When AI Improves the Pitch but Weakens the Signal
This series examines a specific change in hiring and early-stage venture screening: what happens when generative AI helps produce the pitch that an employer or investor uses as evidence. The focus is not whether AI can make sentences smoother. It is whether a smoother application or startup pitch still tells the evaluator who has the relevant expertise, or what a young venture is likely to do under uncertainty. By the end, the useful judgment will be conditional: readers should be able to separate a better-looking output from a better decision signal.
The same screening problem appears in a cover letter and a startup pitch
A hiring manager reading a cover letter is not interested in prose as an isolated artifact. They are trying to infer whether the applicant has the experience and judgment required for a role. An investor reading a venture pitch is making a related inference: does the founder understand the problem, the customers, and the path from an idea to a viable business? The decisions differ, but both ask an evaluator to learn something hidden from something visible.
That relationship is the subject of the evidence considered here. In the hiring and entrepreneurship experiments reported by the ifo/CESifo researchers, job candidates and entrepreneurs wrote pitches and evaluators judged them. The candidates' or entrepreneurs' prior expertise was measured, so the researchers could compare a receiver's impression with the underlying fact. A separate study examined selected large language models as evaluators of 171 new-venture pitches, comparing their assessments with fundraising outcomes, later survival, and a benchmark group of human investors.
The studies do not test one unified AI system. In the first, AI changes the sender's pitch. In the second, an LLM becomes the evaluator. They meet around a common question: when an output looks persuasive, what exactly has the evaluator learned?
The dashed connection is the part that matters. A pitch is supposed to carry information from the hidden capability to the evaluator's judgment. AI may strengthen the visible artifact while leaving that connection unchanged, weakening it, or—under some conditions—making it clearer. The surface result alone cannot tell us which case we are in.
Why a pitch is more than prose
A signal is useful when producing it is not equally easy for everyone. Someone with genuine experience may find it easier to describe a technical problem accurately, anticipate an investor's concern, or choose details that reveal practical understanding. The effort is not necessarily dramatic or visible. What matters is that the relationship between capability and output is different for people with different levels of expertise.
That difference gives an evaluator something to learn from. A strong pitch is not valuable only because it is polished; it is valuable because its quality is harder for an unqualified sender to reproduce. This is why the cost of producing a signal matters. The evaluator is not merely asking, “Does this sound good?” They are asking, “How much should this sound tell me about the person or venture behind it?”
Generative AI changes the production technology. It can help someone organize an argument, supply language, or turn a rough idea into a more finished-looking pitch. That assistance may be genuinely useful. It may help an expert express knowledge that would otherwise remain poorly communicated. It may also allow a non-expert to produce language that resembles the output of an expert. The tool has changed the cost of the visible signal, but it has not automatically changed the sender's underlying expertise.
That distinction is easy to lose because people often use a single word—quality—for several different things. A pitch can have higher linguistic sophistication. It can receive a higher rating from a reader. It can be more accurate about a venture's prospects. Or it can reveal more about the sender's real capability. These are related properties, not interchangeable ones.
For a receiver, the important task is classification, not admiration. They must decide whether a sender belongs to the group with the relevant experience, or whether a venture deserves further attention, while knowing that they cannot observe the underlying trait directly. A polished message can produce a high evaluation without making the classification easier. In fact, if polished language becomes common among people with very different levels of expertise, it can increase uncertainty even while it raises the average impression of quality.
This is why the difference between an average score and a separation score matters. An average score asks whether pitches look better overall. A separation score asks whether the evaluator can tell the stronger cases from the weaker ones. A hiring process can produce more impressive cover letters and still become worse at identifying candidates who have done the work before. A venture process can receive more coherent descriptions and still learn little about how a team will respond when the plan meets an unanticipated constraint.
The venture case also makes the problem harder because the hidden target is not a fixed qualification alone. A founder's expertise may be relevant to a pitch, but the eventual outcome depends on customers, competitors, financing, execution, and events that no pitch fully reveals. The abstract-described LLM study therefore separates realized fundraising from later survival. Matching a recorded fundraising outcome can be useful, but it does not settle what an evaluator knows about a future that was uncertain at the time of the decision.
Lowering the cost of communication is not itself a failure. It can make participation easier for people who know something but struggle to express it in a dominant language or conventional format. The problem begins when an evaluator continues to interpret the new output as if its production cost and diagnostic value were unchanged. The burden then moves from writing to verification: the organization must find another way to test the capability that the pitch no longer reveals as cleanly.
Nor does this mean that every use of AI removes information. The source material supports a conditional account. If a person with real expertise can use the tool to express domain knowledge more effectively than a non-expert can, AI may strengthen the signal. If the tool supplies similar polish to both, the same improvement may carry less information. The visible product cannot tell us which pattern occurred; the relationship between the sender, the tool, and the evaluator has to be examined.
The tension behind “better”
The title's tension is therefore precise. AI may improve the pitch as an object while weakening the pitch as evidence. In the ifo experiments, the relevant comparison was not between a bad human text and an ideal machine text. It was between pitches produced under different conditions, evaluated against known differences in sender expertise. That makes it possible to ask whether the evaluator became better at separating experts from non-experts, not just whether the writing became more attractive.
The result, developed later, is consequential because it reverses a common intuition. A more polished message can make a reader feel that more information has arrived. But if many senders can now reach a similar level of polish, the additional smoothness may contain less information about their differences. The output has improved on one dimension while losing variation on another.
The second study creates a parallel caution from the evaluator side. Its abstract reports that selected LLMs were good at mirroring real fundraising outcomes for the ventures they evaluated. That is a meaningful capability, but it is not identical to reasoning well about an uncertain future. The same abstract says that apparent accuracy in predicting survival largely reflected prior exposure to some ventures' digital footprints rather than genuine reasoning under uncertainty. An evaluator can therefore look accurate for a reason that does not transfer to a new case.
This is not an argument that the models were useless, or that every prediction was contaminated. It is an argument about interpretation. Agreement with a known outcome may reflect useful judgment, accessible information, or prior exposure. A polished pitch may reflect expertise, effective assistance, or both. Before deciding what an output proves, we need to identify the pathway by which the output became persuasive.
Two directions of change, one decision problem
Sender-side and evaluator-side AI alter different links in the screening chain. When a candidate or founder uses ChatGPT, the production of the signal changes before the employer or investor sees it. The evaluator may know that AI was used, but still has to infer capability from a signal whose production cost has shifted. When an LLM evaluates a venture, the evaluator itself changes. The system may combine the pitch with scientific, contextual, or social cues, and its apparent success must be compared with actual outcomes and human judgments.
The comparison has limits. The ifo study is a controlled experiment about sender-side access to ChatGPT in hiring and entrepreneurship. The LLM venture-evaluation study, as available here, is represented by its abstract and concerns an archival set of venture pitches. Their samples, roles, measures, and evidence are not interchangeable. What can be carried across them is narrower and more useful: visible fluency and outcome agreement do not, by themselves, establish that the signal contains the decision-relevant information.
That principle also avoids two opposite mistakes. The first is to treat any AI assistance as deception and conclude that the pitch should be ignored. The second is to treat improved wording or apparent predictive success as proof that the underlying decision has improved. Both mistakes collapse different layers of the problem into one score.
A responsible evaluator needs to keep at least four questions separate. Has the output become more polished? Does it match an observed outcome? Does it identify the capability that matters? And did it reach the conclusion through evidence that would still be available in a new, uncertain case? The sources support asking all four, even though they do not answer all four in the same way.
The question to carry forward
The practical stakes are larger than the fate of one cover letter or pitch deck. Employers and investors allocate opportunities using signals because direct knowledge of a candidate or venture is expensive. If AI changes which signals are cheap, common, or easy to imitate, organizations may either make more screening mistakes or spend more on interviews, background checks, market research, and other forms of validation. The cost of communicating may fall while the cost of knowing rises.
The next step is to examine the mechanism rather than choosing a verdict. The key issue is not simply how much AI improves a pitch on average. It is who benefits from that improvement. If experts gain more than non-experts, the technology can make the distinction easier to see. If non-experts gain more, or everyone is lifted toward the same polished middle, the signal may become weaker. That relationship between expertise and AI's boost is the hinge on which the rest of the evidence turns.