for drafting, coding open-ends, and quick synthesis when you're not presenting findings to a CFO or CMO. The ceiling appears when you need to defend a number: no source attribution, no confidence scoring, no audit trail, and run-to-run drift that can flip findings 180 degrees when a single word in the prompt changes. Merciv's outputs carry a three-tier confidence score, a named source with a retrieval date, and a click-through to the underlying page on every finding. That structure lets anyone pressure-test a claim before the meeting starts, not after it goes sideways.

Can I build an internal RAG tool that passes the same AI washing tests a procurement team would run?

You can get the retrieval layer working, but the governance layer is where internal builds quietly fail the test: no productized audit trail, no confidence scoring applied at the claim level, and no licensed syndicated data rights — generic AI tools legally can't access that corpus, and neither can a RAG build running over your internal files. SOC 2 Type II certification, a zero-training policy enforced at the infrastructure level, and true tenant isolation each require independent engineering work that rarely appears in the original build cost model, and adding them post-hoc is not a sprint.