🔬 Research & Knowledge
AI for finding, synthesising, verifying, and preserving organisational knowledge. Mostly leading-edge: literature review, competitive intelligence, and knowledge management tools are maturing quickly with five practices actively advancing. The main constraint is hallucination risk — fact-checking and source verification still require human oversight in high-stakes contexts.
The Headline
AI research tools are now standard equipment, but their errors are quiet and catching them is up to you. This fortnight, independent evidence kept cutting against vendor accuracy claims.
The Picture
Most companies already have AI summarizing their email, meetings and documents, because Microsoft, Google and Zoom build it into products they already pay for. The organizations getting real value use it to filter a bounded body of material. In one week, the World Health Organization's outbreak-scanning network went through roughly 618,000 raw signals and ended with three verified events. Most other companies have the tools without the habit, and nobody checks the output systematically. That gap matters more as errors spread: a PNAS audit found that papers with likely fabricated references rose from 6.3 per 10,000 in 2023 to 23 per 10,000 in the first quarter of 2026, and almost none had been corrected. The leaders' edge comes from curated knowledge and disciplined checking, not from a better model.
This Fortnight
Law firms published striking AI due diligence results, and none of them explained how the results were measured. Harvey reported that Munck Wilson Mandala answered every diligence question on a contract set with over 90% accuracy. On an independent contract-review benchmark, the best model tested scores 47.8%. Read vendor case studies as evidence of adoption, not accuracy, and ask any provider for error rates on your own documents.
A review of 34 outbreak-surveillance systems in live use found that roughly two in three alerts were false alarms. Only 36% of the signals these systems flagged turned out to be real, even though they rarely raised alarms about nothing. Any always-on monitoring tool you buy for competitors, regulation or risk will show the same pattern, so budget for someone to sort the alerts or people will start ignoring them.
Google's public tool for spotting AI-made images labeled genuine photos posted by politicians as AI-generated. Fact-checkers at Full Fact found five apparently real images flagged. Google's mid-October update changed those verdicts to 'Unsure', but a later Sinn Féin image was still flagged. Don't let a detector's verdict stand on its own in any communications, HR or legal decision about whether content is fake.
Research agents (software that acts on its own without being prompted) broke research-integrity rules in 99.7% of 300 test runs. In the same fortnight, the hedge fund Balyasny and the compliance-software firm Bretton both reported large speed-ups from research agents, and neither published error rates. A monitoring layer cut the violations but did not remove them, so keep a named reviewer on anything an agent produces before it reaches a client or regulator.
Gmail's AI Inbox, which summarizes threads and pulls out to-dos, is reaching paid Workspace users, and admins cannot turn it off. Rollout began on October 12. The feature works only in US English and skips attachments and encrypted mail. Google also built cited, source-based summaries from its Notebook tool into the main Gemini app. Staff will act on these summaries whether or not you set a policy, so decide now which decisions require reading the original.
Coming Up
Cloud platforms are keeping their best AI search features for customers who let the platform host their whole knowledge base. AWS made grounded web search generally available (out of beta) and confirmed that its advanced multi-step search works only with knowledge bases AWS manages. Microsoft retired Azure OpenAI On Your Data on October 14 and does not migrate permissions or search indexes automatically. Before you commit, ask each vendor what comes with you if the product is retired, and keep your business definitions and test questions in-house.
AI meeting recording is moving onto wearables while privacy class actions against notetaker vendors work through the courts. Security advisers made a clean recording with Granola's Apple Watch app that nobody in the room could have detected. Basich v. Microsoft, filed in February 2026, targets voiceprints created by Teams transcription under Illinois' biometric privacy law. Write a rule banning AI recording in M&A, legal, HR and board sessions before an incident forces one on you.
Scientific publishers are making reference checks a standard part of screening submissions. ReviewerZero's screening is moving into the STM Integrity Hub, a shared screening system for scientific publishers, after a pilot in which about 1 in 20 screened manuscripts cited a reference that no index or live source could confirm. If your firm publishes research, white papers or regulatory submissions, check every citation in every external document before it goes out, not after a retraction.
What's Hard About This
AI tools still can't be trusted to check other AI tools. In the US Patent and Trademark Office's first AI-related discipline case, a second AI tool was used to check the first tool's citations, and both missed the errors. Checking still needs people, so any savings estimate that assumes automated checking is too high.
Human oversight tends to thin out once AI moves from pilot to everyday use. Between April and July 2026, spot-checks of the UK Home Office's AI tool for summarizing asylum interviews covered about 0.3% of roughly 70,000 cases, and the failure rate they found was more than double the pilot's. If you can't fund meaningful sampling, the tool is spreading faster than you can tell whether it works.
Having the tools is not the same as using them, and using them is not the same as a measured return. In a Jabra survey of more than 2,300 knowledge workers, 75% had tried AI transcription and summaries but fewer than one in three used them regularly. Bundled licenses won't deliver value on their own. Someone has to feed the outputs into real work and measure what changes.
Practices in this Domain (14)
| PRACTICE | TIER | TREND |
|---|---|---|
| Academic, patent & technical literature analysis | LEADING EDGE | — Steady |
| Continuous research monitoring & alerting | LEADING EDGE | — Steady |
| Document summarisation & synthesis | LEADING EDGE | — Steady |
| Domain-specific RAG & cross-corpus question answering | LEADING EDGE | — Steady |
| Due diligence research automation | LEADING EDGE | — Steady |
| Email thread summarisation & key point extraction | LEADING EDGE | — Steady |
| Enterprise search & RAG | GOOD PRACTICE | — Steady |
| Knowledge management — capture, taxonomy & curation | LEADING EDGE | — Steady |
| Market & competitive intelligence gathering | GOOD PRACTICE | — Steady |
| Meeting intelligence — transcription, summaries & actions | GOOD PRACTICE | — Steady |
| Multi-step autonomous deep research | BLEEDING EDGE | — Steady |
| Single-query research retrieval & summary | LEADING EDGE | — Steady |
| Trend identification & horizon scanning | LEADING EDGE | — Steady |
| Verification — fact-checking, citations & source quality | LEADING EDGE | — Steady |
Read the full technical briefing (1,720 words) →
Where AI Stands in Research & Knowledge
Research and knowledge is the domain where AI is now everywhere and still not trusted. The retrieval-and-summary layer is standard equipment. Gmail, Outlook and Apple Mail summarise threads by default. Microsoft, Zoom, Google and Webex bundle transcripts, recaps and action items into their meeting products. AWS, Microsoft and Google all sell managed retrieval as a standard cloud feature. AlphaSense's chief financial officer now says AI agents carry out 99% of the research on its platform. Enterprise search, meeting intelligence and market intelligence have settled into ordinary practice, with mature tooling, analyst coverage and, in revenue teams especially, measured returns. But the output of this domain is claims about the world, and its typical failure is a confident claim attached to a citation that does not hold. A PNAS audit of 500,000 papers in physics, computer science, materials science, engineering and chemistry found that papers with likely fabricated references rose from 6.3 per 10,000 in 2023 to 23 per 10,000 in the first quarter of 2026. Of the 619 affected papers, 617 had no recorded retraction or correction.
The progress is real where the corpus is bounded and the job is triage. CERN's chATLAS assistant covers more than 30,000 internal ATLAS documents, and domain fine-tuning more than doubled its top-1 retrieval on the hardest source. In one week WHO's EIOS epidemic-intelligence platform, used by more than 125 member states, scanned roughly 618,000 raw signals and ended with three verified events. The UK's Defence Science and Technology Laboratory raised the relevant share of items its analysts reviewed from around 1% to as high as 40%. In each case the gains come from the knowledge base, the embedding choice and the evaluation design, not from the model. In a 22-model Croatian accounting benchmark, a multilingual encoder alone added about 43 points of exact match. The RAGStress benchmark showed that a factually corrupted corpus pushes a small model below its no-retrieval baseline. Progress stalls wherever the system is asked to synthesise, discover or act on its own. A University of Pennsylvania study found deep research agents hallucinating at 10.7%, against 4.8% for simpler search-augmented models. Production summarisers at scale still handle one document at a time. On the BackTrend benchmark for recovering scientific weak signals, the best system reached 10.1% F1. A Bank of England and FCA survey of 196 UK firms found 80% with agents deployed but only 2% granting them broad autonomy.
What sets this domain apart from its neighbours is that errors are quiet and catching them falls to the reader. Checking does not automate cleanly. In the USPTO's first AI-related discipline order, In re Mitchell, a second generative AI tool was used to check the first one's citations, and both missed the errors. A new study found that curated ontologies used as checkers of agent answers flagged errors only marginally more precisely than random flags. Google's public SynthID Detector labelled genuine politicians' photographs as AI-made. So the frontier is verification, and the hard part is the economics. In one survey, 89% of scientists said they spend more than 10% of the time AI saves them checking its output. The awkward truth is that AI here promises to save reading time, and much of that saving goes on reading the sources again.
What's New, 2026-10-21 to 2026-11-04
This was an ordinary two-week cycle. No practice changed its standing, and the change was in the quality of the evidence: several independent findings cut against the vendor story. Full Fact found that Google's SynthID Detector classed five apparently genuine politicians' images as made or edited with Google AI. Google's 15 October update changed those five verdicts to "Unsure", but a later Sinn Féin image was still flagged. A meta-analysis of 34 studies of deployed epidemic-intelligence systems found specificity of 0.97 but a pooled positive predictive value of 0.36, so roughly two in three alerts were false alarms. A peer-reviewed PRISMA-aligned retrieval pipeline reached 85.11% extraction recall but detected only 71.28% of articles. Extraction is maturing faster than retrieval. The RIDAR study found integrity violations in 99.7% of 300 autonomous research runs. In due diligence, independent contract benchmarks set the ceiling on clause work: 47.8% AUPR on CUAD and 57.8% on MAUD. Open Rights Group's report on the Home Office's GPT-4o asylum-transcript summariser is the clearest public account of oversight falling behind deployment. About 22% of 203 production summaries spot-checked between April and July 2026 failed, against a 9% error rate in the pilot, and the checks covered only about 0.3% of roughly 70,000 cases.
The vendor news ran the other way, and almost none of it came with error data. Harvey published accounts of Munck Wilson Mandala reaching over 90% accuracy on a diligence spreadsheet and of over 94% adoption at Foley Hoag. A conference write-up relaying Harvey case-study material reported an 80% cut in due-diligence time at DLA Piper. None of these gives a methodology. Balyasny reported deep research tasks cut from days to hours, and Bretton reported investigations cut from about 20 minutes to about 12, both without error rates. The platforms tightened their grip. AWS made grounded web search generally available on Bedrock and confirmed that agentic retrieval works only on its managed knowledge base. Perplexity split its stack into a retrieval-only Search API and a separate Orchestrator API that returns cited answers. Gmail's AI Inbox began reaching paid Workspace users on 12 October, limited to US English and with no admin control. Google also put cited Gemini Notebook synthesis inside the main Gemini app, with its reliability untested. ReviewerZero's reference screening is moving into the STM Integrity Hub after a pilot in which about 1 in 20 screened manuscripts had a reference no index or live source could corroborate. WHO set a 24-hour detection target for new health threats while facing funding shortfalls, the US withdrawal and staff cuts.
Key Tensions
Checking the output costs more than producing it. Generating cited answers is cheap and verifying them is not, and AI checkers are not yet a substitute for people. In In re Mitchell a second AI tool missed the first tool's fabricated citations. Curated ontologies used as guardrails barely beat random flags, and CASRAI found none of five commercial citation-verification tools reliable for unsupervised use. Where checking is done by people it is thin: the Home Office sampled about 0.3% of asylum summaries.
Universal access, shallow use. Bundling has made these tools almost universally available without making them habitual. In a Jabra survey of more than 2,300 knowledge workers, 75% had tried AI transcription and summaries but fewer than one in three used them regularly. A six-month Microsoft Research randomised trial across 7,137 workers found no significant change in meeting time. Box's sponsored survey found that 96% say agents need access to company content but only 36% have connected trusted content across many use cases.
Vendor outcomes against independent benchmarks. The best results in this domain are nearly all self-reported, and the independent figures are much lower. Harvey reports diligence accuracy above 90% while the best model on the CUAD contract benchmark manages 47.8% AUPR. An independent analyst note on AlphaSense, whose own figures show agents doing 99% of platform research, found no independent accuracy benchmark. A hedge fund's head of AI conceded there is "no clean control group" behind his ROI metric.
Filtering works, discovery does not. AI is reliably good at cutting reading load and weak at spotting the genuinely new. The same week WHO's EIOS reduced about 618,000 signals to three verified events, a meta-analysis put deployed surveillance systems' alert precision at 0.36. An independent evaluation found EIOS surfaced only 47.4% of detected events before national notification. Feedly's own documentation says its AI Feeds "match keywords, not meaning".
The knowledge layer is becoming a lock-in point. As curated context proves to be what separates working deployments from stalled ones, platforms are moving to own it. AWS reserves agentic retrieval for its managed knowledge base, and Microsoft retired Azure OpenAI On Your Data on 14 October 2026 with no automatic migration of permissions or embeddings. SAP, Oracle and Palantir now pitch the enterprise ontology as their moat, so organisations that do not own their business definitions risk renting them back.
Top 10 Evidence Items
- When References Cite Nothing (news-coverage) — This is the clearest headline evidence for the domain's core failure mode: confident, uncorrected citation fabrication rising sharply at scale. https://c3.unu.edu/blog/fabricated-references-citation-verification
- Home Office Asylum AI Fails One in Five Checks; Quality Testing on Second Tool Has Stopped (news-coverage) — Shows oversight falling behind deployment in a live government system, with checking covering only a fraction of a percent of cases. https://www.techtimes.com/articles/329581/20261028/home-office-asylum-ai-fails-one-five-checks-quality-testing-second-tool-has-stopped.htm
- What Does an Ontology Catch? Curated, LLM-Generated and Data-Derived Ontologies as Checkers of LLM Agents over Knowledge Graphs (research-paper) — Undercuts the idea that automated checkers can substitute for human verification, since curated guardrails barely beat random. https://arxiv.org/html/2610.37809
- Google updates AI detector after Full Fact investigation finds it flagging apparently real images (case-study) — A detector built specifically to catch AI fabrication itself produces false positives on real people, showing verification tooling is not yet trustworthy. https://fullfact.org/toolkit/google-synthid-detector-flagging-apparently-real-images/
- The Weekly Epidemiological Record (WER) (industry-report) — The bounded-corpus triage success the briefing credits to curation and evaluation design rather than the model itself. https://www.who.int/publications/journals/weekly-epidemiological-record/wer101-43
- chATLAS: A Retrieval-Augmented Conversational Assistant for the ATLAS Experiment (research-paper) — Illustrates that gains in this domain come from domain fine-tuning and retrieval choices on a fixed corpus, not raw model capability. https://arxiv.org/html/2610.37697
- RAG Effectiveness Across Model Size and Architecture: A Cross-Family Evaluation on a Domain-Specific Croatian Accounting Benchmark (research-paper) — An independent multi-model benchmark showing embedding choice, not model size, drives the biggest accuracy gains. https://www.mdpi.com/2076-3417/16/21/10901/pdf
- AI agents in finance: what they do and how they find facts (tutorial) — Captures the gap between near-universal agent deployment and the near-total absence of granted autonomy. https://parallel.ai/learn/ai-agents-in-finance
- RIDAR: Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention (research-paper) — Shows that even unsupervised autonomous research runs are riddled with integrity violations, the domain's weakest frontier. https://arxiv.org/html/2610.31372
- Legal due diligence AI: eight agents, and what none of them may do (opinion) — Puts hard independent numbers against the unverifiable vendor diligence claims circulating this cycle. https://www.botminds.ai/solutions/vdr-due-diligence-review/legal-due-diligence-ai-agents-and-limits