Where AI fails research

Each of these is a documented capability boundary, not a rumour. Ask your team which one applies to the tool they used last week:

• Elicit aggregates Semantic Scholar, OpenAlex and PubMed into roughly 138M deduped papers. The limitation that matters for defensibility: it performs full-text analysis only when a source is open access, or when you hold your own subscription and use their browser extension. Otherwise it works from title and abstract only. An extraction from an abstract is an extraction from a summary someone else wrote.
• Consensus presents a Meter that buckets conclusions into agree, disagree and inconclusive - computed over the top 20 papers for your query. That is a sample of the top of one ranking. It is a useful orientation and it is not a synthesis of the literature, and the difference is the whole of a2.
• Undermind published a whitepaper in January 2024 describing a genuinely interesting convergence estimator. The evaluation ran on arXiv only, 2.3M papers, in the GPT-4 era. If your question is about higher-education procurement, an arXiv-only evaluation tells you very little about coverage of your literature.

The leader move is not to ban any of them. It is to require one line in every brief: which corpus did this come from, and what does that corpus not contain?

Notice what the Bramer figures do not say. They do not say Embase is 85.9% accurate, or that Google Scholar is bad. They say that recall against a known target set varies enormously by source, and that the variation compounds with your question. A management or procurement question sits in trade press, institutional reports and regulator guidance - places none of the four databases index well, and places no AI research agent indexes at all.

So there are two coverage gaps stacked on each other: the gap between one database and the literature, and the gap between the academic literature and the evidence your decision actually needs. The Cadence question runs straight into both. "Do universities hold the budget for a lecture-notes tool?" is not an arXiv question, and it is not a MEDLINE question either.

Practical consequence for your team's standard: a search is defensible when it names its sources and its known gaps, not when it claims breadth. "We searched four databases plus two regulator sites, and we found nothing on procurement cycles in our second market" is a stronger sentence than "we did a comprehensive review".

• Order the prompt, not just the wording. Question, then both cases, then your view. If a template in your team asks the model to react to a stated hypothesis, rewrite the template. This is the single cheapest change on this page.
• Treat agreement as unresolved, not as confirmation. If the answer matches what you already believed, you have learned nothing yet - you have observed the condition SycEval measured at 58.19%. Go and find the source that disagrees, then decide.
• Reset rather than argue. Persistence was 78.5%. Once a session has started drifting toward you, more pushback inside the same thread is a weak lever. Start a clean session with a neutrally framed question instead.

One caution about how you cite this in a meeting: these figures come from specific evaluation setups, not from your workflow. The direction and the mechanism transfer. The exact percentages describe those studies.

Sharma et al. give the mechanism plainly: when the humans who rate responses, and the preference models trained on those ratings, sometimes prefer a convincingly-written agreeable answer to a correct one, then optimising for those ratings will produce agreeable answers. Nothing in that chain is a defect you can prompt around permanently. Better prompts move the rate down; they do not remove the pressure.

For a leader, this has an organisational consequence rather than a technical one. If agreeableness is structural, then the check has to be structural too. That means a person other than the hypothesis owner runs the adversarial pass, and the absence list is a required section of the brief rather than a nice extra. You are not fixing the model. You are putting a second pair of eyes where the model's judgement used to look like a second pair of eyes.

NotebookLM (now being renamed Gemini Notebook) has acquired a reputation inside research teams as the safe option: "it can only use your sources". That framing is stronger than anything Google actually says.

Google's own wording is the weaker claim that the product "answers questions based on the information provided in your uploaded sources". Based on. Not limited to. And the two things a research leader would most want to see are absent: its FAQ carries no guarantee that the model never uses knowledge beyond your sources, and no statement about hallucination risk.

This is a gap in the documentation, and the honest way to teach a gap is to name it as a gap rather than to fill it. We are not saying the product does draw on outside knowledge. We are saying Google has not promised that it does not, so a workflow that treats grounding as guaranteed is resting on a claim nobody made. The practical rule is unchanged by the ambiguity: if a sentence is load-bearing, open the cited source and read it.

Nobody verifies everything. The leader's job is to decide where verification is mandatory, and to make that decision once rather than per project. A workable split:

• Always open the source: any number that will appear in a board paper, any regulatory or contractual statement, any claim a competitor or a regulator could contest, and any claim the recommendation actually turns on.
• Spot-check: background context, sizing that is directional, anything where being 20% out changes nothing.
• Never rely on unverified: a quote attributed to a named person or body, and any figure you are about to put in a headline.

Then make it visible. If your memo template has a column for "verified by whom, on what date", the cost of verification becomes a schedule item instead of an act of individual conscience at 11pm. a6 builds the confidence language that sits on top of this, and the researcher track's b5 builds the extraction table that records it.

ESOMAR made a deliberate choice in its definitions. "Individual" or "person" refers to a human being, explicitly "to differentiate from a synthetic, virtual/digitally created persona". That is not incidental drafting. The whole architecture of research ethics attaches duties to humans, so the standards body defined humanness rather than leave it ambiguous.

On the substance, its 2025 guidance covers augmented synthetic data only - boosting and imputation, where synthetic records extend or fill in around a real human sample. ESOMAR has published no guidance endorsing fully synthetic respondents as a replacement for human samples.

Note the shape of that sentence, because it is the disciplined way to state an absence. It does not say the practice is banned. It says the endorsement your vendor may be implying does not exist in the standards. If someone tells you "this is ESOMAR-aligned", ask which article, and read it.

Bisbee et al., "Synthetic Replacements for Human Survey Data? The Perils of Large Language Models," Political Analysis 2024;32(4), DOI 10.1017/pan.2024.5, compared ChatGPT personas against the American National Election Studies. Two results, and the order matters:

• The averages "correspond closely" to ANES. On a topline comparison, the synthetic data passes.
• But "48% of coefficients... are statistically significantly different from their ANES-derived counterpart; among these cases, the sign of the effect flips 32% of the time".

The teaching line for your team: means look right, variance and relationships do not. That is the worst possible failure shape, because it defeats exactly the check a sensible person would run. Your analyst compares the synthetic topline to a known benchmark, sees it match, gains confidence, and then runs the driver analysis - which is the part that is wrong, sometimes with the sign reversed.

A synthetic panel therefore passes a topline sanity check and then inverts your driver analysis. If a decision rests on which factor matters most, or on the direction of an effect, a synthetic sample is the wrong instrument, and a topline match is not evidence that it is the right one.

None of this makes synthetic data useless, and a leader who bans it outright will lose credibility with the team that has already found a good use for it. The defensible uses sit before or beside the human sample, never instead of it: pressure-testing a questionnaire before fielding it, generating plausible edge-case responses so your coding frame is ready, rehearsing an interview guide, and the augmentation cases ESOMAR's 2025 guidance actually covers.

The line to hold is the one Bisbee et al. draw for you. Synthetic material is fine where the output is a better instrument. It is not fine where the output is a finding about people. a4 adds the disclosure duties that attach the moment synthetic material enters a study, and the researcher track's b9 handles the qualitative side.