Imagine receiving a polished answer about a product, a payment or a new policy. It sounds official. It uses the right terms. It may even be in your language. The sentence is not the hard part anymore. The hard part is knowing where it came from, whether it applies to you, and what to do if it is wrong.
That distinction has shaped how I think about AI in communications. The promise is real: more people can draft, translate and adapt useful information quickly. But a fluent answer can also erase the clues that used to tell us whether a person or institution stands behind it. A polished tone is no longer a reliable signal of care, expertise or authority.
Three things a trustworthy answer should carry
First, an accountable source. A reader should be able to trace a consequential claim to the institution responsible for it and to the rule or evidence that supports it. Media provenance tools can help show how an image or video was made or changed; the C2PA standards are one attempt to make that history verifiable. Provenance is useful, but it does not decide whether the underlying claim is fair or complete.
Second, a usable route. People need more than an explanation they can admire. If an answer concerns an account, a payment, a deadline or eligibility, it should lead to the relevant service route. An apparently helpful answer that leaves a person stranded is not yet useful communication.
Third, a way to correct the answer. Institutions should be able to hear where an explanation fails, update it and make the correction visible. This is a management practice, not a copywriting flourish. The NIST AI Risk Management Framework treats trustworthiness as something to govern, measure and manage across a system. Communications teams need an equivalent habit for public answers.
Language is part of the test
In East Africa, “accessible” cannot simply mean an English version with shorter sentences. People may meet the same institution in Kinyarwanda, French, Kiswahili or another language, on a website, in a video and in a messaging channel. The translation has to preserve the action a person can take, not only the words. GSMA’s African Trust & Safety LLM Benchmark is a reminder that models need to be tested in African languages and contexts rather than assumed to transfer safely from an English test.
Publish so an answer can be checked
As more people ask an assistant before visiting an organisation’s website, the official site has a second job. It must give machines a current, unambiguous source to retrieve and give people a place to verify the answer they received. That means one maintained explanation for an important rule, a visible date when it changed, clear service links and fewer contradictory PDFs left to circulate without context.
A communications team could test this with a small set of real audience questions. Ask them on the website, in search and in a few AI assistants. Record the answer each route gives, where it cites its information, whether the answer is current and whether a person can act on it. Then fix the source material and test again. This is not a trick for appearing in more AI results. It is a way of finding out where the institution’s own knowledge has become hard to trust.
Communications teams need to become better stewards of source material, review routes and corrections. If a system can generate a thousand convincing answers, the scarce thing is a dependable institution willing to own one.
