Issue 020: Research, Intelligence and the AI Lens

AI-augmented research has split into two populations. One uses the model to accelerate discovery, then verifies the evidence. The other accepts fluent output because it looks finished. The first group is faster and more rigorous than the unaugmented baseline. The second is faster and more wrong.

The evidence on AI productivity supports the split; it is bimodal, not uniform. A Massachusetts Institute of Technology (MIT) study found that access to ChatGPT cut completion time for professional writing tasks by around 40 per cent and raised quality scores by 18 per cent, but the researchers deliberately excluded tasks requiring extensive fact-checking; it measured assisted production inside a controlled boundary. A Stanford and MIT field study of customer support agents found average gains of about 14 per cent, concentrated among less experienced workers. AI did not help every person or task equally.

The sharpest warning comes from the Harvard Business School jagged-frontier experiment. Consultants using AI performed better on tasks inside the model's capability frontier and worse on tasks just outside it, while the output stayed persuasive throughout. The model did not become visibly less fluent as it became less useful. And citation hallucination remains unsolved in general-purpose models: Nature reports that hallucinated citations are polluting the scientific literature. Retrieval reduces the risk when implemented well, but retrieval is not verification; a source can exist and still fail to support the claim attributed to it.

Which is why the multiplier is the habit, not the model. Everyone in your market has access to the same frontier models this quarter. The gap between the two populations is a workflow choice: whether a material claim gets checked against a primary source before it ships. The habit compounds; the models get replaced.

The strongest counter-argument deserves stating plainly. In qualitative synthesis, early-stage strategy and hypothesis generation, verification is hard to define and the productivity gain can outweigh the cost of checking every sentence. Correct. Exploration needs room for unverified possibilities. Client deliverables, investment cases, policy recommendations and published claims do not.

So the recommendation is concrete. Use AI freely to expand the search space, and require evidence before narrowing to a decision. For every material claim: confirm the source exists, is authoritative, is current enough for the decision, and supports the precise wording used. Record that check. The model accelerates the work. The verification trail is what makes the work defensible.

Elicit is a research platform for finding, screening and extracting information from academic papers. It fits consultants, analysts, journalists and policy teams who need a structured literature scan without becoming database specialists. Its strongest feature is not that it writes prose; it organises evidence into a form a human can inspect.

Better suited to research triage than a general chatbot. Users search a question, compare papers, screen larger sets and add extraction columns for specific variables. The Pro plan supports screening up to 5,000 papers, up to 20 table columns, reports drawing on as many as 135 sources, alerts, templates and Application Programming Interface (API) access; Scale raises the limits and adds team features.

Elicit still needs supervision. Database coverage is not universal, relevance ranking is not the same as methodological quality, and an extracted sentence can lose the context that changes its meaning. Use it to build the evidence map, not to outsource the conclusion.

Try Elicit →

Pricing verified on the Elicit pricing page as at 15 July 2026. Approximate GBP conversions at £1 ≈ $1.35; annual billing works out at ~£436 ($588) for Pro and ~£1,502 ($2,028) for Scale. Taxes and card spreads may change the amount charged.

Research firms will soon package verification as a product, not a footnote. By 15 July 2027, I predict at least one major global consultancy will publish a named "verified AI research" methodology and sell its audit trail as part of a paid service tier. The premium will not be for access to the model; everyone will have that. The premium will be for evidence that a qualified person checked the output.

Language will shift with the service model. By 31 December 2026, "augmented research" will appear in serious analyst and consultancy job descriptions more often than "AI research", as employers try to signal human accountability rather than automated production. Both predictions are specific enough to check, and I will mark them against reality in this newsletter when the dates arrive.

If you ran the verification triage on an AI research output this week, hit reply and tell me what the model fabricated. I read every response.