When summaries are free,
reality gets expensive

A strange thing is happening to knowledge work. It is getting easier to sound informed without becoming much better informed. I use AI constantly, and some of the best things it gives me are creative. It helps me develop half-formed ideas, connect thoughts across economics, markets, history and psychology, and turn them into research that is more interesting and more fun to produce. At times, it feels like a small personal renaissance. More ideas seem worth exploring because the distance between having a thought and doing something with it has shortened. It feels like a thinking partner that expands what I can attempt.

That is also why I take its risks seriously. A tool this useful can change how we think, including which parts of thinking we stop practising. A clear summary used to suggest that someone had spent time with the material. Now it may just mean they prompted well. The summary can still be valuable. It simply tells you less about how much its author understands.

There is a second problem, and I think it matters more. Summarising was not only how you demonstrated understanding. Sometimes it was how you developed it. It is the same reason teaching can be such a good way to learn. You think you understand something until you have to explain it clearly to someone else. Then the gaps appear. You cannot quite connect one step to the next, or answer what seemed like an obvious question. Working through that difficulty is part of the learning.

That is what makes me uneasy about being a junior analyst now. You can produce work that looks experienced before you have had enough experience to judge it. Consider a budget release on a trading desk. The documents arrive, prices are moving, and someone wants to know what matters.

You work through the revenue assumptions, spending plans and financing tables. Is the deficit falling because policy has improved, because the economy is larger, or because payments have been pushed into the future? Does the borrowing plan imply more bond supply than the desk expected? You bounce a view off a colleague. They challenge a definition or an assumption. A trader asks what is already priced. You revise the argument and explain which part you are still unsure about. The note is the visible output. Along the way, you are learning how a budget works, how a forecast fails and how to turn an observation into a decision.

Some of that process is inefficient. Some of it is the apprenticeship. AI can help with almost every step. But if you receive the interpretation before trying to form one, you have skipped an opportunity to practise. Reading a convincing answer does not necessarily teach you how to reach it, or recognise when it is wrong. The result can be a junior analyst who writes better notes without becoming proportionately better at analysis.

That outcome is not inevitable. Research has found both weaker unaided learning after students used an answer-giving AI and evidence of durable learning among workers using AI assistance. The settings differ, but they point to an important distinction: a tool can help people practise, or help them bypass practice. The responsibility extends beyond the junior. If a desk rewards speed and polish while removing opportunities to work through problems with colleagues, independent judgment may have less room to develop.[1][2]

A firm can improve this quarter’s output while underinvesting in the people it will depend on five years from now. “Humans will provide the judgment” is an incomplete answer if we automate the work through which humans acquire it. That judgment matters partly because the information needed for a good decision is not always sitting in a document.

Some of it is held by the contractor waiting for payment, the importer trying to obtain dollars, or the official deciding which expenditure can be postponed. It may emerge in a local meeting, an expert webinar or a conversation that only becomes useful after the third follow-up question. AI can help you prepare for those conversations and make sense of what you learn. But a model working from the material available to it cannot know an unreported payment delay merely because you ask the question well. A plausible account of what usually happens is not evidence of what is happening now.

This is where relationships and access matter. Knowing whom to ask, earning a candid answer and recognising what deserves another question are part of research. A contact list alone is not enough. The relationship has to produce information you can evaluate.

Take Sri Lanka. The domestic economy was rebounding while bond negotiations were still going on. Real GDP grew by 5 per cent in 2024. The pace and durability of that recovery mattered for the securities emerging from restructuring. Sri Lanka’s macro-linked bonds tied future payment adjustments to economic outcomes, including average dollar GDP over 2025–27 and a real-growth condition. A stronger recovery could affect both the amounts bondholders would receive and the perceived risk of receiving them. [3][4]

The question was how the growth outlook translated into the different payout scenarios, and how much was already reflected in price. Local knowledge could help assess the recovery. The bond contract determined how that view translated into value. Senegal illustrates the opposite tension. Signs of slowing capital expenditure and the authorities’ efforts to restrain spending called for a closer look beneath the oil-and-gas growth story.

The IMF reported 12.1 per cent headline growth in the first quarter of 2025, but only 3.1 per cent outside hydrocarbons, with payment arrears weighing on construction. A new oil project can lift GDP sharply while domestic businesses remain under pressure. As production levels off, the initial boost to growth fades; government revenues follow their own path, depending on prices, costs and the state’s share.[5]

To understand the consequences, you need to know which contractors are being paid, which projects are slowing and where spending restraint is actually falling. Published data can establish part of the picture. Local contacts and subject experts can help establish what the aggregate figures leave out. None of this replaces the data. But it changes how you read the data.

These are also examples of how human research and AI can strengthen each other. A conversation changes your working assumption. AI helps you trace its implications, compare it with other cases and identify what to check next. You return to the source with a better question. AI can increase the value of a good conversation by helping you do more with what you learn. There is no permanent human monopoly on this process. Once a conversation is recorded or fresh information becomes accessible, AI can help analyse it too. The advantage comes from access to relevant evidence, the questions that draw it out and the judgment used to assess it. It has to be renewed.

Nor is first-hand experience automatically reliable. A memorable meeting can create fake conviction as easily as a polished summary. An official has incentives. A business owner sees part of the economy. A busy restaurant may tell you very little about the country around it. Texture tells you where to investigate. It does not relieve you of the need to investigate.

The same applies to primary documents. A filing is not valuable simply because it is original. The useful detail may be a footnote nobody has connected to the financing plan, or a contradiction between what management says and what customers do. Sometimes the information is new. Sometimes the interpretation is. AI can help with both. The discipline is to keep the explanation connected to evidence that could change it.

That is what worries me about fluency. It can make borrowed opinions sound owned. It can make the difficult part feel finished before you have done it. The danger is not that the model is dumb. The danger is that it is good enough to make you stop checking.

For firms, some of the time saved should become more discussion, more feedback and more opportunities for juniors to make and defend a judgment. Automating tedious work is progress. Preserving the learning it once contained requires thought. The best version of this feels expansive. You can explore more ideas, bring more disciplines into the same conversation and produce work that would previously have been out of reach.

But the quality of that partnership still depends on what you bring to it: knowledge you have worked to understand, people who can tell you something new, and a willingness to discover that your attractive explanation is wrong. The output should improve. So should the person producing it. Use AI to compress the archive. Spend some of the time you save expanding it.