Ikibazo gihuza ibibazo · Biragoye Urwego
Ikibazo cyimyitwarire ya AI
Ibizamini byagize ingaruka ku bantu, amahame akoreshwa, kubazwa, ubutware bwa muntu bufite ireme, ubundi buryo, gusubiramo bikomeje, hamwe ningaruka zisigaye.
Ikibazo 1 Bya 8
Kuki isuzuma ryimyitwarire idashobora kugabanywa kumanota imwe yicyitegererezo?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Kuki isuzuma ryimyitwarire idashobora kugabanywa kumanota imwe yicyitegererezo?
- Amanota ntabwo ari ingirakamaro
- Biterwa nintego, abantu bagizweho ingaruka, gucuruza, no gukwirakwiza ibibi
- Imyitwarire ireba kubahiriza amategeko gusa
- Amanota menshi yerekana ubutabera
- Serivise yorohereza uyikoresha ariko iremerera abantu badashobora guhitamo. Isubiramo ryimyitwarire ikwiye kubaza iki?
- Who benefits, bears extra work, and has influence over the design
- Only whether the interface is fast
- Whether the model is open-weight
- How many features can be automated
- Itsinda rivuga ko gukorera mu mucyo ari ngombwa. Niki cyatuma ihame rikorwa?
- Add “transparent” to the product description
- Publish the model’s parameter count
- Ask the model to explain itself
- Specify what information users need and when they receive it
- Ni ukubera iki uruhare rwabantu rudahita rurinda imyitwarire?
- Humans cannot find model errors
- Human review always makes systems slower
- Reviewers need relevant authority and resources to affect outcomes
- Automation is always more ethical
- Ni iki inzira yo kubazwa igomba kumenya?
- Who can investigate, correct, or stop an inappropriate outcome
- Only the vendor’s contact form
- A model’s confidence score
- The number of review meetings
- When should an ethical review be revisited?
- Only before procurement
- After deployment as uses, data, or feedback change
- Only when the model architecture changes
- Never after approval
- A simpler non-AI process can meet the same need with fewer burdens. What should the team do?
- Choose AI because it is innovative
- Ignore staffing and procedure options
- Compare the alternative instead of assuming more automation is better
- Keep the model but remove user feedback
- How should a review report handle disagreement and remaining risk?
- Remove disagreement from the record
- Describe the review as a safety guarantee
- Record only risks the model can measure
- Document both rather than claiming review proves no harm can occur