PR releases won't fix AI's collapsing reputation

Title: AI's Reputation Collapse: Issuing Press Releases Won't Help

During Friday afternoon group meetings, a student asked me with a case study: Why do search engine AI summaries remain negative even after the company issued dozens of clarification releases? After reviewing the materials, my first reaction was that evaluation standards weren't unified. Traditional PR looks at whether press releases were published, media reposted them, and social platforms cooled down; answer engines look at whether models can retrieve stable, credible, and structurally consistent sources when answering. These two sets of metrics are often not the same thing.

The control design in this experiment is crucial. Over the past month, I led students in running RAG (Retrieval-Augmented Generation), comparing baselines including keyword search, direct generation by standard models, and retrieval-augmented generation with source links. The results were plain: if retrieval evidence contains negative narratives, the model treats them as facts and continues writing. Lewis et al.'s RAG paper stated early on that generation quality largely depends on retrieved evidence. Applied to public opinion, AI summaries act more like amplifiers of corpus distribution.

Gallup data is stark, reportedly showing only 9% of Americans believe AI benefits outweigh drawbacks. This ratio can't be directly extrapolated to the industry, but dataset bias needs consideration. In media, communities, and policy discussions, negative events are more easily recorded; layoffs, energy consumption, erroneous generations, and copyright disputes all enter the corpus. Companies merely issuing positive press releases is equivalent to forcibly adding samples to a skewed distribution; short-term, it might change individual answers, but long-term, it gets pushed back by source weighting and consistency.

Short-term, firefighting requires managing the retrievable fact layer. I agree with a statement in the report: the first 72 hours after a crisis form the primary source records subsequently retrieved by AI engines. From an information retrieval perspective, this isn't exaggerated. Early authoritative sources, structured data, correction pages, and regulatory stances will be repeatedly cited by models. If companies only mass-issue press releases without normalizing sources, ensuring consistent messaging, or marking conflicts, they are essentially leaving the problem for RAG to guess.

Long-term, this PR crisis can't be fixed by the PR department alone. The root of AI companies' reputation issues lies in product promises, data usage, cost spillovers, safety boundaries, and commercialization pressure. I use Dynatrace to monitor agent chains, and the feeling is direct: once the system autonomously calls tools, errors can't be smoothed over by a single piece of copy. Users see where it comes from, what it spent, and what data it touched. Governance mechanisms impact trust more than apology statements.

So, the fire the AI industry needs to put out relies on correcting corpora in the short term and changing behavior in the long term. What's truly actionable is establishing a public opinion assessment system similar to knowledge graph conflict detection, looking at who is saying what, whether sources are traceable, how contradictions are arbitrated, and whether corrections are persistently visible. Otherwise, no matter how many releases are issued, the model summary will still show that one negative line. Before old sources are corrected, issuing supplementary press releases is unlikely to change retrieval results.

---

📌 This article is compiled from Hacker News, original text: https://www.ft.com/content/115c886f-23e4-4a8d-9656-5d7fc9480803

Copyright belongs to the original author. This is a compilation and independent analysis based on public reports.