Why the Same AI Boost Can Clarify or Blur Expertise
This chapter explains the mechanism behind the changing value of an AI-assisted hiring or startup-investing pitch. The central question is not how much generative AI improves a message on average, but which senders receive the larger improvement relative to their underlying expertise. That distinction gives readers a practical way to predict when AI may make screening more accurate and when it may make a polished pitch less informative.
The pitch is a signal-extraction problem
The ifo model starts with two people who want different things. A sender wants an employer or investor to form a favorable belief about their capability. A receiver wants to learn the sender's underlying expertise from a pitch that is visible but imperfect. The sender may be a job candidate writing a cover letter or an entrepreneur describing a business idea. The receiver may be a hiring evaluator or an investor. In either case, the decision depends on a hidden trait that cannot be inspected directly at the moment of screening.
A useful signal is not simply a high score. It is an observable feature that changes the receiver's belief in a reliable way. If experienced people can produce a strong pitch with less effort or greater accuracy than inexperienced people, pitch quality contains information. The receiver can treat the quality as evidence because the two groups face different production costs.
The model represents that difference as expertise. Expertise makes a high signal easier to produce. It does not mean that every expert writes beautifully or that every novice writes badly. It means that, across the relevant population, expertise shifts the relationship between effort and the resulting pitch. The receiver does not observe that relationship for one sender directly. They infer from the signal using what they know about the population.
Generative AI adds another source of variation. It gives a sender a boost that can raise the visible quality of the pitch without requiring the same effort as before. Crucially, that boost is not assumed to be identical for every sender. A person who knows the domain may ask better follow-up questions, recognize a weak suggestion, supply missing context, or judge whether an answer is plausible. A person without the relevant knowledge may still obtain fluent language, but may be less able to direct or correct the system. The model keeps these possibilities open by treating expertise and AI boost as distinct characteristics.
The receiver's problem has now changed. Before AI assistance, a strong pitch could be evidence that the sender had paid the effort cost or possessed the knowledge needed to produce it. After AI assistance, some of the same visible quality may come from the tool. The receiver must learn not only from the pitch but from the way the tool's effect is distributed across senders.
The sequence has a feedback loop that is easy to miss. The receiver's trust in the pitch affects how much weight they put on it. That weight affects the sender's return to effort. If the receiver believes AI has made the pitch less diagnostic, the sender may have less reason to invest in producing an even stronger message. Conversely, if AI makes the difference between experts and non-experts clearer, effort may become more valuable. AI is therefore not just a writing aid inserted between thought and text; it can change the incentives on both sides of the screening relationship.
The crucial variable is not the average AI boost
Suppose every sender's pitch becomes 20 percent more polished. That fact alone does not tell the receiver whether screening improved. The receiver needs to know whether the improvement is correlated with the trait they are trying to identify.
If experts' pitches rise by 20 percent and non-experts' pitches rise by 20 percent, the average level has increased but the gap may not have changed. The receiver sees better messages without gaining a clearer basis for separation. If experts rise by 30 percent and non-experts by 10 percent, the level and the gap both rise. If experts rise by 10 percent and non-experts by 30 percent, the lower-expertise group moves toward the higher-expertise group, and a previously useful distinction becomes harder to read.
The ifo paper describes this relationship through the covariance between expertise and the AI boost. “Covariance” can sound like a technical statistic, but the underlying question is straightforward: do the people who are already more expert tend to receive larger or smaller gains from the technology? It is not a score assigned to an individual prompt. It is a population-level relationship that helps explain whether the technology changes the distribution of signals.
This is why the average improvement can be a poor guide to screening quality. A tool can raise the mean quality of a set of pitches while reducing the information in the differences among those pitches. In a hiring queue, the employer may encounter fewer obviously weak applications, but also fewer reliable clues about which applicants have the relevant experience. In a venture queue, more founders may describe a market in fluent language, while the language becomes less revealing about who has actually understood the customer or operating problem.
The model also explains why the evaluator's reaction matters. If a receiver believes that a signal is more informative, they put more weight on it. If the signal is less informative, they discount it and may seek other evidence. The sender anticipates those beliefs. A change in AI access can therefore produce two forces at once: a direct lift to the visible pitch and an incentive response as senders adjust effort to the receiver's changing trust.
Those forces do not always point in the same direction. AI may make a pitch look as though more effort went into it while also reducing the effort needed to produce that appearance. It may help a sender think more clearly, or it may substitute for some of the work that previously revealed knowledge. The model does not assume in advance which force wins.
Three relationships, three possible outcomes
The simplest way to read the mechanism is to compare three cases.
When experts benefit more. If expertise helps a sender use AI more effectively, the tool can widen the signal gap. An expert may be able to provide precise domain context, challenge a generic answer, and recognize which revision improves the pitch rather than merely making it longer. A non-expert may receive some assistance, but not enough to reproduce the expert's domain-specific advantage. In this case, AI can increase screening accuracy because the visible signals separate more clearly.
The consequence is not that every expert's pitch becomes superior or that the tool “knows” who is an expert. The mechanism is about differential gains. If the gain is systematically larger for people who already possess the relevant capability, the receiver has more information to work with.
When non-experts benefit more. The reverse pattern can compress the signals. A non-expert may use AI to fill a knowledge gap or imitate the language associated with experience. The expert also improves, but by less, because their original pitch already contained much of the relevant knowledge or because the tool adds less beyond it. The two outputs move closer together. Screening accuracy falls because the receiver has more difficulty telling the groups apart.
This is the intuitive meaning of a negative relationship between expertise and AI boost in the study's setting. It does not mean that every non-expert receives a larger boost than every expert. It means that, on average, the technology's gains are tilted in a direction that can make non-experts more similar to experts.
When the relationship is extreme or otherwise changes the ordering. The model is more subtle than “negative is always bad.” It shows that screening accuracy can improve when the relationship between expertise and AI boost is sufficiently high or sufficiently low. If experts benefit much more, their signals become clearly higher. Under an extreme negative relationship, the ordering could remain distinctive in the opposite direction: experts might receive little help or even be harmed relative to non-experts, yet the receiver could still learn which group is which if the resulting signals remain systematically different.
The empirically important case is the middle range, where AI benefits senders in ways that blur rather than reverse the distinction. That is where a technology can make communication more uniformly impressive without creating a new reliable marker of expertise. The model therefore rejects two shortcuts: AI is not guaranteed to destroy information, and a positive average effect is not enough to show that it preserves information.
Why level, effort, and dispersion can move differently
The same logic separates three outcomes that are often folded into one word: level, dispersion, and accuracy.
The level of a signal is its average apparent quality. A direct AI boost can raise this level, making pitches seem more polished or sophisticated. The model also allows an incentive response: if receivers trust the signal more, senders may work harder; if receivers trust it less, they may reduce effort. The observed level is the combination of the tool's direct contribution and the sender's changed behavior.
Dispersion is the amount of variation across pitches. It matters because differences can be informative. If AI helps lower-expertise senders catch up to higher-expertise senders, the distribution narrows. If it helps experts pull farther ahead, the distribution widens. The paper's theoretical prediction is therefore about who benefits relative to whom, not about a mechanical tendency toward uniformity.
Accuracy is whether the receiver's belief matches the sender's actual expertise. A receiver can rate every pitch somewhat higher and still become less accurate if the ratings no longer distinguish the underlying groups. Conversely, a lower average level could coexist with better screening if the remaining differences are more diagnostic. A higher or lower mean is not a substitute for measuring error.
These outcomes can also feed back into one another without being identical. A receiver may recognize that a pitch sounds polished and simultaneously discount it because they know AI was available. In that case, the visible level rises while the belief about expertise does not. The receiver may become more uncertain, not more optimistic. That pattern is precisely why the later experiment records several outcomes rather than asking one rating to carry the entire argument.
The model's practical insight is a change in the evaluator's question. Instead of asking only whether AI made the pitch better, ask what variation remains after AI assistance. Is the variation aligned with domain knowledge, or with ease of producing generic fluency? Does the signal distinguish the people who can perform the job or build the venture, or merely those who can obtain a plausible description?
What the model does not settle
The model is deliberately stylized. It focuses on a principal screened characteristic called expertise, even though real employers and investors may care about several traits at once. A hiring decision may involve technical knowledge, communication, reliability, and the ability to work with a particular team. An investment decision may involve customer insight, execution, financing, and timing. A single pitch can contain information about more than one of these, and AI may affect each dimension differently.
It also treats the relationship between expertise and AI boost as a parameter rather than explaining every reason it arises. Language fluency, domain familiarity, prompting skill, access to information, and the ability to recognize errors may all matter. The model tells us that their combined relationship is decision-relevant; it does not identify one universal cause for that relationship.
Finally, the model describes a screening interaction, not an entire career or venture lifecycle. A poor initial signal might be corrected by an interview, a work sample, customer feedback, or later performance. A misleadingly strong signal might be exposed after hiring or funding. Those later consequences could change incentives, but they are not part of the one-shot inference problem the paper models.
The model's value is therefore not that it predicts every pitch. It gives the experiment a testable question: after AI changes the production of a signal, does the difference between expert and non-expert outputs widen, narrow, or remain otherwise informative? The next chapter turns that question into a research design, showing how the hiring and entrepreneurship experiment observed expertise, AI access, receiver belief, pitch quality, and screening error separately.