When AI Helps the Expert: Context, Language, and Unequal Gains

The pooled experiment showed a screening loss, but the ifo model never predicted that every population would experience the same effect. This chapter compares the estimated relationship between expertise and ChatGPT's signal boost across sender contexts. The practical judgment is conditional: before treating an AI tool as a universal advantage or disadvantage, ask which users can turn it into domain-relevant evidence and whether the observed difference is truly caused by language, access, expertise, or their combination.

One tool, different users

A general-purpose language model does not meet every sender with the same task, vocabulary, or ability to evaluate its suggestions. An entrepreneur who knows an industry may recognize the terms that matter, supply concrete constraints, and reject a plausible but irrelevant answer. Someone without that knowledge may obtain equally fluent sentences but have less basis for deciding which claims are wrong or which details would persuade an informed evaluator.

Language context can intensify that difference. The ifo researchers note that ChatGPT was developed originally in English and that effective use may require follow-up questions, iteration, and the ability to detect hallucinations. A non-native English speaker with deep expertise may still know the English terms used in their industry and use the tool to express domain knowledge more clearly. A non-expert who lacks both the domain knowledge and the language resources may be less able to direct the system toward a genuinely credible pitch.

The reverse pattern can also occur among native or highly fluent users. They may be able to ask ChatGPT enough follow-up questions to make a non-expert pitch sound convincing. If experts already communicate effectively without the tool, the marginal benefit of generic fluency may be larger for the less-experienced sender. The same model then narrows the distinction that an evaluator needs to see.

These are mechanisms, not assumptions that can be assigned to every person in a country. The relevant variable is the relationship between expertise and effective AI use. Language is one possible context for that relationship, alongside domain knowledge, prompting skill, access to information, and the ability to recognize and correct errors.

Reading the three covariance estimates

The ifo study estimates the relationship between expertise and ChatGPT signal boost rather than observing a simple “expert use” score. For the full sample, the point estimate is negative: −0.012, with a bootstrapped standard error of 0.0066 and a p-value of 0.072 when tested against zero. The estimate is consistent with ChatGPT helping non-experts slightly more than experts on average, but the evidence is not as strong as the subgroup estimates. The direction fits the pooled pattern of lower separation and higher screening error.

The subgroup estimates differ sharply. For senders from non-English-speaking countries, the estimated covariance is positive, +0.056, with a standard error of 0.013 and a p-value below 0.001.