KeptKept
← All research

Should I let AI write my messages to friends?

Messages known to involve AI are judged less authentic

Hancock, Naaman & Levy · Journal of Computer-Mediated Communication, and subsequent AI-penalty research

Illustrative — the direction is the finding. Schematic. The penalty grows with the degree of AI involvement.

Messages known to involve AI are judged less trustworthy, less authentic and less useful, with larger penalties for greater involvement. A disclosure paradox compounds it: people say disclosure matters and then penalise it when it happens, which creates an incentive not to disclose.

How it was done

A framework for AI-mediated communication together with experimental work comparing how messages are judged depending on the AI involvement people believe is behind them.

What it means

The penalty attaches to belief about AI involvement, not to the text. The same words land differently depending on what the reader thinks produced them.

That inverts the value proposition for any product that drafts your outbound messages. What it produces is precisely the artefact that is losing its worth.

As machine-generated text becomes ordinary, the scarce signal is evidence that a specific person actually noticed something. Generated warmth cannot supply the fact that your friend remembered the surgery was in March.

What it doesn’t show

A fast-moving area with a short literature, and norms around AI-assisted writing are shifting quickly enough that the size of the penalty may not hold. Most studies test judgements of messages in isolation rather than inside an existing relationship, where trust is already established.

What Kept does about it

It is a hard product constraint. Kept will never draft your message. The AI faces inward only — transcribing and structuring what you said — and the words that reach your friend are yours.

Read the source

Hancock, Naaman & LevyJournal of Computer-Mediated Communication, and subsequent AI-penalty research

Confidence: B — a single strong peer-reviewed study, or a consistent multi-study literature. We don’t publish anything weaker than this.

Related

Kept is what we built from this.

A private voice journal for the people closest to you. Nine seconds of talking after you see someone, and the detail comes back before it matters.