Tell people what your AI can't do
Tell people what your AI can't do
Tell people what your AI can't do
The short answer
Name the limits where they matter: a short "good at / not yet" note at first use, and a plain "I can't do that, but here's what I can" when a request crosses the line. A blanket disclaimer does neither.
Name the limits where they matter: a short "good at / not yet" note at first use, and a plain "I can't do that, but here's what I can" when a request crosses the line. A blanket disclaimer does neither.
Illustrative concept, not client work or measured results.
A disclaimer is not an expectation
Amershi and colleagues distilled over 150 AI design recommendations into 18 guidelines. The first two: "Make clear what the system can do" and "Make clear how well the system can do what it can do." When 49 HCI practitioners checked 20 popular AI products against them, the authors found the second relevant but not widely implemented.
These are tested heuristics, not proof that a capabilities note improves outcomes. Still, "AI may make mistakes" meets neither. It doesn't say what the feature does or where it stops.
A disclaimer is not an expectation
Amershi and colleagues distilled over 150 AI design recommendations into 18 guidelines. The first two: "Make clear what the system can do" and "Make clear how well the system can do what it can do." When 49 HCI practitioners checked 20 popular AI products against them, the authors found the second relevant but not widely implemented.
These are tested heuristics, not proof that a capabilities note improves outcomes. Still, "AI may make mistakes" meets neither. It doesn't say what the feature does or where it stops.
Say it at first use, and again at the edge
Google PAIR advises being "up-front about what your product can and can't do the first time the user interacts with it." Otherwise, people "may find the boundaries of the system by experimenting in ways it isn't prepared to respond to." When the AI can't complete a request, avoid dead ends: give people a way to finish the task manually.
In this illustrative concept, a meeting-notes assistant opens with a short card: what it's good at, what it can't do yet, and what to double-check. Asked to book a follow-up, it says it can't, and offers to draft the invite instead. For the "how well" half, see Turn AI confidence into a next step.
Say it at first use, and again at the edge
Google PAIR advises being "up-front about what your product can and can't do the first time the user interacts with it." Otherwise, people "may find the boundaries of the system by experimenting in ways it isn't prepared to respond to." When the AI can't complete a request, avoid dead ends: give people a way to finish the task manually.
In this illustrative concept, a meeting-notes assistant opens with a short card: what it's good at, what it can't do yet, and what to double-check. Asked to book a follow-up, it says it can't, and offers to draft the invite instead. For the "how well" half, see Turn AI confidence into a next step.
Before you ship the disclaimer
Ask three questions. After first use, could someone name one thing the feature can't do? Does every out-of-scope request end in a next step? Have you actually seen the limits you list in testing?
Before you ship the disclaimer
Ask three questions. After first use, could someone name one thing the feature can't do? Does every out-of-scope request end in a next step? Have you actually seen the limits you list in testing?