01
An ambient AI scribe listens to a clinician–patient visit and drafts the note.
Secondhand AI
Someone else used AI. You received the effect.
Secondhand AI is the downstream effect of one person's AI use on other people in their social network.
We are aware how AI can influence its user, let's call her Alice. Far less attention goes to how AI influences the receiver of Alice's AI use, let's call him Bob. Bob never used AI, and often has no idea Alice did. This is about what happens to Bob.
Let's look at an example:
An AI scribe writes up a single doctor's visit. Scroll sideways to follow where it goes →
02
The clinician reviews it, signs it, and files it into the medical record.
03
From there it travels to specialists, insurers, the patient, and later clinicians.
04
But AI slips: one audit found scribes recording the wrong drug, inventing treatments, or missing details.
05
If the first clinician doesn't catch it, the error becomes part of the record everyone downstream acts on.
06
Neither the patient nor the next doctor used AI, yet both are affected by it.
How can AI impact Bob?
Trust travels through social networks. Long before AI, no one could check everything themselves, we rely on each other to know most of what we act on.
Trust runs on three overlapping signals:
- 1Authority — the roles and expertise that make a judgment worth respecting.
- 2Verification — the evidence of effort and expertise we can check for ourselves.
- 3Influence — the beliefs and norms that travel through the people who pass information on.
Secondhand AI can erode each one.
Three overlapping routes of secondhand AI
AI enters a network through Alice; what it changes travels along ordinary ties through authority, verification, and influence. These modes are neither exhaustive nor discrete. Scroll to explore each →
Authority
“Who should I trust?”
AI can produce confident and authoritative prose, appearing knowledgeable, regardless of whether Alice actually is knowledgeable.
Example
Topic summaries written by your top-scoring friend read as complete and detailed even when they miss the nuance, so Bob trusts it more and questions it less.
Negative effect
Bob's confidence outruns the real understanding behind the note.
Positive effect
Design AI to flag thin or uncertain reasoning, and Bob knows where to look.
Verification
“What must I check?”
AI makes content cheap to produce for Alice but costly for Bob to verify as AI's notes are longer than that of humans.
Example
When clinical notes are written by AI scribes, the next clinician faces more to read with no more time, leading to a shift in their verification behaviour. They review less carefully, and an error slips into the record.
Negative effect
Bob stops checking and starts trusting, or adopts AI to keep up. Allowing errors to stay.
Positive effect
Design AI to write shorter than a human would, and it lightens Bob's load.
Influence
“What should I believe?”
AI can shape Alice's judgement, which she carries into her next conversation, influencing Bob's judgement.
Example
Alice asks an AI to sum up the parties before an election; its slanted, selective framing quietly shapes who she picks. She passes that view to Bob as her own honest judgment, so he ends up holding the AI's framing without either of them realising it came from a model.
Negative effect
An AI-shaped opinion reaches Bob with no artifact to check.
Positive effect
If AI challenges Alice instead of answering for her, she brings sharper thinking to Bob.
Is it like secondhand smoke?
The cost of smoking to one's health does not stay with the user but travels through the air to people in their surroundings. Similarly, the costs and influences of AI use does not stay with the user but travels through social channels carrying knowledge, trust, or professional judgment. Unlike smoke, though, secondhand AI can help as well as harm. Whether a secondhand AI effect is positive or negative comes down to the design of AI tools.
Designing for the recipient
Many current secondhand AI effects are negative. To minimise the number and extent of negative effects, AI tools must design for the recipient.
The design fixes depend on whether Bob receives Alice's output or talks to a changed Alice. Content generated with the help of AI could shorten instead of lengthen work, could flag what parts were edited or left unchanged, and could express instead of suppress uncertainties. AI tools could challenge Alice's opinions and prompt critical thinking, instead of strengthening her opinions.
Disclosure of AI use is a start, but Bob needs enough information to actually calibrate what he receives. And where AI may have changed Alice's judgment, institutions can check she can still explain and defend it independently.
The problem isn't AI itself. It's how the models are designed and how they're deployed.
Why it matters
Because these effects travel through people, they don't stop at Bob. They reach Dave, Carol, and whoever's next, and can shift the standards of a whole community, not just one person's understanding. They strain the relationships trust is built on: doctor and patient, teacher and student. None of it is intended; it's the spillover of AI use that was never meant to reach anyone else — which is exactly why it's so easy to miss, and why it needs attention now.
Research
The studies behind this work will be collected here — more soon.
Real life examples
If you spot it, tag it and it appears here. Filter by effect, mode, and domain, and open each card for a closer look.
Use the term.
If you spot it, name it. Tag it and it lands in the examples above.