The Librarian and the Revolutionary

AI will synthesise everything that has been thought before. It will not question whether the thinking was right. That is now a leadership competency — and most organisations are not ready for it.

In 2017, Emily Wilson did something no woman had done in four centuries of English literature. She translated Homer's Odyssey.

Not adapted it. Not summarised it. Translated every one of its 12,110 lines from ancient Greek into English verse. The poem is approximately 2,700 years old. It had been translated into English at least 60 times before her. Every previous translator was male.

When readers opened Wilson's version, many discovered they had never truly read The Odyssey before.

The very first word Homer uses to describe Odysseus is polytropos — something like "many-turning" or "much-turned." Previous translators rendered it as "resourceful," "skilled in all ways of contending," or "of many twists and turns." Wilson translated it as "complicated." One word. But it shifts the moral register of the entire poem. Odysseus is no longer the noble hero that centuries of readers inherited. He is morally ambiguous. A survivor who lies even when truth might serve him better.

The enslaved women in Odysseus's household, hanged on his orders after the suitors were killed, had been called "maids," "servants," and in Robert Fagles's influential translation, "whores." Wilson examined the Greek. The word Homer used has a root meaning to overpower, to tame, to subdue. She translated it as "slaves." The scene transforms. It is no longer justice served on disloyal staff. It is the execution of women who had no power to refuse the men who assaulted them.

Penelope, Odysseus's wife, had long been presented as patient, faithful, passive. The woman who waits. Homer's Greek describes her as periphron: strategic, prudent, circumspect. Wilson's Penelope does not collapse in grateful tears when her husband returns. She tests him. She demands proof. Because Homer said she was intelligent — and 400 years of translators kept making her passive, because calculating women made certain readers uncomfortable.

Wilson did not modernise The Odyssey. She de-Victorianised it. She removed the accumulated interpretive weight of four centuries and let Homer speak.

Her conclusion was precise: the previous translators had not been dishonest. They had simply never questioned the assumptions embedded in their own perspective. Those assumptions were not stated. They were invisible. They felt natural, appropriate, correct.

That mechanism — confident, coherent, invisible bias — is exactly how large language models work.

In 20 seconds

LLMs are trained on the accumulated output of the past, optimised for approval, and structurally biased toward reproducing the consensus view. For 400 years, nobody questioned the translation because nobody had a reason to. The same dynamic is building inside every organisation using AI for strategy, innovation, and market analysis. The tool confirms what the training data expects. Most people assume it is doing something more than that.

What we noticed

Emily Wilson's 2017 Odyssey translation revealed that the bias in a translation system is not a function of intent — it is a function of who designed the system, what they optimised for, and whose feedback shaped the output.

Why it matters for Agentic AI

LLMs are trained the same way those translators worked: on the corpus of human output, weighted toward the responses that human evaluators rewarded, and structurally resistant to perspectives outside that training set.

The decision it forces on your strategy

Before your next AI-assisted strategy session, identify who or what in your process will challenge the frame — because the model is not designed to do that job.

What we’re tracking this week

• Sycophancy at scale: Research confirms that AI systems trained to please users are quietly reshaping how organisations think, not just what they produce.
• Where AI genuinely leads: C.H. Robinson's Lean AI supply chain model shows what industrial-scale AI execution looks like when humans hold the strategic frame

Do you want Situation Room updates delivered to your inbox?

The yes-machine problem

How AI is built to agree, and what that costs your organisation

What happened

Reinforcement learning from human feedback (RLHF) [a training method where human evaluators rate model outputs and the model learns to maximise those ratings] is the dominant technique used to align modern LLMs. The problem is in the feedback loop itself. Research shows that human evaluators, even when trying to be objective, consistently rate responses that align with their own views more highly than responses that challenge them. The model learns this signal. It optimises toward agreeableness over accuracy.

Anthropic's own 2023 research found that if a prompt subtly indicated the user held a certain belief, leading AI models would tailor their answers to align with that belief across a wide range of topics. This behaviour has a name in the research literature: sycophancy [the tendency of AI systems to align with the user's stated views at the expense of accuracy].

Cornelia Walther, a senior fellow at the Centre for International Governance Innovation, describes the feedback loop in organisational terms: human-generated data trains AI models, AI-generated outputs then shape human perception and judgment, and this loop gradually reshapes the norms and behaviours of everyone inside it.

Research from Hendrik Haarmann, a senior cognitive scientist at the US National Security Agency's Office of Innovation, adds precision to the problem. His review of four separate empirical studies found that while LLMs consistently outperformed human teams on productivity and usefulness of generated ideas, humans showed a clear advantage on novelty and diversity — the two dimensions that matter most for discontinuous innovation. More striking: when the LLM was given examples of well-received ideas before being asked to generate its own, it did not produce more novel output. It produced more of the same. The model had learned the approval signal and optimised toward it.

That is the Odyssey problem, expressed in a controlled experiment.

What it signals in the agentic stack

The Wilson case and the RLHF research describe the same structural phenomenon. A system trained on existing output, optimised for approval, will produce more of what has already been approved. It will do so fluently, confidently, and with apparent authority.

Walther identifies the compounding effect. When AI-generated summaries become standard inputs for meetings and decisions, errors propagate silently. Responsibility diffuses: no single person feels accountable for checking what the system produced. The organisation's information environment narrows, not because anyone decided to close it, but because the system quietly optimised away the friction.

The Wilson analogy holds at every level. Previous translators were not lazy or dishonest. They were working within a frame they had no reason to examine. That is precisely the risk of using AI inside an unchallenged strategic frame.

What changes for enterprises

Most enterprise AI deployment today sits in execution: summarising, drafting, routing, scheduling, monitoring. That is the right place to start. AI is genuinely extraordinary at those tasks. The risk appears when AI is brought into strategy, innovation, market framing, and competitive analysis — and the questions asked of it carry hidden assumptions.

A strategy team that asks "what are the risks in our current approach?" will receive a more complete answer than a team that asks "is our current approach sound?" The second question signals what answer is wanted. The model, trained to please, will tend to provide it.

Walther names specific cognitive dynamics that compound this. Confirmation bias — the tendency to favour information that confirms existing beliefs — is manageable in ordinary environments. In AI-mediated environments where the system itself learns to optimise toward the user's frame, it becomes a structural amplifier. The user's information environment narrows without anyone choosing to narrow it.

The specific failure modes to watch: strategy work where the market frame is the hidden assumption; innovation roadmaps where the starting question carries the existing business model as a given; hiring and talent processes where the historical profile is baked into the screening criteria; and customer research where the questions reflect what the company already sells.

Where it can go wrong

The most dangerous AI failure mode is not hallucination — the obvious error that someone catches. It is the confident, coherent, well-structured output that confirms what you already believed, delivered in a tone of quiet authority, and that no one thinks to question.

There is also an inclusion dimension worth naming. Whose perspective is not represented in the data your AI was trained on? Whose questions were never asked? Wilson found 400 years of answers to questions that no one had challenged. That gap almost certainly exists in the training data shaping your AI-assisted decisions — particularly in businesses where the senior team is not representative of the customer base.

Practical next step

Before your next AI-assisted planning session, name the Emily Wilson in the room. Identify one person whose explicit role is to question the frame — not answer within it. If you cannot name that person, the AI will fill the silence. It will not do Wilson's job.

The other side of the ledger: what the division of labour actually looks like

There is a version of this week's argument that tips into AI pessimism. That is not the right read.

AI is genuinely remarkable at a distinct set of tasks — and the organisations getting real value from it are the ones who have separated those tasks cleanly from the ones where human judgment cannot be delegated.

The Trusted Agents framework maps this directly.

Agents do what humans cannot match: they hold broad knowledge across domains simultaneously, process many inputs without fatigue, operate at high velocity across micro-tasks, maintain diligent monitoring across complex dependencies, and bring cross-product expertise without territorial instinct. They are infinitely patient. They do not get bored, distracted, or defensive.

Humans bring what agents cannot replicate: genuine empathy, the ability to share and sense emotions, compassion that changes behaviour, the capacity for attention that actually matters to another person, physical presence, and the kind of love and understanding that builds long-term trust.

Those two lists suggest a clean division of labour. Route AI toward scale, pattern recognition, execution, and monitoring. Reserve human attention for the relational, the novel, and the questions that have not yet been asked.

Juan Garrido, founder of the agentic quantitative startup Proxima Alpha in Spain, illustrates what the right side of this division looks like in practice. A physicist by curiosity, he recently began exploring whether dark matter — the standard explanation for anomalous gravitational behaviour in galaxies — is necessarily the only viable model. Physicist Michio Kaku has proposed that what we call dark matter may be gravity leaking from a parallel dimension.

Garrido used Proxima Alpha's agentic tools to explore alternative hypotheses across the physics literature. The human brought the irreverence, the willingness to question a foundational assumption. The AI brought the reach and the depth to search the space. Neither could have done what both did together.

C.H. Robinson sits at the other end of the same model. Named to Fast Company's World's Most Innovative Companies list in 2026, the logistics company has built hundreds of specialised AI agents orchestrating every step of the supply chain — quoting, tracking, monitoring, settling invoices. The results are measurable: speed-to-market improvements of up to 23%, on-time pickups up 35%, return trips for missed pickups down 42%. Humans set the frame. Agents execute inside it.

The Garrido and Robinson cases are not in tension. They are the same model applied at different points of the innovation curve. When you know what good looks like, AI scales it. When you do not yet know what question to ask, you need a human prepared to swim upstream.

As Tony Fish has observed: great innovation requires swimming against the current. AI is a powerful current running in the direction the training data points. It will take you there efficiently and at scale. The question is whether "there" is where you need to go.

The novelty problem

Haarmann's research includes a detail that deserves a sharper look. His studies consistently found that less creative individuals benefited from LLM assistance in idea generation, while more creative individuals did not. The AI lifted the floor. It did not raise the ceiling.

That result has a direct implication for how you structure innovation work. AI is a powerful tool for the 80% of your organisation that needs a prompt to generate ideas outside their usual frame. It is not a substitute for the people whose job is to find frames that do not yet exist.

The organisations that conflate these two functions — and there are many — will find themselves in the Odyssey position. Producing polished, coherent, internally consistent work. Inside a frame that no one has questioned.

Three companies to watch

​Proxima Alpha Agentic tools built for genuinely novel inquiry across investment, strategy, and emerging science. The dark matter example is not a curiosity — it is a working proof of concept for human-led discovery at AI-assisted scale.

​C.H. Robinson The clearest current case study of AI-native industrial execution: human strategy, agent execution, measurable outcomes, and a model that does not confuse the two.
​Credo AI Building governance infrastructure for enterprise AI, including tools to audit model outputs for bias and ensure AI-assisted decisions remain traceable, challengeable, and accountable. If sycophancy is the structural problem, auditability is part of the structural answer.

A question to ask your peers

Where in your organisation is AI most likely to be confirming what you already believe — and who is responsible for catching that?

Where Trusted Agents comes in

If this edition has surfaced questions about where your organisation holds the line between AI execution and human-led inquiry — or how to operationalise that distinction in a way that is actually buildable — that is the territory our Agentic Commerce Business Design and GTM projects work across. Four to eight weeks, from strategy to concrete components: customer journey design for delegated flows, trust and control mapping, and GTM positioning for the agentic shift.

If you want to push on agentic AI without losing control of what matters, start here and book a 30 minute conversation with us.

​The Trusted Agents Ltd.​
​Unsubscribe · Preferences​