The Day Knowledge Gets Cheap: 2027–2029, Value Shifts from "Knowing" to "Answering For It" ── Reading the slope of model progress: where trust, authenticity, and conscience belong in the AI era

The cheaper knowledge becomes to produce, the scarcer the thing that matters: not the knowledge itself, but the person who takes responsibility for it. Certainty over speed, context over volume, and the plain fact that a human did it. This article separates what will change from what will not, and looks for where trust and conscience belong in the AI era. It also spells out the conditions under which the forecast could be wrong.

01What we can already confirm: the price of knowledge has fallen

Before making predictions, let me list what can be confirmed today. With the spread of generative AI (= artificial intelligence that automatically produces text, images, and program code), the production cost (= the time and money it takes to make one item) of writing, images, and code has plunged. A draft introduction that would have cost tens of thousands of yen from an outside writer a few years ago can now be had in tens of seconds for a few dozen yen in API fees (= pay-per-use charges for calling AI from a program). Quality is still debatable, but the fact that the price of a "first draft" has dropped by an order of magnitude (to less than one-tenth) can be confirmed from published price lists. The figures above are typical examples, not exact measurements, but the direction is beyond doubt.

Access to knowledge has followed the same path. Knowledge that once lived only in an expert's head or an expensive reference book can now be pulled out of a conversational AI in seconds: an overview of a drug's mechanism of action (= how a medicine works inside the body), when to use which statistical method, a summary of a regulatory framework. Verifying accuracy is still human work, but the distance to knowledge has shrunk — not all at once, but steadily. Here an old principle of economics kicks in: what becomes abundant becomes cheap. As with water and salt, knowledge that is easy to obtain becomes hard to charge for on its own.

Brought down to the pharmaceutical workplace, it means this. "Knowing the relevant section of the guideline" and "having read the prior literature" have, until now, explained part of a professional's salary. That scarcity is eroding. The major conversational AI models have raised their coverage of knowledge and their speed of response with every generation. I can find no reason, so far, to think this pace of improvement will suddenly stall in 2027.

Everything up to this point is observation, not speculation. The rest of this article places an extrapolation (= extending the current trend forward) on top of that foundation. When knowledge becomes cheap, what will people pay for — with money and with trust? As a way of framing 2027–2029, let us first look back at history once.

02What handcrafts after industrialization teach us: a reversal of value

In nineteenth-century Britain, when machine-woven cloth and mass-produced tableware flooded the market, handmade goods lost their value for a time. For the same purpose, machines could make things cheaper, faster, and more uniformly. Craft was pushed off the economic stage. Then something unexpected happened: the fact of being "made by human hands" was itself re-valued. The Arts and Crafts movement of William Morris and others (= a nineteenth-century British design movement that reasserted the worth of handwork) is the emblem of this. Once machines took over "quantity," human involvement became a scarce good (= something valuable precisely because it is hard to obtain).

Strip out the structure and you get three stages.

• Stage one: machines (AI) take over mass production and the price of the product falls. Being able to make it stops being anyone's strength.
• Stage two: near-identical mass goods flood in and people tire of them; attention turns to provenance (= the record of how something came to be) — who made it, why, and how.
• Stage three: human involvement, judgment, and the location of responsibility get priced in. "A human did this" becomes something displayed, and a reason to choose.

I expect the relationship between generative AI and knowledge to follow the same structure. The more AI-written text and AI-made summaries fade into the everyday background, the more people will ask: who answers for this conclusion? Whose conscience did this judgment pass through? Already, debates over labeling requirements for AI-generated content, and moves to mark work as human-written, have begun in several countries. This reads like the doorstep of stage two.

The historical analogy has limits, though. Cloth and tableware are consumables — a choice occurs at every repurchase — whereas knowledge, once spread, can be copied at near-zero cost, and a market for "handmade knowledge" may not form in the same shape. The re-valuation of handcrafts also took decades; the reversal around AI may be faster, and there will remain domains where "cheap and good enough" keeps winning. History gives a hint about direction, but not about speed or scope.

03From "knowing" to "answering for it"

A line in a package insert (= the official document that sets out how a medicine is used and what to watch for) is still written by a human. Even when AI can produce the draft, if that line is misread in the field and a patient is harmed, "the AI wrote it" is no defense. The same goes for approved materials (= promotional and informational print pieces and slides that have passed internal review): the name of the person who signed the approval box stays on it to the end. Here is the true identity of work whose price does not fall even as knowledge gets cheap.

AI can produce answers. What it cannot do is bear the consequences of an answer. To bear them means standing in the position of the one who apologizes when things go wrong, putting your own name on a judgment, and being able to decide "we could release this, but we won't." Given the pace at which answer quality has improved over the past few model generations, by around 2027 the situations where a human beats AI on answer quality should be quite narrow. But the location of responsibility is fixed to humans by law, contract, and organizational design, independent of technical progress. When Japan's Pharmaceuticals and Medical Devices Act holds someone accountable for advertising, that someone is a company and its staff — not a model.

This is not a grand matter of virtue or ethics. It is a more everyday question: "Could I send this document out under my own name?" The difference between someone who forwards AI output as-is and someone who runs it through that question first is almost invisible in ordinary times. It becomes visible when something goes wrong — and at that moment, the first person has nothing they can explain. In a world of cheap knowledge, the habit of honestly passing every piece of work through that question becomes a kind of credit you cannot list on a résumé.

Put the other way around: what loses value inside organizations in 2027–2029 is the person who "knows the answer," not the person who "can answer for the answer." The value of remembering what can be looked up has been cut twice already, first by search engines and then by AI. What remains is the position of someone who makes fallible judgments at their own discretion, corrects them when they miss, and can stop a project by the light of conscience (= the inner sense that says "this is wrong"). If this prediction fails, it will be because a legal system emerges that grants AI output something like corporate personhood as a bearer of responsibility — and at present I can find no concrete legislative movement in that direction.

04Certainty over speed, context over volume: the axes of evaluation swap

Suppose a first draft of a material that used to take three weeks now takes three minutes with AI. The advantage of "the fast person" disappears — everyone takes three minutes. Once generation cost (= the effort of producing text or images) approaches zero, volume stops being a differentiator too; anyone can produce a hundred drafts. The axis of evaluation then moves from the side that makes to the side that verifies. Is this information sound? May this be said in this context? Those two questions.

Certainty is a matter of verification: did you check the source, did you reconcile the numbers against the original (= the underlying paper or official document)? The more fluent (= smooth and natural-sounding) an AI's output, the less an error looks like an error. Catching a plausible fabrication (= an invented fact) with human eyes becomes relatively heavier work the faster generation gets. Context means that the very same content can be permissible or not depending on the reader, the regulation, and the timing. A sentence that is fine for healthcare professionals can turn into advertising of an unapproved indication (= promoting a benefit that has not been approved) when aimed at the general public. Correctness cannot be judged outside its context.

Pharmaceutical material review has been doing exactly these two things for a long time: grading the evidence (= the scientific basis) and judging whether an expression is permissible in light of the intended reader and the regulations. So my read is that the cheaper AI makes generation, the sharper — not fainter — the outline of this occupation becomes. On one condition. If review stays at the level of correcting particles and phrasing, that is precisely what AI will replace. When the axes swap, the only reviews that benefit are the ones that have shifted their weight onto the substance: verifying certainty and judging context.

So what should you count, starting tomorrow? Not "how much did I produce today," but "of what I sent out today, how much did I verify down to the evidence?" The answer to the first question rises automatically once AI is adopted. The second does not. What does not increase is what becomes scarce.

05The value of "a human did this": authenticity as a new label

Several studies, as of 2026, estimate what share of text on the web is AI-generated or AI-assisted. The numbers swing widely with the method of measurement, but the direction is single: text, images, and summaries can now be mass-produced by machines. Then something odd happens. The bare facts — "a human wrote this," "this person actually experienced it" — begin to carry value as scarce information. Authenticity (= being genuine; knowing for certain who made something) rises as a label on an axis separate from quality.

The signs are already here: a return to services that sell handwork, face-to-face contact, and real names; the spreading adoption of Content Credentials (= an international standard for embedding a tamper-evident record of who made something and when into photos and documents); and regulatory debates in several countries requiring AI-generated content to be labeled as such. All of these are attempts to re-ask "where did this come from" in a world awash in generated output. The EU AI Act (= the EU's comprehensive AI regulation, adopted in 2024) mandates transparency labeling for generated content, with phased application beginning in 2026. That reads as a first step toward institutionalization.

Picture a meeting with an MR (= a medical representative, the pharmaceutical company's information liaison to physicians). The value of that meeting has never been determined by the information alone. Even for the same package-insert content, the physician finds meaning in the fact that "this person explained it in front of me, in the context of my patients." A Q&A session at a medical congress is the same: a speaker answering an unanticipated question on the spot carries trust information that reading from a script cannot replace. The structure in which people seek explanations from people, and the desire to hear it face to face, did not disappear with search engines or with AI. I do not expect it to disappear after 2027 either. If anything, the more AI floods the market with the "quantity" of explanation, the scarcer "who took it on and explained it" becomes.

• Authenticity of experience: first-hand experience — "I actually saw that case," "I worked on that trial." AI cannot possess it in principle, and it becomes the core of a speaker's worth.
• Authenticity of relationship: information from someone you have met repeatedly is received with more weight than a flawless document from a stranger. Trust accumulates as a history, not per transaction.
• Authenticity of signature: putting your name to "I verified this." It is a declaration that you accept the risk of being blamed if you are wrong — a weight anonymous generated output does not have.

Two sober caveats, though. First, a "human-made" label does not guarantee quality. Humans err, tire, and carry bias. A document drafted by AI and carefully verified is often more accurate than one a human dashed off alone. Authenticity is information about who takes responsibility, not information about whether something is correct. Second, the means of proof are immature. Claiming "a human wrote this" is easy; verifying it is hard. Provenance standards are emerging, but their application to plain text lags behind. The years 2027–2029 will likely be a build-out period for the institutions and technology of proving authenticity. Until the labels themselves become trustworthy, the most practical proof of authenticity will remain what it has always been: the ongoing reputation of an identifiable individual.

06How this could be wrong — and what stays the same anyway

The predictions so far are nothing more than extrapolations of current trends. Let me state explicitly the conditions under which they fail. When reading any forecast, it is more honest to look at the conditions for failure before the grounds for success.

Conversely, two things stay the same in every scenario. One: a human bears final responsibility. A human signs the marketing application for a medicine; a human answers for the judgment in an adverse-event report; a human explains the prescription to the patient. Even when AI drafts, cross-checks, and raises alerts, the question "who takes it on when it goes wrong" remains at the bottom of every institution. Two: the structure in which people seek explanations from people. The more important the decision, the less people accept "the system said so" — they want to hear it from the mouth of someone responsible. This desire has persisted as a property of human beings, independent of technical progress.

So this forecast is neither pessimism nor optimism. Knowledge getting cheap looks like a threat to those who have made a living from knowledge. But shift the viewpoint: exactly to the degree that the price of knowledge falls, the worth that lies beyond it comes into focus for the first time — the eye that vouches for certainty, the experience that reads context, and the resolve to take on fallible judgments with a conscience. The difference between the person who merely "knows" and the person who "answers for it" becomes visible, in price and in reputation. If that is what 2029 looks like, I would say it is not a bad year to arrive at.