You receive a kind reply. It acknowledges the awkward part of your message, offers a clear answer, and ends warmly. Then you wonder whether the sender chose the words or selected a suggestion.
The message has not changed. Your reading of the relationship may have.
That small uncertainty sits at the center of a useful 2023 study of AI-mediated communication, later published in Scientific Reports. Researchers including Jess Hohenstein and Malte Jung studied suggested replies in an experimental messaging system. In one experiment, 219 pairs discussed a policy issue with different access to suggestions. The researchers found more efficient communication and more positive language with the tool, alongside positive interpersonal evaluations associated with its use.
They also found a different pattern around suspicion: people who thought a partner used suggestions evaluated that partner less favorably. The association with perceived use does not prove that disclosure caused the negative response. Cornell's account of the research explains this tension. These were bounded experiments with an earlier smart-reply system, not a survey of every contemporary workplace or friendship.
The distinction still matters. What a tool does to language and what people think its use means can move in different directions.
We read effort between the lines
A message carries content, but it also offers a clue about attention. Someone noticed the detail. Someone took a moment. Someone found words for a difficult answer.
Those judgments have never been perfect. A long reply can be careless. A short reply can reflect deep familiarity. A person writing in a second language may need help expressing a thought they have considered carefully. An assistant can remove friction that has little to do with care.
Yet the feeling of having been attended to remains important. If we imagine the reply arrived with no human engagement, its warmth can start to feel generic. This is an interpretation of the social problem, not a claim that the study measured every form of attention.
The temptation is to turn authorship into a binary test: either the person wrote every word or the exchange is inauthentic. That leaves out translation, editing, accessibility tools, advice from friends, and all the other ways language has always been supported.
A suggestion can quietly become a stance
Consider a fictional exchange between colleagues. One proposes a change. The other feels uncertain but is offered a bright, agreeable reply. Accepting it may save time while making the response sound more enthusiastic than intended.
The important question is what the sender means to communicate. A friendly phrase can be appropriate; agreement that the person does not feel is a different matter. Before sending, the writer can ask whether the message preserves uncertainty, boundaries, and the promise they actually intend to make.
This is not an argument for making every message visibly rough. Clear, thoughtful writing can be polished. The issue is whether the polish has changed the relationship signal. An apology that assumes responsibility, an invitation that implies commitment, or praise that overstates conviction carries more than tone.
Suggested language becomes useful when it helps the sender say what they mean. The sender still has to recognize the meaning before allowing it to travel.
Transparency needs a purpose
The research does not establish one disclosure rule for every message. Adding a label to a routine scheduling reply and explaining substantial assistance on a sensitive letter serve different needs.
Ask what the recipient reasonably needs to understand. Are they relying on personal judgment, a considered commitment, or merely a time and place? Would the assistance change their interpretation of the sender's involvement? Can the sender explain what they reviewed and why they stand behind it?
Organizations can discuss those questions without making colleagues prove that their prose looks sufficiently human. Style is a poor witness to the amount of care someone invested. A predictable phrase may be sincere; an unusual phrase may be generated.
For sensitive exchanges, the strongest signal may be behavior after the message: listening to the response, answering a specific question, correcting a misunderstanding, or following through. Those actions give the recipient more information than a polished opening line.
Keep the relationship available for correction
The experiment's social puzzle should make us curious, not suspicious of every courteous sentence. People can misread assistance, and assistance can help people participate more fully. Both possibilities deserve room.
A useful habit is to reread a generated reply as a statement of intention. Would you say this in the next conversation? Can you explain the promise it contains? If the recipient challenges it, will you engage rather than point to the tool?
The hidden coauthor matters most when it becomes a hidden substitute for attention. A warm reply can still belong to a person who selected it thoughtfully. The evidence of that care is the person who remains present when the conversation continues.
Sources and limitations
- Hohenstein et al., “Artificial intelligence in communication impacts language and social relationships”: checked October 15, 2026. Author-posted primary research later published in Scientific Reports; experimental findings do not establish universal social effects or the causal effect of disclosing AI use.
- Cornell Chronicle, “Study uncovers social cost of using AI in conversations”: checked October 15, 2026. First-party university account of the 2023 study, including experimental design and the distinction between actual and perceived use.