How to design messages for both human and artificial readers.
David N. Silverman is president of North Coast Communications LLC.
Apple Intelligence doesn’t just summarize your emails; it interprets them. Our research found that Apple’s AI summaries differ even when the underlying email is identical. We also uncovered guidelines to help preserve and convey key messages to both human and machine readers.
Apple controls roughly 57% of the U.S. mobile phone market and accounts for nearly half of all email opens globally. With the rapid adoption of AI tools, a future in which many emails may be summarized — and in which 50% are summarized by Apple — is possible.
Despite a few problematic examples, most PR pros check the AI output (and — ahem! — remove the em dashes) before hitting send, post or publish. But because email summarization takes place after distribution, the carefully crafted messages embedded in emails to journalists, clients or the public will read differently than written when summarized by the reader’s iPhone.
When it comes to email, AI will be speaking to your stakeholders for you.
This is a situation that warrants better understanding, so we investigated.
First, we wanted to see if identical email inputs would produce relatively identical summaries in iOS. We used Federalist 51, the Gettysburg Address, the Magna Carta and a recent State Department press release as inputs. We copied and pasted the unchanged text of each document into 3 separate emails — that’s 12 emails total, 3 emails for each of the 4 documents. Then, we sent these emails at different times to different devices to mimic real-world conditions.
What did we find? The AI summarizer is like a tortured genius at an awards show: When you give them the mic, you don’t know what they are going to say next.
This “control” test showed AI summaries of identical emails were considerably different. For example, the first sentences of each summary of Federalist 51:
A: “The Constitution establishes a system of checks and balances to prevent any single branch of government from becoming too powerful.”
B: “To prevent power concentration, the government should be designed with independent departments that check each other.”
C: “The Federalist Papers argue that the Constitution’s separation of powers, with checks and balances, is essential to prevent tyranny.”
These sentences are, respectively, observational, prescriptive and analytical — and convey irrefutably different messages to the reader.
PR pros should be acutely aware of this possibility.
Now on to the good news.
Our research also identified guidelines that preserve email message fidelity for readers both human and artificial. To test the summarizer, we created variations of those well-known documents to send and then summarize and compare with the control. The variations included:
- Moving a paragraph from the end of the document to the beginning.
- Removing the first sentence.
- Adding bold headings as section titles.
- Removing immaterial words and phrases while preserving core meaning.
Here are our findings and recommendations:
First, make sure your key message is in the first paragraph. When we moved a paragraph from the end to the beginning of the document, the summary reflected the elevated content in 75% of tests — even if that wasn’t the text’s main message.
Second, your first sentence should be the key sentence of your key paragraph. In 4 of 12 tests, a change to the first sentence altered the AI summary significantly from the control. Avoid narrative setups. Use a journalistic lead.
Third, use headers judiciously, not just to guide eyeballs or break up text. When testing bold headers as section titles, 50% of the time they significantly altered the summary, including their verbatim insertion into the summary itself. If it shouldn’t be in the summary, avoid making it a header.
Fourth, don’t cut important details for the sake of brevity. In 7 of 12 tests, shorter emails led the AI to drop important context (and likely overweigh whatever remained), sometimes shifting the message entirely.
To be clear, sometimes the differences we found were minor. But minor differences risk major consequences in any email of societal impact, like a health or weather advisory, a federal policy announcement, an investor relations letter or a Fortune 100 press release. Political, financial, diplomatic and crisis communicators should seek to derisk, protect message fidelity, and limit the likelihood of AI misinterpretation. To do so, our research suggests communicators pay close attention to text order, structure and style.
Email remains a core channel for PR, crisis response, investor relations and public affairs, even as AI summarization introduces a new layer of interpretation outside the sender’s control. We’ll have to adjust, and thankfully, we’ve adjusted before: from desktop to mobile-first design and from chronological feeds to algorithmic ones. This shift is unique, however, because it isn’t deciding who sees your message. It decides what your message says.
Your email has two audiences: the human reader and the AI that will remix your message. The sooner our industry begins designing for both, the better the writer and the reader will be.
The post AI email summaries are rewriting your message – here’s what to do about it appeared first on PR Daily.
