Glide Live: London 2026 - AI in practice when trust is paramount

A scientific publisher saw 5.7 billion bot requests in 16 days. An AI site scraped a drug formulary and recommended a fatal child dosage. Two trust-critical brands on where AI helps and where it cannot be used.

by Rich Fairbairn

Published: 19:47, 01 July 2026

"We co-publish a formulary used by pharmacists to determine dosages. An AI-driven site scraped it and recommend an adult dosage to a child, which would have been fatal." 

That quote reflects the tone for what happens when AI meets content where accuracy is life-or-death. 

In this fascinating talk the discussion centred on actually using AI in an environment where trust is critical, and where it is NOT being used. It also touched on the tactics publishers are being forced to take when their content is taken by AI companies and at times muddled or misattributed with potentially deadly effects.

A consumer trust brand (subscription-funded, no advertising, tests thousands of products per year) and a scientific publisher (65+ journals, content influences clinical guidelines globally) shared their approaches to AI - where it helps, where it cannot be used, and how to draw lines that hold.

The unifying principle: AI is a tool, not a decision-maker and human-in-the-loop is non-negotiable. But the panel also argued that the biggest risk most publishers face is under-investment, not over-investment.

The scale of AI scraping is not theoretical

One scientific publisher recorded 5.7 billion bot requests in a 16-day window, approximately 300x normal traffic levels. This isn't background noise. It's industrial-scale extraction of content that influences clinical guidelines.

50% of consumers prefer brands that avoid GenAI in consumer-facing content. Even if publishers could guarantee zero errors, consumer resentment may constrain what's possible. Source: Gartner, March 2026

Stat: Bot traffic vs normal (scientific publisher, 16-day window):

• Bot hits: 5.7 billion
• Expected: ~19 million
• That's ~300x normal traffic in just 16 days.

The "hollow papers" problem: Academics using AI signifiers now publish 30% more papers - they are technically correct, but feel lacking in stance or genuine insight. The output is hollow - accurate but purposelss. This same risk applies to any publisher that uses AI for content creation without editorial judgement layered on top.

Trust as a licensable product

"Code is no longer a moat. The moat is unique data, clean data, understood data, and trust and brand."

A human-curated knowledge graph is being licensed to third parties precisely because it's guaranteed AI-free. Trust has become a licensable product in itself. One mentioned licensing deal involves a major tech company using publisher knowledge to power health recommendations.

The four-tier model for AI use

AI use discussed in this session is organised into four tiers: 

1. Everyday productivity - data summaries, spreadsheets, internal communication 
2. CMS-embedded features - headline suggestions, content summaries, taxonomy tagging 
3. API-scale work - bulk data gathering, automated processing pipelines 
4. Bespoke tools - consumer-facing chatbots, internal review assistants

A consumer-facing chatbot was deliberately launched with a very narrow scope, giving product recommendations only, not "ask us anything." Scope management controls cost and risk. Token monitoring from day one. LLM-model-agnostic approach across multiple providers.

Under-investment is the bigger risk

AI rollout is a people problem rather than a technology one. The existing tools can do sophisticated things today. The bottleneck is building human confidence in the output, both internally and with external audiences.

Staff adoption metric: "If we took away the AI tools tomorrow, how upset would people be?" - tracking whether adoption is genuinely useful regardless of what people say about the technology.

Data quality is the real bottleneck

Internal tools that surface insights are only as good as the clean data beneath them. One organisation built an analytics tool in half a day - but can't release it widely because the underlying data tagging isn't clean enough. Tacit organisational knowledge doesn't transfer automatically.

An AI peer review assistant can extract every statistical claim from a research paper and re-check whether the evidence supports the conclusions. Still in proof of concept - editors require evidence equivalent to a randomised controlled trial before trusting it.

Quotes from the room

"AI is confidently wrong. Lots of people are beginning to look at content which isn't very good. One trend we might see is a return to trusted media because the generative sources will be incorrect."

"We found a website that is a complete rip-off of our site with our branding. Our Head of Research Integrity's profile is on there talking about the high research integrity of the site. It's nuts."

"The papers are technically correct, accurate, but they don't have a stance in the world. They don't really have any position. They're hollow."

"You can't take the stance of banning AI use. A controlled study showed reviewers who pledged not to use AI used it anyway."

"50% of consumers prefer brands that avoid GenAI in consumer-facing content. Even if you could guarantee zero errors, does that consumer resentment constrain what you can do?"

More from Glide Live: London 2026

• Keynote: what's ahead for publishing in an AI-dominated web
  
• SEO in an AI world: what's still working for publishers in 2026
• Revenue beyond the pageview: new business models for media companies
• A modern subscription success story: from 7-minute checkout to all-time highs
• Building audiences you actually own: habits, frequency, and the Rule of Three
• Sports clubs are publishers now: content operations, player audiences, and owned-channel revenue
• Eight themes from across the day

Glide Live operates under Chatham House Rules. No speakers or organisations are identified in these summaries unless specifically requested or authorised.